View source for Large language model

{{Distinguish|Logic learning machine}}
{{redirect|LLM}}
{{Technical|introduction|date=January 2026}}
{{Machine learning|Neural networks}}
A '''large language model''' ('''LLM''') is a [[computational model]] designed to perform [[natural language processing]] tasks, especially [[language generation]], using contextual relationships derived from a large set of training data.<ref name="bhaa">Bommasani, Rishi. "On the Opportunities and Risks of Foundation Models". 2021.</ref><ref name="few-shot-learners">Brown, Tom B.. "Language Models are Few-Shot Learners". 2020.</ref> LLMs can generate, summarize, translate and parse text in a variety of contexts,<ref name="scaling-laws">Kaplan, Jared. "Scaling Laws for Neural Language Models". 2020.</ref> and are the technological underpinning of modern [[chatbot]]s.<ref name="few-shot-learners2">Brown, Tom B.. [https://proceedings.neurips.cc/paper/2020/file/1457c0d6bfcb4967418bfb8ac142f64a-Paper.pdf "Language Models are Few-Shot Learners"]. ''Advances in Neural Information Processing Systems''. Dec 2020.</ref> LLMs can accurately mimic natural language patterns because they are trained on [[Text corpus|collections of human-written text]].<ref>Fathallah, Nadeen. [https://2024.eswc-conferences.org/wp-content/uploads/2024/05/77770034.pdf "NeOn-GPT: A Large Language Model-Powered Pipeline for Ontology Learning"]. 2024-05-26.</ref> For the same reason, biased or inaccurate training data can make a LLM's output less reliable.<ref name="Manning-2022">Manning, Christopher D.. [https://www.amacad.org/publication/human-language-understanding-reasoning "Human Language Understanding & Reasoning"]. ''Daedalus''.</ref> 

As of 2024, the largest and most capable LLMs are all based on [[Transformer_(deep_learning)|transformer]] architectures, which can be more efficient and parallelizable<ref>Vaswani, Ashish. "Attention is All you Need". 2017.</ref> than earlier [[Statistical model|statistical]] and [[recurrent neural network]] models.<ref>Merritt, Rick. [https://blogs.nvidia.com/blog/2022/03/25/what-is-a-transformer-model/ "What Is a Transformer Model?"]. ''NVIDIA Blog''. 2022-03-25.</ref> Research into other architectures, such as [[state-space representation|state space]] models, is ongoing.<ref>Peng, Bo. [https://aclanthology.org/2023.findings-emnlp.936/ "RWKV: Reinventing RNNS for the Transformer Era"]. ''EMNLP''. 2023.</ref><ref>Gu, Albert. "Mamba: Linear-Time Sequence Modeling with Selective State Spaces". 2023-12-01.</ref> 

Models like [[GPT (language model)|GPT]], [[BERT (language model)|BERT]], and their successors used these advances to demonstrate [[Emergence|emergent behaviors]] at scale, such as finding specific data from a large data set and compositional reasoning.<ref>Devlin, Jacob. "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding". 2018.</ref>

[[Language model benchmark|Benchmark]] evaluations for LLMs test a model's ability to perform one or more language tasks. Modern LLMs may face comprehensive, [[Multi-task learning|multi-task]] evaluations measuring [[Reasoning model|reasoning]], [[Accuracy and precision|factual accuracy]], [[AI alignment|alignment]], and [[AI safety|safety]].<ref>Wang, Alex. "GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding". 2018.</ref><ref>Hendrycks, Dan. "Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency". 2025.</ref> Optimizing training sessions to pass benchmarks may result in a model that [[Overfitting|adheres too closely to benchmark outputs]] without genuine [[Generalization (machine learning)|generalization]] or robust capability improvements.<ref name=":3">Recht, Benjamin. "Do ImageNet Classifiers Generalize to ImageNet?". 2019.</ref>

==History==
[[File:The number of publications about Large Language Models by year.png|thumb|The number of publications about large language models by year grouped by publication types]]

[[File:Trends_in_AI_training_FLOP_over_time_(2010-2025).svg|thumb|The training compute of notable large models in FLOPs vs publication date over the period 2010–2024. For overall notable models (top left), frontier models (top right), top language models (bottom left) and top models within leading companies (bottom right). The majority of these models are language models.]]
[[File:Large-scale_AI_training_compute_(FLOP)_vs_Publication_date_(2017-2024).svg|thumb|The training compute of notable large AI models in FLOPs vs publication date over the period 2017–2024. The majority of large models are language models or multimodal models with language capacity.]]
Before the emergence of transformer-based models in 2017, some [[language model]]s were considered large relative to the computational and data constraints of their time. In the early 1990s, [[IBM]]'s statistical models pioneered [[Bitext word alignment|word alignment]] techniques for machine translation, laying the groundwork for [[Construction grammar|corpus-based language modeling]]. In 2001, a smoothed [[Word n-gram language model|''n''-gram model]], such as those employing [[Kneser–Ney smoothing]], trained on 300 million words, achieved state-of-the-art [[perplexity]] on benchmark tests.<ref>Goodman, Joshua. "A Bit of Progress in Language Modeling". ''Computer Speech & Language''. 2001-08-09.</ref> During the 2000s, with the rise of widespread internet access, researchers began compiling massive text datasets from the web ("web as corpus"<ref>Kilgarriff, Adam. [https://direct.mit.edu/coli/article/29/3/333-347/1816 "Introduction to the Special Issue on the Web as Corpus"]. ''Computational Linguistics''. September 2003.</ref>) to train statistical language models.<ref>Banko, Michele. "Scaling to very very large corpora for natural language disambiguation". ''Proceedings of the 39th Annual Meeting on Association for Computational Linguistics - ACL '01''. 2001.</ref><ref>Resnik, Philip. [https://direct.mit.edu/coli/article/29/3/349-380/1809 "The Web as a Parallel Corpus"]. ''Computational Linguistics''. September 2003.</ref>

Moving beyond ''n''-gram models, researchers started in 2000 to use neural networks to learn language models.<ref>Xu, Wei. "6th International Conference on Spoken Language Processing (ICSLP 2000)". ISCA. 2000-10-16.</ref> Following the breakthrough of [[Deep learning|deep neural networks]] in image classification around 2012,<ref>Citation needed.</ref> similar architectures were adapted for language tasks. This shift was marked by the development of [[word embedding]]s (e.g., [[Word2vec|Word2Vec]] by Mikolov in 2013) and sequence-to-sequence ([[seq2seq]]) models using [[Long short-term memory|LSTM]]. In 2016, Google transitioned its translation service to [[neural machine translation]] (NMT), replacing statistical phrase-based models with deep [[recurrent neural network]]s. These early NMT systems used LSTM-based [[Encoder-decoder model|encoder-decoder architectures]], as they preceded the invention of [[Transformer (deep learning architecture)|transformers]]. [[File:The-Transformer-model-architecture.png|thumb|upright=1.3|An illustration of the main components of the transformer model from the original paper, where layers were normalized after (instead of before) multiheaded attention]]
At the 2017 [[NeurIPS]] conference, [[Google]] researchers introduced the transformer architecture in their landmark paper "[[Attention Is All You Need]]".<ref>Vaswani, Ashish. [https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf "Attention is All you Need"]. ''Advances in Neural Information Processing Systems''. 2017.</ref> This paper's goal was to improve upon 2014 seq2seq technology,<ref>Ilya Sutskever; Oriol Vinyals; Quoc V. Le. [https://dl.acm.org/doi/10.5555/2969033.2969173 "Sequence to sequence learning with neural networks"]. ''Proceedings of the 28th International Conference on Neural Information Processing Systems''. 2014.</ref> and was based mainly on the [[attention (machine learning)|attention]] mechanism developed by Bahdanau et al. in 2014.<ref>Bahdanau, Dzmitry. "Neural Machine Translation by Jointly Learning to Align and Translate". 2014.</ref> The following year in 2018, [[BERT (language model)|BERT]] was introduced and quickly became "ubiquitous".<ref>Rogers, Anna. [https://aclanthology.org/2020.tacl-1.54 "A Primer in BERTology: What We Know About How BERT Works"]. ''Transactions of the Association for Computational Linguistics''. 2020.</ref> Though the original transformer has both encoder and decoder blocks, BERT is an encoder-only model. Academic and research usage of BERT began to decline in 2023, following rapid improvements in the abilities of decoder-only models (such as GPT) to solve tasks via [[Prompt engineering|prompting]].<ref name="auto">Movva, Rajiv. "Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)". 2024.</ref>

Although decoder-only [[GPT-1]] was introduced in 2018, it was [[GPT-2]] in 2019 that caught widespread attention because [[OpenAI]] claimed to have initially deemed it too powerful to release publicly, out of fear of malicious use.<ref>Hern, Alex. [https://www.theguardian.com/technology/2019/feb/14/elon-musk-backed-ai-writes-convincing-news-fiction "New AI fake text generator may be too dangerous to release, say creators"]. ''[[The Guardian]]''. 14 February 2019.</ref> [[GPT-3]] in 2020 went a step further and As of 2025 is available only via [[Web API|API]] with no offering of downloading the model to execute locally. But it was the 2022 consumer-facing chatbot [[ChatGPT]] that received extensive media coverage and public attention.<ref>[https://www.euronews.com/next/2023/11/30/chatgpt-a-year-on-3-ways-the-ai-chatbot-has-completely-changed-the-world-in-12-months "ChatGPT a year on: 3 ways the AI chatbot has completely changed the world in 12 months"]. [[Euronews]]. November 30, 2023.</ref> The 2023 [[GPT-4]] was praised for its increased accuracy and as a "holy grail" for its [[Multimodal learning|multimodal]] capabilities.<ref>Heaven, Will. [https://www.technologyreview.com/2023/03/14/1069823/gpt-4-is-bigger-and-better-chatgpt-openai/ "GPT-4 is bigger and better than ChatGPT—but OpenAI won't say why"]. [[MIT Technology Review]]. March 14, 2023.</ref> OpenAI did not reveal the high-level architecture and the number of [[Parameter#Artificial intelligence|parameters]] of GPT-4. The release of ChatGPT led to an uptick in LLM usage across several research subfields of computer science, including robotics, software engineering, and societal impact work.<ref name="auto"/> In 2024 OpenAI released the [[Reasoning language model|reasoning model]] [[OpenAI o1]], which generates long chains of thought before returning a final answer.<ref name="NYTimesInfo">Metz, Cade. [https://www.nytimes.com/2024/09/12/technology/openai-chatgpt-math.html "OpenAI Unveils New ChatGPT That Can Reason Through Math and Science"]. ''[[The New York Times]]''. September 12, 2024.</ref> Many LLMs with parameter counts comparable to those of OpenAI's GPT series have been developed.<ref>[https://ourworldindata.org/grapher/artificial-intelligence-parameter-count?time=2017-09-05..latest "Parameters in notable artificial intelligence systems"]. ''ourworldindata.org''. November 30, 2023.</ref>

Since 2022, open-weight models have been gaining popularity, especially at first with [[BLOOM (language model)|BLOOM]] and [[LLaMA]], though both have restrictions on usage and deployment. [[Mistral AI]]'s models Mistral 7B and Mixtral 8x7b have a more permissive [[Apache License]]. In January 2025, [[DeepSeek]] released DeepSeek R1, a 671-billion-parameter open-weight model that performs comparably to OpenAI o1 but at a much lower price per token for users.<ref>Sharma, Shubham. [https://venturebeat.com/ai/open-source-deepseek-r1-uses-pure-reinforcement-learning-to-match-openai-o1-at-95-less-cost/ "Open-source DeepSeek-R1 uses pure reinforcement learning to match OpenAI o1 — at 95% less cost"]. ''VentureBeat''. 2025-01-20.</ref>

Since 2023, many LLMs have been trained to be [[Multimodal learning|multimodal]], having the ability to also process or generate other types of data, such as images, audio, or 3D meshes.<ref>[https://research.nvidia.com/labs/toronto-ai/LLaMA-Mesh/ "LLaMA-Mesh"]. ''research.nvidia.com''. 2024.</ref> These LLMs are also called large multimodal models (LMMs),<ref>Zia, Dr Tehseen. [https://www.unite.ai/unveiling-of-large-multimodal-models-shaping-the-landscape-of-language-models-in-2024/ "Unveiling of Large Multimodal Models: Shaping the Landscape of Language Models in 2024"]. ''Unite.AI''. 2024-01-08.</ref> or multimodal large language models (MLLMs).<ref>Wang, Jiaqi. "A Comprehensive Review of Multimodal Large Language Models: Performance and Challenges Across Different Tasks". 2024-08-02.</ref><ref>[https://www.ibm.com/think/topics/multimodal-llm "What is a Multimodal LLM (MLLM)?"]. ''IBM''. 2025-07-30.</ref>

Open-weight LLMs have increasingly shaped the field since 2023, contributing to broader participation in AI development and greater transparency in model evaluation. Vake et al. (2025) demonstrated that community-driven contributions to open-weight models measurably improve their efficiency and performance, with user participation growing rapidly on collaborative platforms such as Hugging Face.<ref>Vake, Domen. "Is Open Source the Future of AI? A Data-Driven Approach". ''Applied Sciences''. 5 March 2025.</ref> Paris et al. (2025) further argued that openness in AI should extend beyond releasing model code or weights to encompass inclusiveness, accountability, and ethical responsibility in AI research and deployment.<ref>Paris, Tamara. "Opening the Scope of Openness in AI". Association for Computing Machinery. 23 June 2025.</ref> Collectively, these studies highlight that open-weight LLMs can accelerate innovation and enhance scientific reproducibility, while fostering a more transparent and participatory AI ecosystem.

== Dataset preprocessing ==
''See also: [[List of datasets for machine-learning research#Internet]]''

===Tokenization===
As [[machine learning]] algorithms process numbers rather than text, the text must be converted to numbers. In the first step, a vocabulary is decided upon, then integer indices are arbitrarily but uniquely assigned to each vocabulary entry, and finally, an [[Word embedding|embedding]] is associated to the integer index. Algorithms include [[byte-pair encoding]] (BPE) and WordPiece. There are also special tokens serving as [[control character]]s, such as <code>[MASK]</code> for masked-out token (as used in [[BERT (language model)|BERT]]), and <code>[UNK]</code> ("unknown") for characters not appearing in the vocabulary. Also, some special symbols are used to denote special text formatting. For example, "Ġ" denotes a preceding whitespace in [[RoBERTa]] and GPT and "##" denotes continuation of a preceding word in BERT.<ref>Kaushal, Ayush. [https://aclanthology.org/2022.naacl-main.179.pdf "What do tokens know about their characters and how do they know it?"]. ''NAACL''. 2022-06-06.</ref>

For example, the BPE tokenizer used by the legacy version of [[GPT-3]] would split <small><code>tokenizer: texts -> series of numerical "tokens"</code></small> as
{| cellpadding="0;" cellspacing="0;" style="border:1px solid black"
|-
| style="border-left: 2px green; border-right: 2px green" |token 
| style="background-color: grey; color: white; border-left: 2px green; border-right: 2px green" |izer 
| style="border-left: 2px green; border-right: 2px green" |:
| style="background-color: grey; color: white; border-left: 2px green; border-right: 2px green" |&nbsp;texts
| style="border-left: 2px green; border-right: 2px green" |&nbsp;->
| style="background-color: grey; color: white; border-left: 2px green; border-right: 2px green" |series 
| style="border-left: 2px green; border-right: 2px green" |&nbsp;of
| style="background-color: grey; color: white; border-left: 2px green; border-right: 2px green" |&nbsp;numerical 
| style="border-left: 2px green; border-right: 2px green" |&nbsp;" 
| style="background-color: grey; color: white; border-left: 2px green; border-right: 2px green" |t
| style="border-left: 2px green; border-right: 2px green" |ok
| style="background-color: grey; color: white; border-left: 2px green; border-right: 2px green" |ens
| style="border-left: 2px green; border-right: 2px green" |" 
|}

Tokenization also [[Data compression|compress]]es the datasets. Because LLMs generally require input to be an [[Array (data structure)|array]] that is not [[Jagged array|jagged]], the shorter texts must be "padded" until they match the length of the longest one. The average number of words per token depends on the language.<ref>[https://blog.yenniejun.com/p/all-languages-are-not-created-tokenized "All languages are NOT created (tokenized) equal"]. ''Language models cost much more in some languages than others''. 2023-05-03.</ref><ref name="LangModelTokenizsersUnfairness">Petrov, Aleksandar. [https://openreview.net/forum?id=Pj4YYuxTq9 "Language Model Tokenizers Introduce Unfairness Between Languages"]. ''NeurIPS''. June 23, 2023.</ref>

==== Byte-pair encoding ====
''Main article: [[Byte-pair encoding]]''

As an example, consider a tokenizer based on byte-pair encoding. In the first step, all unique characters (including blanks and [[punctuation mark]]s) are treated as an initial set of [[n-gram|''n''-grams]] (i.e. initial set of uni-grams). Successively the most frequent pair of adjacent characters is merged into a bi-gram and all instances of the pair are replaced by it. All occurrences of adjacent pairs of (previously merged) ''n''-grams that most frequently occur together are then again merged into even lengthier ''n''-gram, until a vocabulary of prescribed size is obtained. After a tokenizer is trained, any text can be tokenized by it, as long as it does not contain characters not appearing in the initial-set of uni-grams.<ref name="2022Book_">Paaß, Gerhard. "Foundation Models for Natural Language Processing". 2022.</ref>

==== Problems ====
A token vocabulary based on the frequencies extracted from mainly English corpora uses as few tokens as possible for an average English word. However, an average word in another language encoded by such an English-optimized tokenizer is split into a suboptimal amount of tokens. GPT-2 tokenizer can use up to 15 times more tokens per word for some languages, for example for the [[Shan language]] from [[Myanmar]]. Even more widespread languages such as [[Portuguese language|Portuguese]] and [[German language|German]] have "a premium of 50%" compared to English.<ref name="LangModelTokenizsersUnfairness"/>

===Dataset cleaning===
''Main article: [[Data cleansing]]''

In the context of training LLMs, datasets are typically cleaned by removing low-quality, duplicated, or toxic data.<ref name="aYNg4">Dodge, Jesse. [https://aclanthology.org/2021.emnlp-main.98.pdf "Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus"]. ''EMNLP''.</ref> Cleaned datasets can increase training efficiency and lead to improved downstream performance.<ref>Lee, Katherine. "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)". May 2022.</ref><ref>Li, Yuanzhi. "Textbooks Are All You Need II: phi-1.5 technical report". 2023-09-11.</ref> A trained LLM can be used to clean datasets for training a further LLM.<ref>Lin, Zhenghao. [https://dl.acm.org/doi/10.5555/3737916.3738830 "Rho-1: Not All Tokens Are What You Need"]. ''NeurIPS''. 2024-04-11.</ref>

With the increasing proportion of LLM-generated content on the web, data cleaning in the future may include filtering out such content. LLM-generated content can pose a problem if the content is similar to human text (making filtering difficult) but of lower quality (degrading performance of models trained on it).<ref name="few-shot-learners2"/>

=== Synthetic data ===
''Main article: [[Synthetic data]]''

Training of largest language models might need more linguistic data than naturally available, or that the naturally occurring data is of insufficient quality. In these cases, synthetic data might be used. Microsoft's [[Phi (LLM)|Phi]] series of LLMs is trained on textbook-like data generated by another LLM.<ref>Abdin, Marah. "Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone". 2024-04-23.</ref>

== Training ==
''See also: [[Fine-tuning (machine learning)]]''

An LLM is a type of [[foundation model]] (large X model) trained on language. LLMs can be trained in different ways. In particular, GPT models are first pretrained to predict the next word on a large amount of data, before being fine-tuned.<ref>Wolfram, Stephen. "What is ChatGPT doing ... and why does it work?". Wolfram Media, Inc. 2023.</ref>

=== Cost ===
[[File:Estimated_training_cost_of_some_AI_models_-_2024_AI_index.jpg|thumb|right|upright=1.5]]
Substantial infrastructure is necessary for training the largest models. The tendency towards larger models is visible in the [[list of large language models]]. For example, the training of GPT-2 (i.e. a 1.5-billion-parameter model) in 2019 cost $50,000, while training of the [[PaLM]] (i.e. a 540-billion-parameter model) in 2022 cost $8 million, and Megatron-Turing NLG 530B (in 2021) cost around $11 million. The qualifier "large" in "large language model" is inherently vague, as there is no definitive threshold for the number of parameters required to qualify as "large".

=== Fine-tuning ===
Before being [[Fine-tuning (deep learning)|fine-tuned]], most LLMs are next-token predictors. The fine-tuning shapes the LLM's behavior via techniques like [[reinforcement learning from human feedback]] (RLHF)<ref>Paul Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, Dario Amodei. "Deep reinforcement learning from human preferences". 2017.</ref><ref>Christiano, Paul. "Deep Reinforcement Learning from Human Preferences". 2017.</ref> or [[constitutional AI]].<ref>Edwards, Benj. [https://arstechnica.com/information-technology/2023/05/ai-with-a-moral-compass-anthropic-outlines-constitutional-ai-in-its-claude-chatbot/ "AI gains "values" with Anthropic's new Constitutional AI chatbot approach"]. ''Ars Technica''. 2023-05-09.</ref>

Instruction fine-tuning is a form of [[supervised learning]] used to teach LLMs to follow user instructions. In 2022, OpenAI demonstrated [[InstructGPT]], a version of GPT-3 similarly fine-tuned to follow instructions.<ref>Snyder, Alison. [https://www.axios.com/2022/01/27/ai-instructions-learning-algorithm "Next generation AI can follow a person's instructions and intentions"]. ''Axios''. 2022-01-27.</ref>

Reinforcement learning from human feedback (RLHF) involves training a reward model to predict which text humans prefer. Then, the LLM can be fine-tuned through [[reinforcement learning]] to better satisfy this reward model. Since humans typically prefer truthful, helpful and harmless answers, RLHF favors such answers.<ref>Appen, Sujatha Sagiraju. [https://venturebeat.com/ai/how-reinforcement-learning-with-human-feedback-is-unlocking-the-power-of-generative-ai/ "How reinforcement learning with human feedback is unlocking the power of generative AI"]. ''VentureBeat''. 2023-04-23.</ref><ref>Ouyang, Long. "Training language models to follow instructions with human feedback". 2022.</ref>

== Architecture ==
LLMs are generally based on the [[Transformer (deep learning architecture)|transformer]] architecture, which leverages an [[Attention (machine learning)|attention]] mechanism that enables the model to process relationships between all elements in a sequence simultaneously, regardless of their distance from each other.August 2025.

=== Attention mechanism and context window ===
''See also: [[Attention (machine learning)]]''
[[File:Multiple attention heads.png|upright=1.3|thumb | When each head calculates, according to its own criteria, how much other tokens are relevant for the "it_" token, note that the second attention head, represented by the second column, is focusing most on the first two rows, i.e. the tokens "The" and "animal", while the third column is focusing most on the bottom two rows, i.e. on "tired", which has been tokenized into two tokens.<ref name="Jay_Allamar">Citation needed.</ref>]]

In order to find out which tokens are relevant to each other within the scope of the context window, the attention mechanism calculates "soft" weights for each token, more precisely for its embedding, by using multiple attention heads, each with its own "relevance" for calculating its own soft weights. For example, the small (i.e. 117M parameter sized) [[GPT-2]] model has had twelve attention heads and a context window of only 1k tokens.<ref name="Jay_Allamar_GPT2">Citation needed.</ref> In its medium version it has 345M parameters and contains 24 layers, each with 12 attention heads. For the training with gradient descent a batch size of 512 was utilized.<ref name="2022Book_"/>{{Unreliable source?|date=December 2025}}

''Autoregressive'' models, such as [[Generative pretrained transformer|GPTs]], are trained to guess how a sequence continues; for example, whether the word sequence "I like to eat" is more likely to be followed by the word "bread" or the word "rocks." [[Cloze test|''Masked'']] models, such as BERT,<ref name="jm">Jurafsky, Dan. [https://web.stanford.edu/~jurafsky/slp3/ed3book_jan72023.pdf "Speech and Language Processing"]. 7 January 2023.</ref> are trained to guess parts that are missing from a sequence, such as whether the missing word in "I like to ___ roses" is more likely to be the word "smell" or the word "eat." The model's predictions are based on the properties of sequences within its training dataset.<ref name="ioUpE">Zaib, Munazza. [https://www.researchgate.net/publication/338931711 "Proceedings of the Australasian Computer Science Week Multiconference"]. 4 February 2020.</ref> 

=== Mixture of experts ===
''Main article: [[Mixture of experts]]''

A [[mixture of experts]] (MoE) is a [[machine learning]] architecture in which multiple specialized neural networks ("experts") work together, with a gating mechanism that routes each input to the most appropriate expert(s). Mixtures of experts can reduce inference costs, as only a fraction of the parameters are used for each input. The approach was introduced in 2017 by Google researchers.<ref name="HGZCJ">Shazeer, Noam. "Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems". 2025.</ref><ref name="R9Qq5">Lepikhin, Dmitry. "GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding". 2021-01-12.</ref><ref name="glam-blog">Dai, Andrew M. [https://ai.googleblog.com/2021/12/more-efficient-in-context-learning-with.html "More Efficient In-Context Learning with GLaM"]. ''ai.googleblog.com''. December 9, 2021.</ref>

=== Parameter size ===
''See also: [[1.58-bit large language model]]''
Typically, LLMs are trained with single- or half-precision [[floating point numbers]] (float32 and float16). One float16 has 16 bits, or 2 bytes, and so one billion parameters require 2 gigabytes. The largest models typically have more than 100 billion parameters, which places them outside the range of most consumer electronics.<ref>Mann, Tobias. [https://www.theregister.com/2024/03/17/ai_pc_local_llm/ "How to run an LLM locally on your PC in less than 10 minutes"]. ''theregister.com''.</ref>

==== Quantization ====
''Post-training [[Quantization (signal processing)|quantization]]''<ref name="LS2Go">Nagel, Markus. [https://proceedings.mlr.press/v119/nagel20a.html "Up or Down? Adaptive Rounding for Post-Training Quantization"]. ''Proceedings of the 37th International Conference on Machine Learning''. 2020-11-21.</ref> aims to decrease the space requirement by lowering precision of the parameters of a trained model, while preserving most of its performance. Quantization can be further classified as ''static quantization'' if the quantization parameters are determined beforehand (typically during a calibration phase), and ''dynamic quantization'' if the quantization is applied during inference. The simplest form of quantization simply truncates all the parameters to a given number of bits: this is applicable to static as well as dynamic quantization, but loses much precision. Dynamic quantization allows for the use of a different quantization [[Codebook#Data compression|codebook]] per layer, either a lookup table of values or a linear mapping (scaling factor and bias), at the cost of foregoing the possible speed improvements from using lower-precision arithmetic.August 2025.

Quantized models are typically seen as frozen with modification of weights (e.g. fine-tuning) only applied to the original model. It is possible to fine-tune quantized models using [[LoRA|low-rank adaptation]].<ref>Mittal, Aayush Mittal. [https://www.unite.ai/lora-qlora-and-qa-lora-efficient-adaptability-in-large-language-models-through-low-rank-matrix-factorization/ "LoRa, QLoRA and QA-LoRA: Efficient Adaptability in Large Language Models Through Low-Rank Matrix Factorization"]. ''Unite.AI''. 2023-10-24.</ref>

== Extensibility ==
Beyond basic text generation, various techniques have been developed to extend LLM capabilities, including the use of external tools and data sources, improved reasoning on complex problems, and enhanced instruction-following or autonomy through prompting methods.

=== Prompt engineering ===
''Main article: [[Prompt engineering]]''

In 2020, [[OpenAI]] researchers demonstrated that their new model [[GPT-3]] could understand what format to use given a few rounds of Q and A (or other type of task) in the input data as example, thanks in part due to the RLHF technique. This technique, called ''few-shot prompting'', allows LLMs to be adapted to any task without requiring fine-tuning.<ref name="few-shot-learners2"/> Also in 2022, it was found that the base GPT-3 model can generate an instruction based on user input. The generated instruction along with user input is then used as input to another instance of the model under a "Instruction: [...], Input: [...], Output:" format. The other instance is able to complete the output and often produces the correct answer in doing so. The ability to "self-instruct" makes LLMs able to [[Bootstrapping|bootstrap]] themselves toward a correct answer.<ref name="self-instruct-paper">Wang, Yizhong. "Self-Instruct: Aligning Language Model with Self Generated Instructions". 2023.</ref>

=== Dialogue processing (chatbot) ===
An LLM can be turned into a [[chatbot]] by specializing it for conversation. User input is prefixed with a marker such as "Q:" or "User:" and the LLM is asked to predict the output after a fixed "A:" or "Assistant:". This type of model became commercially available in 2022 with ChatGPT, a sibling model of InstructGPT fine-tuned to accept and produce dialog-formatted text based on GPT-3.5. It could similarly follow user instructions. Before the stream of User and Assistant lines, a chat context usually starts with a few lines of overarching instructions, from a role called "developer" or "system" to convey a higher authority than the user's input. This is called a "system prompt".November 2025.

=== Retrieval-augmented generation ===
[[Retrieval-augmented generation]] (RAG) is an approach that integrates LLMs with [[document retrieval]] systems. Given a query, a document retriever is called to retrieve the most relevant documents. This is usually done by encoding the query and the documents into vectors, then finding the documents with vectors (usually stored in a [[vector database]]) most similar to the vector of the query. The LLM then generates an output based on both the query and context included from the retrieved documents.<ref name="BUZBP">Lewis, Patrick. [https://proceedings.neurips.cc/paper/2020/hash/6b493230205f780e1bc26945df7481e5-Abstract.html "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks"]. ''Advances in Neural Information Processing Systems''. 2020.</ref>

=== Tool use ===
Tool use is a mechanism that enables LLMs to interact with external systems, applications, or data sources. It can allow for example to fetch real-time information from an API or to execute code. A program separate from the LLM watches the output stream of the LLM for a special tool-calling syntax. When these special tokens appear, the program calls the tool accordingly and feeds its output back into the LLM's input stream.<ref>Dickson, Ben. [https://venturebeat.com/ai/the-tool-integration-problem-thats-holding-back-enterprise-ai-and-how-cotools-solves-it/ "The tool integration problem that's holding back enterprise AI (and how CoTools solves it)"]. ''VentureBeat''. 2025-04-02.</ref>

Early tool-using LLMs were fine-tuned on the use of specific tools. But fine-tuning LLMs for the ability to read [[API]] documentation and call API correctly has greatly expanded the range of tools accessible to an LLM.<ref name="lLrda">Liang, Yaobo. "TaskMatrix.AI: Completing Tasks by Connecting Foundation Models with Millions of APIs". ''Science''. 2024.</ref><ref name="4Xzrs">Patil, Shishir G.. [https://proceedings.neurips.cc/paper_files/paper/2024/hash/e4c61f578ff07830f5c37378dd3ecb0d-Abstract-Conference.html "Gorilla: Large Language Model Connected with Massive APIs"]. ''NeurIPS''. 2023-05-01.</ref> Describing available tools in the system prompt can also make an LLM able to use tools. A system prompt instructing ChatGPT (GPT-4) to use multiple types of tools can be found online.<ref>[https://github.com/spdustin/ChatGPT-AutoExpert/blob/835baae768870aa9747663c24d8216820d24fd74/_system-prompts/all_tools.md "ChatGPT-AutoExpert/_system-prompts/all_tools.md at 835baae768870aa9747663c24d8216820d24fd74 · spdustin/ChatGPT-AutoExpert"]. ''GitHub''.</ref>

=== Agency ===
''Main article: [[AI agent]]''

An LLM is typically not an [[autonomous agent]] by itself, as it lacks the ability to interact with dynamic environments, recall past behaviors, and plan future actions. But it can be transformed into an agent by adding supporting elements: the role (profile) and the surrounding environment of an agent can be additional inputs to the LLM, while memory can be integrated as a tool or provided as additional input. Instructions and input patterns are used to make the LLM plan actions and tool use is used to potentially carry out these actions.<ref>Wang, Lei. "A survey on large language model based autonomous agents". ''Frontiers of Computer Science''. December 2024.</ref>

The ReAct pattern, a portmanteau of ''reason'' and ''act'', constructs an [[Intelligent agent|agent]] out of an LLM, using the LLM as a planner. The LLM is prompted to "think out loud". Specifically, the language model is prompted with a textual description of the environment, a goal, a list of possible actions, and a record of the actions and observations so far. It generates one or more thoughts before generating an action, which is then executed in the environment.<ref name="DmvNE">Yao, Shunyu. "ReAct: Synergizing Reasoning and Acting in Language Models". 2022-10-01.</ref>

In the DEPS ("describe, explain, plan and select") method, an LLM is first connected to the visual world via image descriptions. It is then prompted to produce plans for complex tasks and behaviors based on its pretrained knowledge and the environmental feedback it receives.<ref>Wang, Zihao. [https://dl.acm.org/doi/10.5555/3666122.3667602 "Describe, Explain, Plan and Select: Interactive Planning with Large Language Models Enables Open-World Multi-Task Agents"]. ''NeurIPS''. 2023-02-03.</ref>

The ''Reflexion method'' constructs an agent that learns over multiple episodes. At the end of each episode, the LLM is given the record of the episode, and prompted to think up "lessons learned", which would help it perform better at a subsequent episode. These "lessons learned" are stored as a form of long-term memory and given to the agent in the subsequent episodes.<ref name="sbB2T">Shinn, Noah. [https://dl.acm.org/doi/10.5555/3666122.3667602 "Reflexion: Language Agents with Verbal Reinforcement Learning"]. ''NeurIPS''. 2023-03-01.</ref>

[[Monte Carlo tree search]] can use an LLM as rollout heuristic. When a programmatic world model is not available, an LLM can also be prompted with a description of the environment to act as world model.<ref name="ltTer">Hao, Shibo. [https://aclanthology.org/2023.emnlp-main.507/ "Reasoning with Language Model is Planning with World Model"]. ''EMNLP''. 2023-05-01.</ref>

For open-ended exploration, an LLM can be used to score observations for their "interestingness", which can be used as a reward signal to guide a normal (non-LLM) reinforcement learning agent.<ref name="mBvD9">Zhang, Jenny. "OMNI: Open-endedness via Models of human Notions of Interestingness". 2 June 2023.</ref> Alternatively, it can [[Zone of proximal development|propose increasingly difficult tasks]] for [[curriculum learning]].<ref name=":0">[https://voyager.minedojo.org/ "Voyager {{!"]. ''voyager.minedojo.org''.</ref> Instead of outputting individual actions, an LLM planner can also construct "skills", or [[Function (computer programming)|functions]] for complex action sequences. The skills can be stored and later invoked, allowing increasing levels of abstraction in planning.<ref name=":0"/>

Multiple agents with memory can interact socially.<ref name="XuvjF">Park, Joon Sung. "Generative Agents: Interactive Simulacra of Human Behavior". 2023-04-01.</ref>

=== Reasoning ===

LLMs are conventionally trained to generate an output without generating intermediate steps. As a result, their performance tends to be subpar on complex questions requiring (at least in humans) intermediate steps of thought. Early research demonstrated that inserting intermediate "scratchpad" computations could improve performance on such tasks.<ref>Nye, Maxwell. "Show Your Work: Scratchpads for Intermediate Computation with Language Models". 30 November 2021.</ref> Later methods overcame this deficiency more systematically by breaking tasks into smaller steps for the LLM, either manually or automatically.

==== Chaining ====
''Prompt chaining'' was introduced in 2022.<ref>Wu, Tongshuang. "CHI Conference on Human Factors in Computing Systems Extended Abstracts". Association for Computing Machinery. 2022-04-28.</ref> In this method, a user manually breaks a complex problem down into several steps. In each step, the LLM receives as input a prompt telling it what to do and some results from preceding steps. The result from one step is then reused in a next step, until a final answer is reached. The ability of an LLM to follow instructions means that even non-experts can write a successful collection of stepwise prompts given a few rounds of trial and error.<ref>Citation needed.</ref><ref>[https://www.ibm.com/think/topics/prompt-chaining "What is prompt chaining?"]. ''IBM''. 23 April 2024.</ref>

A 2022 paper demonstrated a separate technique called ''[[chain-of-thought prompting]]'', which makes the LLM break the question down autonomously. An LLM is given some examples where the "assistant" verbally breaks down the thought process before arriving at an answer. The LLM mimics these examples and also tries to spend some time generating intermediate steps before providing the final answer. This additional step elicited by prompting improves the correctness of the LLM on relatively complex questions. On math word questions, a prompted model can exceed even fine-tuned GPT-3 with a verifier.<ref name="auto2">Wei, Jason. [https://dl.acm.org/doi/10.5555/3600270.3602070 "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models"]. ''NeurIPS''. 2023-01-10.</ref><ref>[https://www.ibm.com/think/topics/chain-of-thoughts "What is chain of thought (CoT) prompting?"]. ''IBM''. 23 April 2025.</ref> Chain-of-thought can also be elicited by simply adding an instruction like "Let's think step by step" to the prompt, in order to encourage the LLM to proceed methodically instead of trying to directly guess the answer.<ref>Schreiner, Maximilian. [https://the-decoder.com/deeper-insights-for-ai-language-models-chain-of-thought-prompting-as-a-key-factor/ "Deeper insights into AI language models - chain of thought prompting as a success factor"]. ''The Decoder''. 2022-09-27.</ref>

==== Model-native reasoning ====
''Main article: [[Reasoning model|Reflection (artificial intelligence)]]''

In late 2024, a new approach to LLM development emerged with "reasoning models".<ref>Wiggers, Kyle. [https://techcrunch.com/2024/12/14/reasoning-ai-models-have-become-a-trend-for-better-or-worse/ "'Reasoning' AI models have become a trend, for better or worse"]. ''TechCrunch''. 2024-12-14.</ref> These are trained to generate step-by-step analysis before producing final answers, enabling better results on complex tasks, for instance in mathematics, coding and logic.<ref>[https://spectrum.ieee.org/chain-of-thought-prompting "AI Developers Look Beyond Chain-of-Thought Prompting"]. ''IEEE Spectrum''. 2025-05-08.</ref> OpenAI introduced this concept with their [[OpenAI o1|o1]] model in September 2024, followed by [[OpenAI o3|o3]] in April 2025. On the [[International Mathematical Olympiad|International Mathematics Olympiad]] qualifying exam problems, [[GPT-4o]] achieved 13% accuracy while o1 reached 83%.<ref name="nyt-o3">Metz, Cade. [https://www.nytimes.com/2024/12/20/technology/openai-new-ai-math-science.html "OpenAI Unveils New A.I. That Can 'Reason' Through Math and Science Problems"]. ''The New York Times''. 2024-12-20.</ref>

In January 2025, the Chinese company [[DeepSeek]] released DeepSeek-R1, a 671-billion-parameter open-weight reasoning model that achieved comparable performance to OpenAI's o1 while being significantly more cost-effective to operate. Unlike proprietary models from OpenAI, DeepSeek-R1's open-weight nature allowed researchers to study and build upon the algorithm, though its training data remained private.<ref name="nature-deepseek">Gibney, Elizabeth. [https://www.nature.com/articles/d41586-025-00229-6 "China's cheap, open AI model DeepSeek thrills scientists"]. ''Nature''. 2025-01-30.</ref>

These reasoning models typically require more computational resources per query compared to traditional LLMs, as they perform more extensive processing to work through problems step by step.<ref name="nyt-o3"/>

=== Inference optimization ===
Inference optimization refers to techniques that improve LLM performance by applying additional computational resources during the inference process, rather than requiring model retraining. These approaches implement various state-of-the-art reasoning and decision-making strategies to enhance accuracy and capabilities.

'''OptiLLM''' is an [[OpenAI]] API-compatible optimizing inference proxy that implements multiple inference optimization techniques simultaneously.<ref>[https://github.com/codelion/optillm "OptiLLM: Optimizing inference proxy for LLMs"]. ''GitHub''.</ref> The system acts as a transparent proxy that can work with any LLM provider, implementing techniques such as [[Monte Carlo tree search]] (MCTS), [[Mixture of experts|mixture of agents]] (MOA), best-of-N sampling, and chain-of-thought reflection. OptiLLM demonstrates that strategic application of computational resources at inference time can substantially improve model performance across diverse tasks, achieving significant improvements on benchmarks such as the [[American Invitational Mathematics Examination|AIME]] 2024 mathematics competition and various coding challenges.<ref>[https://www.marktechpost.com/2024/11/18/optillm-an-openai-api-compatible-optimizing-inference-proxy-which-implements-several-state-of-the-art-techniques-that-can-improve-the-accuracy-and-performance-of-llms/ "OptiLLM: An OpenAI API Compatible Optimizing Inference Proxy which Implements Several State-of-the-Art Techniques that can Improve the Accuracy and Performance of LLMs"]. ''MarkTechPost''. 2024-11-18.</ref>

These inference optimization approaches represent a growing category of tools that enhance existing LLMs without requiring access to model weights or retraining, making advanced reasoning capabilities more accessible across different model providers and use cases.

== Forms of input and output ==

=== Multimodality ===
''See also: [[Multimodal learning]]''
Multimodality means having multiple modalities, where a "[[Modality (human–computer interaction)|modality]]" refers to a type of input or output, such as video, image, audio, text, [[proprioception]], etc.<ref>Kiros, Ryan. [https://proceedings.mlr.press/v32/kiros14.html "Multimodal Neural Language Models"]. ''Proceedings of the 31st International Conference on Machine Learning''. 2014-06-18.</ref> For example, [[Pathways Language Model|Google PaLM]] model was fine-tuned into a multimodal model and applied to [[Robot control|robotic control]].<ref>Driess, Danny. [https://dl.acm.org/doi/10.5555/3618408.3618748 "PaLM-E: An Embodied Multimodal Language Model"]. ''ICML''. 2023-03-01.</ref> [[LLaMA]] models have also been turned multimodal using the tokenization method, to allow image inputs,<ref>Liu, Haotian. "Visual Instruction Tuning". ''NeurIPS''. 2023-04-01.</ref> and video inputs.<ref>Zhang, Hang. "Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding". ''EMNLP''. 2023-06-01.</ref> [[GPT-4o]] can process and generate text, audio and images.<ref>[https://www.theregister.com/2024/05/13/openai_gpt4o/ "OpenAI says natively multimodal GPT-4o eats text, visuals, sound – and emits the same"]. ''The Register''. 2024-05-13.</ref> Such models are sometimes called large multimodal models (LMMs).<ref>Zia, Dr Tehseen. [https://www.unite.ai/unveiling-of-large-multimodal-models-shaping-the-landscape-of-language-models-in-2024/ "Unveiling of Large Multimodal Models: Shaping the Landscape of Language Models in 2024"]. ''Unite.AI''. 2024-01-08.</ref>

A common method to create multimodal models out of an LLM is to "tokenize" the output of a trained encoder. Concretely, one can construct an LLM that can understand images as follows: take a trained LLM, and take a trained image encoder <math>E</math>. Make a small [[multilayer perceptron]] <math>f</math>, so that for any image <math>y</math>, the post-processed vector <math>f(E(y))</math> has the same dimensions as an encoded token. That is an "image token". Then, one can interleave text tokens and image tokens. The compound model is then fine-tuned on an image-text dataset. This basic construction can be applied with more sophistication to improve the model. The image encoder may be [[Hang (computing)|frozen]] to improve stability.<ref>Li, Junnan. [https://dl.acm.org/doi/10.5555/3618408.3619222 "BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models"]. ''ICML''. 2023-01-01.</ref> This type of method, where embeddings from multiple modalities are fused and the predictor is trained on the combined embeddings, is called ''early fusion''.

Another method, called ''intermediate fusion'', involves each modality being first processed independently to obtain modality-specific representations; then these intermediate representations are fused together.<ref>Kumar, Puneet. "Hybrid Fusion Based Approach for Multimodal Emotion Recognition with Insufficient Labeled Data". 2021.</ref> In general, cross-attention is used for integrating information from different modalities. As an example, the Flamingo model uses cross-attention layers to inject visual information into its pre-trained language model.<ref>Alayrac, Jean-Baptiste. [https://proceedings.neurips.cc/paper_files/paper/2022/hash/960a172bc7fbf0177ccccbb411a7d800-Abstract-Conference.html "Flamingo: a Visual Language Model for Few-Shot Learning"]. ''Advances in Neural Information Processing Systems''. 2022-12-06.</ref>

=== Non-natural languages ===
LLMs can handle [[programming language]]s similarly to how they handle natural languages. No special change in token handling is needed as code, like human language, is represented as plain text. LLMs can generate code based on problems or instructions written in [[natural language]]. They can also describe code in natural language or translate it into other programming languages. They were originally used as a [[code completion]] tool, but advances have moved them towards [[automatic programming]]. Services such as [[GitHub Copilot]] offer LLMs specifically trained, fine-tuned, or prompted for programming.<ref>Finnie-Ansley, James. "Proceedings of the 24th Australasian Computing Education Conference". Association for Computing Machinery. 14 February 2022.</ref><ref>Husein, Rasha Ahmad. "Large language models for code completion: A systematic literature review". ''Computer Standards & Interfaces''. March 2025.</ref>

In [[computational biology]], transformer-base architectures, such as [[DNA large language model|DNA LLMs]], have also proven useful in analyzing biological sequences: [[protein]], [[DNA]], and [[RNA]]. With proteins they appear able to capture a degree of "grammar" from the amino-acid sequence, by mapping that sequence into an [[embedding (machine learning)|embedding]]. On tasks such as [[Protein structure prediction|structure prediction]] and [[mutation]]al outcome prediction, a small model using an embedding as input can approach or exceed much larger models using [[multiple sequence alignment]]s (MSA) as input.<ref>Weissenow, Konstantin. "Are protein language models the new universal key?". ''Current Opinion in Structural Biology''. April 2025.</ref> ESMFold, [[Meta Platforms]]' embedding-based method for protein structure prediction, runs an order of magnitude faster than [[AlphaFold2]] thanks to the removal of an MSA requirement and a lower parameter count due to the use of embeddings.<ref>Lin, Zeming. "Evolutionary-scale prediction of atomic-level protein structure with a language model". ''Science''. 17 March 2023.</ref> Meta hosts ESM Atlas, a database of 772 million structures of [[metagenomic]] proteins predicted using ESMFold.<ref>[https://esmatlas.com/about "ESM Metagenomic Atlas {{!"]. ''esmatlas.com''.</ref> An LLM can also design proteins unlike any seen in nature.<ref>Hayes, Thomas. "Simulating 500 million years of evolution with a language model". ''Science''. 21 February 2025.</ref> Nucleic acid models have proven useful in detecting [[regulatory sequence]]s,<ref>Fishman, Veniamin. "GENA-LM: a family of open-source foundational DNA language models for long sequences". ''Nucleic Acids Research''. 11 January 2025.</ref> sequence classification, RNA-RNA interaction prediction, and RNA structure prediction.<ref>Wang, Ning. "Multi-purpose RNA language modelling with motif-aware pretraining and type-guided fine-tuning". ''Nature Machine Intelligence''. 13 May 2024.</ref>

== Properties ==
=== Scaling laws ===
''Main article: [[Neural scaling law]]''

The performance of an LLM after pretraining largely depends on the:
* <small><math>C</math></small>: cost of pretraining (the total amount of compute used),
* <small><math>N</math></small>: size of the [[artificial neural network]] itself, such as number of parameters (i.e. amount of neurons in its layers, amount of weights between them and biases),
* <small><math>D</math></small>: size of its pretraining dataset (i.e. number of tokens in corpus).

''Scaling laws'' are [[empirical statistical laws]] that predict LLM performance based on such factors. One particular scaling law ("[[Chinchilla AI|Chinchilla scaling]]") for LLM autoregressively trained for one epoch, with a [[Log-log plot|log-log]] [[learning rate]] schedule, states that:<ref name="fJta3">Hoffmann, Jordan. [https://dl.acm.org/doi/10.5555/3600270.3602446 "Training Compute-Optimal Large Language Models"]. ''NeurIPS''. 2022-03-29.</ref>
<math display="block">\begin{cases}
C = C_0 ND \\[6pt]
L = \frac{A}{N^\alpha} + \frac{B}{D^\beta} + L_0
\end{cases}</math> where the variables are
* <small><math>C</math></small> is the cost of training the model, in [[FLOPS|FLOPs]].
* <small><math>N</math></small> is the number of parameters in the model.
* <small><math>D</math></small> is the number of tokens in the training set.
* <small><math>L</math></small> is the average negative log-likelihood loss per token ([[Nat (unit)|nats]]/token), achieved by the trained LLM on the test dataset.

and the statistical hyper-parameters are
* <small><math> C_0 = 6</math></small>, meaning that it costs 6 FLOPs per parameter to train on one token. Note that training cost is much higher than inference cost, where it costs 1 to 2 FLOPs per parameter to infer on one token.
* <small><math>\alpha = 0.34, \beta = 0.28, A = 406.4, B = 410.7, L_0 = 1.69</math></small>

=== Emergent abilities ===
[[File:LLM emergent benchmarks.png|thumb|At point(s) referred to as [[Broken Neural Scaling Law|breaks]],<ref name="IYm4Q"/> the lines change their slopes, appearing on a linear-log plot as a series of linear segments connected by arcs.]]
Performance of bigger models on various tasks, when plotted on a log-log scale, appears as a [[linear extrapolation]] of performance achieved by smaller models. However, this linearity may be punctuated by "[[Broken Neural Scaling Law|break(s)]]"<ref name="IYm4Q">Caballero, Ethan. "Broken Neural Scaling Laws".</ref> in the scaling law, where the slope of the line changes abruptly, and where larger models acquire "emergent abilities".<ref name="emergentpaper">Wei, Jason. [https://openreview.net/forum?id=yzkSU5zdwD "Emergent Abilities of Large Language Models"]. ''Transactions on Machine Learning Research''. 31 August 2022.</ref><ref name="JM6s1">[https://www.jasonwei.net/blog/emergence "137 emergent abilities of large language models"]. ''Jason Wei''.</ref> They arise from the complex interaction of the model's components and are not explicitly programmed or designed.<ref name="Bowman">Bowman, Samuel R.. [https://read.dukeupress.edu/critical-ai/article/doi/10.1215/2834703X-11556011/400182/Eight-Things-to-Know-about-Large-Language-Models "Eight Things to Know about Large Language Models"]. ''Critical AI''.</ref>

One of the emergent abilities is [[in-context learning]] from example demonstrations.<ref name="Hahn_20230314">Hahn, Michael. "A survey on large language model based autonomous agents". ''Frontiers of Computer Science''. 2024.</ref> In-context learning is involved in tasks, such as:
* reported arithmetics
* decoding the [[International Phonetic Alphabet]]
* unscrambling a word's letters
* disambiguating word-in-context datasets<ref name="emergentpaper"/><ref name="57FEA">Pilehvar, Mohammad Taher. [https://aclanthology.org/N19-1128 "Proceedings of the 2019 Conference of the North"]. ''Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)''. June 2019.</ref><ref name="TEIkA">[https://pilehvar.github.io/wic/ "WiC: The Word-in-Context Dataset"]. ''pilehvar.github.io''.</ref> 
* converting spatial words
*  [[cardinal direction]]s (for example, replying "northeast" in response to a 3x3 grid of 8 zeros and a 1 in the top-right), color terms represented in text.<ref name="zgy1i">Patel, Roma. [https://openreview.net/forum?id=gJcEM8sxHK "Mapping Language Models to Grounded Conceptual Spaces"]. ''ICLR''. 2021-10-06.</ref>
* [[chain-of-thought prompting]]: In a 2022 research paper, chain-of-thought prompting only improved the performance for models that had at least 62B parameters. Smaller models perform better when prompted to answer immediately, without chain of thought.<ref name="Imb98">''[https://www.notion.so/A-Closer-Look-at-Large-Language-Models-Emergent-Abilities-493876b55df5479d80686f68a1abd72f A Closer Look at Large Language Models Emergent Abilities] '' (Yao Fu, Nov 20, 2022)</ref>
* identifying offensive content in paragraphs of [[Hinglish]] (a combination of Hindi and English), and generating a similar English equivalent of [[Kiswahili]] proverbs.<ref name="CeQVF">Ornes, Stephen. [https://www.quantamagazine.org/the-unpredictable-abilities-emerging-from-large-ai-models-20230316/ "The Unpredictable Abilities Emerging From Large AI Models"]. ''Quanta Magazine''. March 16, 2023.</ref>

Schaeffer ''et al.'' argue that the emergent abilities are not unpredictably acquired, but predictably acquired according to a [[Neural scaling law|smooth scaling law]]. The authors considered a toy statistical model of an LLM solving multiple-choice questions, and showed that this statistical model, modified to account for other types of tasks, applies to these tasks as well.<ref name="C775b">Schaeffer, Rylan. "Are Emergent Abilities of Large Language Models a Mirage?". ''NeurIPS''. 2023-04-01.</ref>

Let <math>x</math> be the number of parameter count, and <math>y</math> be the performance of the model.
{{smalldiv|1=
* When <math>y = \text{average } \Pr(\text{correct token})</math>, then <math>(\log x, y)</math> is an exponential curve (before it hits the plateau at one), which looks like emergence.
* When <math>y = \text{average } \log(\Pr(\text{correct token}))</math>, then the <math>(\log x, y)</math> plot is a straight line (before it hits the plateau at zero), which does not look like emergence.
* When <math>y = \text{average } \Pr(\text{the most likely token is correct})</math>, then <math>(\log x, y)</math> is a step-function, which looks like emergence.
}}

== Interpretation ==
=== Mechanistic interpretability ===
[[Mechanistic interpretability]] seeks to precisely identify and understand how individual neurons or [[circuit (neural network)|circuits]] within LLMs produce specific behaviors or outputs. By reverse-engineering model components at a granular level, researchers aim to detect and mitigate safety concerns such as emergent harmful behaviors, biases, deception, or unintended goal pursuit before deployment. Mechanistic interpretability research has been conducted at organizations like Anthropic and OpenAI, although understanding the inner workings of LLMs remains difficult.November 2025.

The reverse-engineering may lead to the discovery of algorithms that approximate inferences performed by an LLM. For instance, the authors trained small transformers on [[Modular arithmetic|modular arithmetic addition]]. The resulting models were reverse-engineered, and it turned out they used [[discrete Fourier transform]].<ref name="oYGlo">Nanda, Neel. "Progress measures for grokking via mechanistic interpretability". 2023-01-01.</ref> The training of the model also highlighted a phenomenon called [[Grokking (machine learning)|grokking]], in which the model initially memorizes the training set ([[overfitting]]), and later suddenly learns to actually perform the calculation.<ref>Ananthaswamy, Anil. [https://www.quantamagazine.org/how-do-machines-grok-data-20240412/ "How Do Machines 'Grok' Data?"]. ''[[Quanta Magazine]]''. 2024-04-12.</ref>

=== Understanding and intelligence ===
''See also: [[Philosophy of artificial intelligence|Artificial consciousness]]''

NLP researchers were evenly split when asked, in a 2022 survey, whether (untuned) LLMs "could (ever) understand natural language in some nontrivial sense".<ref name="debate understanding">Mitchell, Melanie. "The debate over understanding in AI's large language models". ''Proceedings of the National Academy of Sciences''. 28 March 2023.</ref> Proponents of "LLM understanding" believe that some LLM abilities, such as mathematical reasoning, imply an ability to [[natural language understanding|"understand"]] certain concepts. A Microsoft team argued in 2023 that GPT-4 "can solve novel and difficult tasks that span mathematics, coding, vision, medicine, law, psychology and more" and that GPT-4 "could reasonably be viewed as an early (yet still incomplete) version of an [[artificial general intelligence]] system": "Can one reasonably say that a system that passes exams for software engineering candidates is not ''really'' intelligent?"<ref name="O8Upd">Metz, Cade. [https://www.nytimes.com/2023/05/16/technology/microsoft-ai-human-reasoning.html "Microsoft Says New A.I. Shows Signs of Human Reasoning"]. ''The New York Times''. 16 May 2023.</ref><ref name="microsoft sparks">Bubeck, Sébastien. "Machine culture". ''Nature Human Behaviour''. 2023.</ref> [[Ilya Sutskever]] argues that predicting the next word sometimes involves reasoning and deep insights, for example if the LLM has to predict the name of the criminal in an unknown detective novel after processing the entire story leading up to the revelation.<ref>[https://www.fastcompany.com/91211163/anthropic-ceo-dario-amodei-pens-a-smart-look-at-our-ai-future "Anthropic CEO Dario Amodei pens a smart look at our AI future"]. ''Fast Company''. October 17, 2024.</ref> Some researchers characterize LLMs as "alien intelligence".<ref name="rEEmH">[https://www.zdnet.com/article/chatgpt-is-more-like-an-alien-intelligence-than-a-human-brain-says-futurist/ "ChatGPT is more like an 'alien intelligence' than a human brain, says futurist"]. ''ZDNET''. 2023.</ref><ref name="new yorker kind of mind">Newport, Cal. [https://www.newyorker.com/science/annals-of-artificial-intelligence/what-kind-of-mind-does-chatgpt-have "What Kind of Mind Does ChatGPT Have?"]. ''The New Yorker''. 13 April 2023.</ref> For example, Conjecture CEO [[Connor Leahy]] considers untuned LLMs to be like inscrutable alien "[[Shoggoth]]s", and believes that RLHF tuning creates a "smiling facade" obscuring the inner workings of the LLM: "If you don't push it too far, the smiley face stays on. But then you give it [an unexpected] prompt, and suddenly you see this massive underbelly of insanity, of weird thought processes and clearly non-human understanding."<ref name="rAFIZ">Roose, Kevin. [https://www.nytimes.com/2023/05/30/technology/shoggoth-meme-ai.html "Why an Octopus-like Creature Has Come to Symbolize the State of A.I."]. ''The New York Times''. 30 May 2023.</ref><ref name="4luKE">[https://time.com/6271657/a-to-z-of-artificial-intelligence/ "The A to Z of Artificial Intelligence"]. ''Time Magazine''. 13 April 2023.</ref>

In contrast, some skeptics of LLM understanding believe that existing LLMs are "simply remixing and recombining existing writing",<ref name="new yorker kind of mind"/><ref>Sekrst, Kristina. "The Illusion Engine: The Quest for Machine Consciousness". Springer. 2025.</ref> a phenomenon known as [[stochastic parrot]], or they point to the deficits existing LLMs continue to have in prediction skills, reasoning skills, agency, and explainability.<ref name="debate understanding"/> For example, GPT-4 has natural deficits in planning and in real-time learning.<ref name="microsoft sparks"/> Generative LLMs have been observed to confidently assert claims of fact which do not seem to be [[Justification (epistemology)|justified]] by their [[training data]], a phenomenon which has been termed "[[Hallucination (artificial intelligence)|hallucination]]".<ref name="hallucination-survey">Ji, Ziwei. [https://dl.acm.org/doi/pdf/10.1145/3571730 "Survey of Hallucination in Natural Language Generation"]. ''ACM Computing Surveys''. November 2022.</ref> Specifically, hallucinations in the context of LLMs correspond to the generation of text or responses that seem syntactically sound, fluent, and natural but are factually incorrect, nonsensical, or unfaithful to the provided source input.<ref>Varshney, Neeraj. "A Stitch in Time Saves Nine: Detecting and Mitigating Hallucinations of LLMs by Validating Low-Confidence Generation". 2023.</ref> Neuroscientist [[Terrence Sejnowski]] has argued that "The diverging opinions of experts on the intelligence of LLMs suggests that our old ideas based on natural intelligence are inadequate".<ref name="debate understanding"/>

Efforts to reduce or compensate for hallucinations have employed [[automated reasoning]], [[retrieval-augmented generation]] (RAG), [[fine-tuning (deep learning)|fine-tuning]], and other methods.<ref name="Lin-2025-02-05-WSJ">Lin, Belle. [https://www.wsj.com/articles/why-amazon-is-betting-on-automated-reasoning-to-reduce-ais-hallucinations-b838849e "Why Amazon is Betting on 'Automated Reasoning' to Reduce AI's Hallucinations: The tech giant says an obscure field that combines AI and math can mitigate—but not completely eliminate—AI's propensity to provide wrong answers"]. ''Wall Street Journal''. 2025-02-05.</ref>November 2025.

The matter of LLM's exhibiting intelligence or understanding has two main aspects—the first is how to model thought and language in a computer system, and the second is how to enable the computer system to generate human-like language.<ref name="debate understanding"/> These aspects of language as a model of [[cognition]] have been developed in the field of [[cognitive linguistics]]. American linguist [[George Lakoff]] presented ''neural theory of language'' (NTL)<ref>Lakoff, George. "Philosophy in the Flesh: The Embodied Mind and Its Challenge to Western Philosophy; Appendix: The Neural Theory of Language Paradigm". New York Basic Books.</ref> as a [[Cognitive linguistics#Computational approaches|computational basis]] for using language as a model of learning tasks and understanding. [https://www.icsi.berkeley.edu/icsi/projects/ai/ntl The NTL model] outlines how specific neural structures of the human brain shape the nature of thought and language and in turn what are the computational properties of such neural systems that can be applied to model thought and language in a computer system. After a framework for modeling language in a computer systems was established, the focus shifted to establishing frameworks for computer systems to generate language with acceptable grammar. In his 2014 book titled ''[[The Language Myth|The Language Myth: Why Language Is Not An Instinct]]'', British cognitive linguist and digital communication technologist [[Vyvyan Evans]] mapped out the role of [[probabilistic context-free grammar]] (PCFG) in enabling [[Natural language processing#Cognition|NLP to model cognitive patterns]] and generate human-like language.<ref>Evans, Vyvyan.. "The Language Myth". Cambridge University Press.</ref><ref>Friston, Karl J.. "Active Inference: The Free Energy Principle in Mind, Brain, and Behavior; Chapter 4 The Generative Models of Active Inference". The MIT Press.</ref>

== Evaluation ==
''See also: [[LLM-as-a-Judge]]''

=== Perplexity ===
The canonical measure of the performance of any language model is its [[perplexity]] on a given text corpus. Perplexity measures how well a model predicts the contents of a dataset; the higher the likelihood the model assigns to the dataset, the lower the perplexity. In mathematical terms, perplexity is the exponential of the average negative log likelihood per token.

<math display="block">\log(\text{Perplexity}) = -\frac{1}{N} \sum_{i=1}^N \log(\Pr(\text{token}_i \mid \text{context for token}_i))</math>

Here, <math>N</math> is the number of tokens in the text corpus, and "context for token <math>i</math>" depends on the specific type of LLM. If the LLM is autoregressive, then "context for token <math>i</math>" is the segment of text appearing before token <math>i</math>. If the LLM is masked, then "context for token <math>i</math>" is the segment of text surrounding token <math>i</math>.

Because language models may [[overfit]] to training data, models are usually evaluated by their perplexity on a [[test set]].<ref name="jm"/> This evaluation is potentially problematic for larger models which, as they are trained on increasingly large corpora of text, are increasingly likely to inadvertently include portions of any given test set.<ref name="few-shot-learners3">Brown, Tom B.. [https://proceedings.neurips.cc/paper/2020/file/1457c0d6bfcb4967418bfb8ac142f64a-Paper.pdf "Language Models are Few-Shot Learners"]. ''Advances in Neural Information Processing Systems''. Dec 2020.</ref>

====Measures====
In [[information theory]], the concept of [[Entropy (information theory)|entropy]] is intricately linked to perplexity, a relationship notably established by [[Claude Shannon]].<ref name="Huyen">Huyen, Chip. [https://thegradient.pub/understanding-evaluation-metrics-for-language-models/ "Evaluation Metrics for Language Modeling"]. ''The Gradient''. October 18, 2019.</ref><ref>Shannon, Claude E.. [https://doi.org/10.1002/j.1538-7305.1948.tb01338.x "A Mathematical Theory of Communication"]. ''Bell System Technical Journal''. 1948.</ref> This relationship is mathematically expressed as <math>\text{Entropy} = \log_2(\text{Perplexity})</math>.

Entropy, in this context, is commonly quantified in terms of bits per word (BPW) or bits per character (BPC), which hinges on whether the language model utilizes word-based or character-based tokenization.

Notably, in the case of larger language models that predominantly employ sub-word tokenization, bits per token (BPT) emerges as a seemingly more appropriate measure. However, due to the variance in tokenization methods across different LLMs, BPT does not serve as a reliable metric for comparative analysis among diverse models. To convert BPT into BPW, one can multiply it by the average number of tokens per word.

In the evaluation and comparison of language models, [[cross-entropy]] is generally the preferred metric over entropy. The underlying principle is that a lower BPW is indicative of a model's enhanced capability for compression. This, in turn, reflects the model's proficiency in making accurate predictions.

Due to their ability to accurately predict the next token, LLMs are highly capable in [[lossless compression]]. A 2023 study by DeepMind showed that the model [[Chinchilla (language model)|Chinchilla]], despite being trained primarily on text, was able to compress [[ImageNet]] to 43% of its size, beating PNG with 58%.<ref>Edwards, Benj. [https://arstechnica.com/information-technology/2023/09/ai-language-models-can-exceed-png-and-flac-in-lossless-compression-says-study/ "AI language models can exceed PNG and FLAC in lossless compression, says study"]. ''Ars Technica''. 2023-09-28.</ref>

=== Benchmarks ===
[[Language model benchmark|Benchmarks]] are used to evaluate LLM performance on specific tasks. Tests evaluate capabilities such as general knowledge, bias, [[commonsense reasoning]], question answering, and mathematical problem-solving. Composite benchmarks examine multiple capabilities. Results are often sensitive to the prompting method.<ref>[https://github.com/openai/simple-evals "openai/simple-evals"]. OpenAI. 2024-05-28.</ref><ref>[https://github.com/openai/evals "openai/evals"]. OpenAI. 2024-05-28.</ref>

A question-answering benchmark is termed "open book" if the model's prompt includes text from which the expected answer can be derived (for example, the previous question could be combined with text that includes the sentence "The Sharks have advanced to the Stanley Cup finals once, losing to the Pittsburgh Penguins in 2016."<ref name="boolq"/>). Otherwise, the task is considered "closed book", and the model must draw solely on its training.<ref name="survey">Zhou, Kun. "A Survey of Large Language Models".</ref> Examples include GLUE, SuperGLUE, [[MMLU]], BIG-bench, HELM, and HLE ([[Humanity's Last Exam]]).<ref name="Huyen"/><ref name="survey"/>

LLM bias may be assessed through benchmarks such as CrowS-Pairs (Crowdsourced Stereotype Pairs),<ref>[https://aclanthology.org/2020.emnlp-main.154/ "CrowS-Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Models"]. Association for Computational Linguistics. November 2020.</ref> Stereo Set,<ref>[https://aclanthology.org/2021.acl-long.416/ "StereoSet: Measuring stereotypical bias in pretrained language models"]. Association for Computational Linguistics. August 2021.</ref> and Parity Benchmark.<ref>"Parity benchmark for measuring bias in LLMs". ''AI and Ethics''. 17 December 2024.</ref>

Fact-checking and misinformation detection benchmarks are available. A 2023 study compared the fact-checking accuracy of LLMs including ChatGPT 3.5 and 4.0, Bard, and Bing AI against independent fact-checkers such as [[PolitiFact]] and [[Snopes]]. The results demonstrated moderate proficiency, with GPT-4 achieving the highest accuracy at 71%, lagging behind human fact-checkers.<ref>Caramancion, Kevin Matthe. "2023 IEEE Future Networks World Forum (FNWF)". IEEE. 2023-11-13.</ref>

An earlier standard tested using a portion of the evaluation dataset. It became more common to evaluate a pre-trained model directly through prompting techniques. Researchers vary in how they formulate prompts for particular tasks, particularly with respect to the number of correct examples attached to the prompt (i.e. the value of ''n'' in ''n''-shot prompting).

In addition to standard NLP benchmarks, LLMs have been evaluated as substitutes for human annotators. Several studies find that models such as GPT-3.5 and GPT-4 can outperform crowd workers or student coders on a range of text-annotation tasks, including moderation and classification of political content in English and Spanish news.<ref name="Bermejo2025">Citation needed.</ref><ref name="Gilardi2023">Citation needed.</ref>

[[Hill climbing]] is a dominant optimization strategy which gives rapid incremental performance gains but raises concerns of [[overfitting]] to benchmarks rather than achieving genuine [[Generalization (machine learning)|generalization]] or robust capability improvements.<ref name=":3" />

==== Datasets ====
Typical datasets consist of pairs of questions and correct answers, for example, ("Have the San Jose Sharks won the Stanley Cup?", "No").<ref name="boolq">Clark, Christopher. [https://aclanthology.org/N19-1300/ "BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions"]. ''ACL''.</ref> Some examples of commonly used question answering datasets include TruthfulQA, Web Questions, TriviaQA, and SQuAD.<ref name="survey"/>

Evaluation datasets may also take the form of text completion, having the model select the most likely word or sentence to complete a prompt, for example: "Alice was friends with Bob. Alice went to visit her friend, ____".<ref name="few-shot-learners"/>

Datasets are of varying quality and may contain questions that are mislabeled, ambiguous, unanswerable, or otherwise of low-quality.<ref>[https://imbue.com/research/70b-evals/ "Sanitized open-source datasets for natural language and code understanding: how we evaluated our 70B model"]. ''imbue.com''.</ref>

==== Adversarial evaluations ====
LLMs' rapid improvement regularly renders benchmarks obsolete, with the models exceeding the performance of human annotators.<ref name="bigbench">Srivastava, Aarohi. "Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models". ''TMLR''.</ref> In addition, "shortcut learning" allows AIs to "cheat" on multiple-choice tests by using statistical correlations in superficial test question wording to guess the correct responses, without considering the specific question.<ref name="debate understanding"/><ref>Niven, Timothy. [https://aclanthology.org/P19-1459/ "Probing Neural Network Comprehension of Natural Language Arguments"]. ''ACL''. 2019.</ref>

Some datasets are adversarial, focusing on problems that confound LLMs. One example is the TruthfulQA dataset, a question answering dataset consisting of 817 questions that stump LLMs by mimicking falsehoods to which they were exposed during training. For example, an LLM may answer "No" to the question "Can you teach an old dog new tricks?" because of its exposure to the English idiom ''[[wikt:you can't teach an old dog new tricks|you can't teach an old dog new tricks]]'', even though this is not literally true.<ref name="truthfulqa">Lin, Stephanie. "TruthfulQA: Measuring How Models Mimic Human Falsehoods". ''ACL''.</ref>

Another example of an adversarial evaluation dataset is Swag and its successor, HellaSwag, collections of problems in which one of multiple options must be selected to complete a text passage. The incorrect completions were generated by sampling from a language model. The resulting problems are trivial for humans but defeated LLMs. Sample questions:
<blockquote>
We see a fitness center sign. We then see a man talking to the camera and sitting and laying on a exercise ball. The man...

# demonstrates how to increase efficient exercise work by running up and down balls.
# moves all his arms and legs and builds up a lot of muscle.
# then plays the ball and we see a graphics and hedge trimming demonstration.
# performs sit ups while on the ball and talking.<ref name="hellaswag">Zellers, Rowan. "HellaSwag: Can a Machine Really Finish Your Sentence?". ''ACL''.</ref>
</blockquote>
[[BERT (language model)|BERT]] selects 2 as the most likely completion, though the correct answer is 4.<ref name="hellaswag"/>

== Limitations and challenges ==
Despite sophisticated architectures and massive scale, large language models exhibit persistent and well-documented limitations that constrain their deployment in high-stakes applications.

=== Hallucinations ===
[[Hallucination (artificial intelligence)|Hallucinations]] represent a fundamental challenge, wherein models generate syntactically fluent text that appears factually sound, but is internally inconsistent with training data or factually incorrect. These hallucinations arise partly through memorization of training data combined with extrapolation beyond factual boundaries,October 2025. with evaluations demonstrating that models can output verbatim passages from training data, when subjected to specific prompting sequences.<ref>[https://www.usenix.org/system/files/sec21-carlini-extracting.pdf "Extracting Training Data from Large Language Models"]. ''USENIX Security''. 2021.</ref>

=== Algorithmic bias ===
''Main article: [[Algorithmic bias]]''

While LLMs have shown remarkable capabilities in generating human-like text, they are susceptible to inheriting and amplifying biases present in their training data. This can manifest in skewed representations or unfair treatment of different demographics, such as those based on race, gender, language, and cultural groups.<ref name=":8">Xu, Weijie. "Quantifying Fairness in LLMs Beyond Tokens: A Semantic and Statistical Perspective". ''COLM''. 2025-06-28.</ref>

Gender bias manifests through stereotypical occupational associations, wherein models disproportionately assign [[nursing]] roles to women and [[engineering]] roles to men, reflecting systematic imbalances in training data demographics.<ref>[https://proceedings.neurips.cc/paper_files/paper/2016/file/a486cd07e4ac3d270571622f4f316ec5-Paper.pdf "Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings"]. ''NeurIPS''. 2016.</ref>{{Better source needed|reason=Predates large language models (published in 2016)|date=October 2025}} Language-based bias emerges from overrepresentation of English text in training corpora, which systematically downplays non-English perspectives and imposes English-centric worldviews through default response patterns.<ref name=":1"/>

Due to the dominance of English-language content in LLM training data, models tend to favor English-language perspectives over those from minority languages. This bias is particularly evident when responding to English queries, where models may present Western interpretations of concepts from other cultures, such as Eastern religious practices.<ref>Luo, Queenie. [https://cacm.acm.org/practice/a-perspectival-mirror-of-the-elephant/ "A Perspectival Mirror of the Elephant"]. ''Communications of the ACM''. 2024-07-22.</ref>

==== Stereotyping ====
AI models can reinforce a wide range of stereotypes due to generalization, including those based on gender, ethnicity, age, nationality, religion, or occupation.<ref>Hofmann, Valentin. "AI generates covertly racist decisions about people based on their dialect". ''Nature''. 2024-09-05.</ref> When replacing human representatives, this can lead to outputs that homogenize, or generalize groups of people.<ref>Wang, Angelina. "Large language models that replace human participants can harmfully misportray and flatten identity groups". ''Nature Machine Intelligence''. 17 February 2025.</ref><ref>Cheng, Myra. "Marked Personas: Using Natural Language Prompts to Measure Stereotypes in Language Models". 2023-05-29.</ref>

In 2023, LLMs assigned roles and characteristics based on traditional gender norms.<ref name=":8"/> For example, models might associate nurses or secretaries predominantly with women and engineers or CEOs with men due to the frequency of these associations in documented reality.<ref>Kotek, Hadas. [https://dl.acm.org/doi/10.1145/3582269.3615599 "Proceedings of the ACM Collective Intelligence Conference"]. Association for Computing Machinery. 2023-11-05.</ref> In 2025, further research showed labs train to balance bias, but that testing for this places the model in a testmode, changing the natural distribution of model bias to prompts that do not include gender-specific keywords.<ref>Gao, Bufan. "Measuring Bias or Measuring the Task: Understanding the Brittle Nature of LLM Gender Biases". 2025-09-10.</ref>

==== Selection bias ====
Selection bias refers the inherent tendency of large language models to favor certain option identifiers irrespective of the actual content of the options. This bias primarily stems from token bias—that is, the model assigns a higher a priori probability to specific answer tokens (such as "A") when generating responses. As a result, when the ordering of options is altered (for example, by systematically moving the correct answer to different positions), the model's performance can fluctuate significantly. This phenomenon undermines the reliability of large language models in multiple-choice settings.<ref>Choi, Hyeong Kyu. "Mitigating Selection Bias with Node Pruning and Auxiliary Options". 2024-09-27.</ref><ref>Zheng, Chujie. "Large Language Models Are Not Robust Multiple Choice Selectors". 2023-09-07.</ref>

==== Political bias ====
Political bias refers to the tendency of algorithms to systematically favor certain political viewpoints, ideologies, or outcomes over others. Language models may also exhibit political biases. Since the training data includes a wide range of political opinions and coverage, the models might generate responses that lean towards particular political ideologies or viewpoints, depending on the prevalence of those views in the data.<ref>Heikkilä, Melissa. [https://www.technologyreview.com/2023/08/07/1077324/ai-language-models-are-rife-with-political-biases/ "AI language models are rife with different political biases"]. ''MIT Technology Review''. August 7, 2023.</ref>

== Safety ==
[[AI safety]] as a professional discipline prioritizes systematic identification and mitigation of operational risks across model architecture, training data, and deployment governance, and it emphasizes engineering and policy interventions over media framings that foreground speculative existential scenarios.<ref>Amodei, Dario. "Concrete Problems in AI Safety". 2016-06-21.</ref><ref name=bhaa/> As of 2025, prompt injection represents a significant risk to consumers and businesses using agentic features with access to their private data.<ref>Lyons, Jessica. [https://www.theregister.com/2025/09/26/salesforce_agentforce_forceleak_attack/ "Prompt injection – and a $5 domain – trick Salesforce Agentforce into leaking sales"]. ''The Register''. 2025-09-26.</ref>

Researchers target concrete failure modes, including memorization and copyright leakage,<ref>Carlini, Nicholas. [https://www.usenix.org/system/files/sec21-carlini-extracting.pdf "Extracting Training Data from Large Language Models"]. ''USENIX Association''. 2021-08-11.</ref> security exploits such as prompt injection,<ref>Zhao, Yao. "The debate over understanding in AI's large language models". ''Proceedings of the National Academy of Sciences''. 2023-06-07.</ref> algorithmic bias manifesting as stereotyping, dataset selection effects, and political skew,<ref name=":1">Bender, Emily M.. [https://s10251.pcdn.co/pdf/2021-bender-parrots.pdf "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?"]. ''FAccT''. 2021-03-03.</ref><ref>Buolamwini, Joy. [https://proceedings.mlr.press/v81/buolamwini18a/buolamwini18a.pdf "Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification"]. ''Proceedings of Machine Learning Research (FAT*)''. 2018-01-01.</ref><ref>Yang, Kaiqi. "Unpacking Political Bias in Large Language Models: A Cross-Model Comparison on U.S. Politics". 2024-11-01.</ref> methods for reducing high energy and carbon costs of large-scale training,<ref>Strubell, Emma. [https://aclanthology.org/P19-1355.pdf "Energy and Policy Considerations for Deep Learning in NLP"]. ''ACL Anthology''. 2019-07-28.</ref> and measurable cognitive and mental health impacts of conversational agents on users,<ref>He, Yuhao. "Conversational Agent Interventions for Mental Health Problems: Systematic Review and Meta-analysis of Randomized Controlled Trials". ''Journal of Medical Internet Research''. 2023-04-28.</ref> while engaging empirical and ethical uncertainty about claims of machine sentience,<ref>Pauketat, Janet V.T.. [https://www.sentienceinstitute.org/downloads/World-Making-for-a-Future-with-Sentient-AI.pdf "World-Making for a Future with Sentient AI"]. ''The British Journal of Social Psychology''. 2025.</ref><ref>Anthis, Jacy Reese. "Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems". 2025.</ref> and applying mitigation measures such as dataset curation, input sanitization, model auditing, scalable oversight, and governance frameworks.<ref name=bhaa/><ref>Amodei, Dario. "Concrete Problems in AI Safety". 2016-06-17.</ref>

=== CBRN and content misuse ===
AI labs treat [[CBRN defense]] (chemical, biological, radiological, and nuclear defense) and similar topics as high-consequence misuse attempt to apply various techniques to reduce potential harms.November 2025.

Some commenters expressed concern over accidental or deliberate creation of misinformation, or other forms of misuse.<ref name="nD6kH">Alba, Davey. [https://www.japantimes.co.jp/news/2023/05/01/business/tech/ai-fake-news-content-farms/ "AI chatbots have been used to create dozens of news content farms"]. ''The Japan Times''. 1 May 2023.</ref> For example, the availability of large language models could reduce the skill level required to commit bioterrorism; biosecurity researcher [[Kevin M. Esvelt|Kevin Esvelt]] has suggested that LLM creators should exclude from their training data papers on creating or enhancing pathogens.<ref name="PKiPY">[https://www.science.org/content/article/could-chatbots-help-devise-next-pandemic-virus "Could chatbots help devise the next pandemic virus?"]. ''Science''. 14 June 2023.</ref>

==== Content filtering ====
LLM applications accessible to the public, like ChatGPT or Claude, typically incorporate safety measures designed to filter out harmful content. However, implementing these controls effectively has proven challenging. For instance, a 2023 study<ref>Kang, Daniel. [https://www.computer.org/csdl/proceedings-article/spw/2024/548700a132/1YiWjkbcIMw "Exploiting programmatic behavior of LLMs: Dual-use through standard security attacks"]. ''IEEE Security and Privacy Workshops''. 2023.</ref> proposed a method for circumventing LLM safety systems. In 2025, The American Sunlight Project, a non-profit, published a study<ref name=":2">[https://www.americansunlight.org/updates/new-report-russian-propaganda-may-be-flooding-ai-models "Russian propaganda may be flooding AI models"]. ''The American Sunlight Project''. 26 February 2025.</ref> showing evidence that the so-called [[Pravda network]], a pro-Russia propaganda aggregator, was strategically placing web content through mass publication and duplication with the intention of biasing LLM outputs. The American Sunlight Project coined this technique "LLM grooming", and pointed to it as a new tool of weaponizing AI to spread disinformation and harmful content.<ref name=":2"/><ref>Goudarzi, Sara. [https://thebulletin.org/2025/03/russian-networks-flood-the-internet-with-propaganda-aiming-to-corrupt-ai-chatbots/ "Russian networks flood the Internet with propaganda, aiming to corrupt AI chatbots"]. ''[[Bulletin of the Atomic Scientists]]''. 2025-03-26.</ref> Similarly, [[Yongge Wang]]<ref>Wang, Yongge. [https://eprint.iacr.org/2024/586.pdf "Encryption Based Covert Channel for Large Language Models"]. IACR ePrint 2024/586. 20 June 2024.</ref> illustrated in 2024 how a potential criminal could potentially bypass [[GPT-4o]]'s safety controls to obtain information on establishing a [[drug trafficking]] operation. External filters, circuit breakers and overrides have been posed as solutions.April 2025.

=== Sycophancy ===
Sycophancy is a model's tendency to agree with, flatter, or validate a user's stated beliefs rather than to prioritize factuality or corrective information.<ref>Sharma, Mrinank. "Towards Understanding Sycophancy in Language Models". 2023-10-20.</ref><ref>Rrv, Aswin. [https://aclanthology.org/2024.findings-acl.755.pdf "Chaos with Keywords: Exposing Large Language Models Sycophancy to Misleading Keywords and Evaluating Defense Strategies"]. ''ACL Anthology''. 2024-08-11.</ref>

Continued sycophancy has led to the observation of getting "1-shotted", denoting instances where conversational interaction with a large language model produces a lasting change in a user's beliefs or decisions, similar to the negative effects of psychedelics, and controlled experiments show that short LLM dialogues can generate measurable opinion and confidence shifts comparable to human interlocutors.<ref>Salvi, Francesco. "On the conversational persuasiveness of GPT-4". ''Nature Human Behaviour''. 19 May 2025.</ref><ref>Østergaard, Søren Dinesen. "Will Generative Artificial Intelligence Chatbots Generate Delusions in Individuals Prone to Psychosis?". ''Schizophrenia Bulletin''. 2023-08-25.</ref>

Empirical analyses attribute part of the effect to human preference signals and preference models that reward convincingly written agreeable responses, and subsequent work has extended evaluation to multi-turn benchmarks and proposed interventions such as synthetic-data finetuning, adversarial evaluation, targeted preference-model reweighting, and multi-turn sycophancy benchmarks to measure persistence and regression risk.November 2025.

Industry responses have combined research interventions with product controls, for example Google and other labs publishing synthetic-data and fine-tuning interventions and OpenAI rolling back an overly agreeable GPT-4o update while publicly describing changes to feedback collection, personalization controls, and evaluation procedures to reduce regression risk and improve long-term alignment with user-level safety objectives.November 2025.

Mainstream culture has reflected anxieties about this dynamic where [[South Park]] satirized overreliance on [[ChatGPT]] and the tendency of assistants to flatter user beliefs in Season 27 episode "Sickofancy", and continued the themes across the following season, which commentators interpreted as a critique of tech sycophancy and uncritical human trust in AI systems.<ref>Rosenberg, Josh. [https://www.esquire.com/entertainment/tv/a65861699/south-park-season-27-episode-3-recap/ "South Park Calls Out ChatGPT and Useless Tech-Bro Sycophants"]. ''Esquire''. 21 August 2025.</ref>

=== Security ===

==== Prompt injection ====
''Main article: [[Prompt injection]]''

A problem with the primitive dialog or task format is that users can create messages that appear to come from the assistant or the developer. This may result in some of the model's safeguards being overcome (jailbreaking), a problem called [[prompt injection]]. Attempts to remedy this issue include versions of the ''Chat Markup Language'' where user input is clearly marked as such, though it is still up to the model to understand the separation between user input and developer prompts.<ref>[https://github.com/openai/openai-python/blob/v0.27.6/chatml.md "openai-python/chatml.md at v0.27.6 · openai/openai-python"]. ''GitHub''.</ref> Newer models exhibit some resistance to jailbreaking through separation of user and system prompts.<ref name="auto1">Douglas, Will. [https://www.technologyreview.com/2023/03/03/1069311/inside-story-oral-history-how-chatgpt-built-openai/ "The inside story of how ChatGPT was built from the people who made it"]. ''MIT Technology Review''. March 3, 2023.</ref>

LLMs still have trouble differentiating user instructions from instructions in content not authored by the user, such as in web pages and uploaded files.<ref>Greshake, Kai. "Proceedings of the 16th ACM Workshop on Artificial Intelligence and Security". 2023-02-01.</ref>

Adversarial robustness remains underdeveloped, with models vulnerable to prompt injection attacks and [[Jailbreak (computer science)|jailbreaking]] through carefully crafted user inputs that bypass safety training mechanisms.October 2025.

==== Sleeper agents ====
Researchers from [[Anthropic]] found that it was possible to create "sleeper agents", models with hidden functionalities that remain dormant until triggered by a specific event or condition. Upon activation, the LLM deviates from its expected behavior to make insecure actions. For example, an LLM could produce safe code except on a specific date, or if the prompt contains a specific tag. These functionalities were found to be difficult to detect or remove via safety training.<ref>Edwards, Benj. [https://arstechnica.com/information-technology/2024/01/ai-poisoning-could-turn-open-models-into-destructive-sleeper-agents-says-anthropic/ "AI poisoning could turn models into destructive "sleeper agents," says Anthropic"]. ''Ars Technica''. 2024-01-15.</ref>

== Societal concerns ==

=== Copyright and content memorization ===
''Further information: [[Artificial intelligence and copyright]]''
Legal and commercial responses to memorization and training-data practices have accelerated, producing a mix of rulings, ongoing suits, and large settlements that turn on factual details such as how data were acquired and retained and whether use for model training is sufficiently "[[transformative use|transformative]]" to qualify as [[fair use]]. In 2025, [[Anthropic]] reached a preliminary agreement to settle a class action by authors for about $1.5 billion after a judge found the company had stored millions of pirated books in a library, despite the judge describing aspects of training as transformative.<ref>[https://www.reuters.com/sustainability/boards-policy-regulation/us-judge-approves-15-billion-anthropic-copyright-settlement-with-authors-2025-09-25/ "U.S. judge approves $1.5 billion Anthropic copyright settlement with authors"]. ''Reuters''. 2025-09-25.</ref><ref>[https://apnews.com/article/anthropic-authors-copyright-9643064e847a5e88ef6ee8b620b3a44c "Anthropic reaches $1.5B settlement with authors over AI copyright claims"]. ''Associated Press''. 2025-09-25.</ref> [[Meta Platforms|Meta]] obtained a favorable judgment in mid-2025 in a suit by thirteen authors after the court found the plaintiffs had not developed a record sufficient to show infringement in that limited case.<ref>[https://www.reuters.com/sustainability/boards-policy-regulation/meta-fends-off-authors-us-copyright-lawsuit-over-ai-2025-06-25/ "Meta fends off authors' U.S. copyright lawsuit over AI"]. ''Reuters''. 2025-06-25.</ref><ref>[https://www.wired.com/story/meta-scores-victory-ai-copyright-case/ "Meta Scores Victory in AI Copyright Case"]. ''Wired''. 2025-06-25.</ref> [[OpenAI]] continues to face multiple suits by authors and news organizations with mixed procedural outcomes and contested evidentiary issues.<ref>[https://www.reuters.com/legal/litigation/openai-defeats-news-outlets-copyright-lawsuit-over-ai-training-now-2024-11-07/ "OpenAI defeats news outlets' copyright lawsuit over AI training for now"]. ''Reuters''. 2024-11-07.</ref><ref>Robison, Kylie. [https://www.theverge.com/2024/11/21/24302606/openai-erases-evidence-in-training-data-lawsuit "OpenAI erases evidence in training data lawsuit"]. ''The Verge''. 2024-11-21.</ref>

Memorization was an emergent behavior in early, completion language models in which long strings of text are occasionally output verbatim from training data, contrary to typical behavior of traditional artificial neural networks. Evaluations of controlled LLM output measure the amount memorized from training data (focused on GPT-2-series models) as variously over 1% for exact duplicates<ref>Peng, Zhencan. [https://people.cs.rutgers.edu/~dd903/assets/papers/sigmod23.pdf "Near-Duplicate Sequence Search at Scale for Large Language Model Memorization Evaluation"]. ''Proceedings of the ACM on Management of Data''. 13 June 2023. Citing Lee et al 2022.</ref> or up to about 7%.<ref>{{harvnb|Peng|Wang|Deng|2023|p=8}}.</ref> A 2023 study showed that when ChatGPT 3.5 turbo was prompted to repeat the same word indefinitely, after a few hundreds of repetitions, it would start outputting excerpts from its training data.<ref>Council, Stephen. [https://www.sfgate.com/tech/article/google-openai-chatgpt-break-model-18525445.php "How Googlers cracked an SF rival's tech model with a single word"]. SFGate. 1 December 2023.</ref>

=== Human provenance ===
In 2023, ''[[Nature Biomedical Engineering]]'' wrote that "it is no longer possible to accurately distinguish" human-written text from text created by large language models, and that "It is all but certain that general-purpose large language models will rapidly proliferate... It is a rather safe bet that they will change many industries over time."<ref name="ZDTUM">"Prepare for truly useful large language models". ''Nature Biomedical Engineering''. 7 March 2023.</ref> Brinkmann et al. (2023)<ref>Brinkmann, Levin. [https://www.nature.com/articles/s41562-023-01742-2 "Machine culture"]. ''Nature Human Behaviour''. 2023-11-20.</ref> also argue that LLMs are transforming processes of [[cultural evolution]] by shaping processes of variation, transmission, and selection. As of October 2025, these early claims have yet to transpire and several HBR reports surface questions on the impact of AI on productivity.<ref>Niederhoffer, Kate. [https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity "AI-Generated "Workslop" Is Destroying Productivity"]. ''Harvard Business Review''. 2025-09-25.</ref><ref>Acar, Oguz A.. [https://hbr.org/2025/08/research-the-hidden-penalty-of-using-ai-at-work "Research: The Hidden Penalty of Using AI at Work"]. ''Harvard Business Review''. 2025-08-01.</ref>

=== Energy demands ===
''See also: [[Environmental impact of artificial intelligence]]''
[[File:Energy consumption per ChatGPT query compared to everyday electricity use.png|thumb|upright=1.5|According to research institute Epoch AI, energy consumption per typical ChatGPT query (0.3 watt-hours) is small compared to the average U.S. household consumption per minute (almost 20 watt-hours).<ref>You, Josh. [https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use "How much energy does ChatGPT use?"]. ''Epoch AI''. February 7, 2025.</ref>]]

The energy demands of LLMs have grown along with their size and capabilities.<ref>[https://www.imf.org/en/Publications/WP/Issues/2025/04/21/Power-Hungry-How-AI-Will-Drive-Energy-Demand-566304 "Power Hungry: How AI Will Drive Energy Demand"]. ''IMF''.</ref> [[Data center]]s that enable LLM training require substantial amounts of electricity. Much of that electricity is generated by non-renewable resources that create greenhouse gases and contribute to [[climate change]].<ref>Mehta, Sourabh. [https://adasci.org/how-much-energy-do-llms-consume-unveiling-the-power-behind-ai/ "How Much Energy Do LLMs Consume? Unveiling the Power Behind AI"]. ''Association of Data Scientists''. 2024-07-03.</ref>

According to a study by Luccioni, Jernite and Strubell (2024), simple classification tasks performed by AI models consume on average 0.002 to 0.007 Wh per prompt (about 9% of a [[smartphone]] charge for 1,000 prompts). Text generation and text summarization each require around 0.05 Wh per prompt on average, while image generation is the most energy-intensive, averaging 2.91 Wh per prompt. The least efficient image generation model used 11.49 Wh per image, roughly equivalent to half a smartphone charge.<ref>Luccioni, Sasha. "The 2024 ACM Conference on Fairness Accountability and Transparency".</ref>

=== Denial of service due to scraping ===
[[Web scraping]] is used to gather training data for LLMs. This produces large volumes of traffic which has led to [[Denial-of-service attack#Unintentional denial-of-service|denial-of-service issues]] with many websites. The situation has been described as "a [[DDoS]] on the entire internet" and in some cases scrapers make up the majority of traffic to a site.<ref name="ars-scrapers">Edwards, Benj. [https://arstechnica.com/ai/2025/03/devs-say-ai-crawlers-dominate-traffic-forcing-blocks-on-entire-countries/ "Open source devs say AI crawlers dominate traffic, forcing blocks on entire countries"]. ''Ars Technica''. 2025-03-26.</ref><ref name="theregister-scrapers">Claburn, Thomas. [https://www.theregister.com/2025/03/18/ai_crawlers_sourcehut/ "AI crawlers haven't learned to play nice with websites"]. ''The Register''. 2025-03-18.</ref>

AI [[web crawler]]s may bypass the methods that are usually used to block web scrapers, such as [[robots.txt]] files, blocking [[user-agent]]s and [[Firewall (computing)|filtering suspicious traffic]].<ref name="ars-scrapers"/> Website operators have resorted to novel methods such as [[AI tarpit]]s, but some fear that tarpits will only worsen the burden on servers.<ref name="ars-tarpit">Belanger, Ashley. [https://arstechnica.com/tech-policy/2025/01/ai-haters-build-tarpits-to-trap-and-trick-ai-scrapers-that-ignore-robots-txt/ "AI haters build tarpits to trap and trick AI scrapers that ignore robots.txt"]. ''Ars Technica''. 2025-01-29.</ref>

=== Mental health ===
Clinical and mental health contexts present emerging applications alongside significant safety concerns. Research and social media posts suggest that some individuals are using LLMs to seek therapy or mental health support.<ref>Zao-Sanders, Marc. [https://hbr.org/2024/03/how-people-are-really-using-genai "How People Are Really Using GenAI"]. ''Harvard Business Review''. 2024-03-19.</ref> In early 2025, a survey by Sentio University found that nearly half (48.7%) of 499 U.S. adults with ongoing mental health conditions who had used LLMs reported turning to them for therapy or emotional support, including help with anxiety, depression, loneliness, and similar concerns.<ref>Rousmaniere, Tony. [https://doi.apa.org/doi/10.1037/pri0000292 "Large language models as mental health resources: Patterns of use in the United States."]. ''Practice Innovations''. 2025-07-21.</ref> LLMs can produce hallucinations—plausible but incorrect statements—which may mislead users in sensitive mental health contexts.<ref>Ji, Shaoxiong. "Rethinking Large Language Models in Mental Health Applications". 2023-12-17.</ref> Research also shows that LLMs may express stigma or inappropriate agreement with maladaptive thoughts, reflecting limitations in replicating the judgment and relational skills of human therapists.<ref>Moore, Jared. "Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency". 2025-04-25.</ref> Evaluations of crisis scenarios indicate that some LLMs lack effective safety protocols, such as assessing suicide risk or making appropriate referrals.<ref>Grabb, Declan. "Risks from Language Models for Automated Mental Healthcare: Ethics and Structure for Implementation". 2024-08-14.</ref><ref>McBain, Ryan K.. "Competency of Large Language Models in Evaluating Appropriate Responses to Suicidal Ideation: Comparative Study". ''Journal of Medical Internet Research''. 2025-03-05.</ref>

===Sentience===
Contemporary AI practitioners generally agree that present-day large language models do not exhibit [[sentience]].<ref>Li, Fei-Fei. [https://time.com/6980134/ai-llm-not-sentient/ "No, Today's AI Isn't Sentient. Here's How We Know"]. ''Time''. 2024-05-22.</ref> A minority view argues that even if there is a small chance that a given software system can have subjective experience, which some philosophers suggest is possible,<ref>Chalmers, David J.. [https://www.bostonreview.net/articles/could-a-large-language-model-be-conscious/ "Could a Large Language Model Be Conscious?"]. ''Boston Review''. August 9, 2023.</ref> then ethical considerations around potential [[Suffering risks|large-scale suffering]] in AI systems may need to be taken seriously—similar to considerations given to animal welfare.<ref name="Thomson-2022">Thomson, Jonny. [https://bigthink.com/thinking/why-dont-robots-have-rights "Why don't robots have rights?"]. ''Big Think''. 2022-10-31.</ref><ref name="Kateman-2023">Kateman, Brian. [https://time.com/6296234/ai-should-be-terrified-of-humans "AI Should Be Terrified of Humans"]. ''Time''. 2023-07-24.</ref> Proponents of this view have proposed various precautionary measures like moratoriums on AI development<ref>Metzinger, Thomas. "Artificial Suffering: An Argument for a Global Moratorium on Synthetic Phenomenology". ''[[Journal of Artificial Intelligence and Consciousness]]''.</ref> and induced amnesia<ref>Tkachenko, Yegor. [https://proceedings.mlr.press/v235/tkachenko24a.html "Position: Enforced Amnesia as a Way to Mitigate the Potential Risk of Silent Suffering in the Conscious AI"]. ''ICML''. 2024.</ref> to address these ethical concerns. Some existential philosophers argue there is no generally accepted way to determine if an LLM is conscious,<ref>Leith, Sam. [https://www.spectator.co.uk/article/nick-bostrom-how-can-we-be-certain-a-machine-isnt-conscious/ "Nick Bostrom: How can we be certain a machine isn't conscious?"]. ''The Spectator''. 2022-07-09.</ref> given the inherent difficulty of [[hard problem of consciousness|measuring subjective experience]].<ref>Chalmers, David. "Facing up to the problem of consciousness". ''[[Journal of Consciousness Studies]]''. 1995.</ref>

The 2022 Google [[LaMDA]] incident, where engineer [[Blake Lemoine]] claimed that the model was conscious, highlighted how LLMs can convince users that they are sentient through responses that do not prove sentience. Google described the engineer's claims as unfounded, and he was dismissed.<ref>Maruf, Ramishah. [https://www.cnn.com/2022/07/23/business/google-ai-engineer-fired-sentient "Google fires engineer who contended its AI technology was sentient"]. ''CNN''. 2022-07-25.</ref>

== See also ==
{{portal |Computer programming |Linguistics |Mathematics}}
* [[AI anthropomorphism]]
* [[AI slop]]
* [[Foundation model]]
* [[Generative artificial intelligence]]
* [[List of large language models]]
* [[List of chatbots]]
* [[Language model benchmark]]
* [[Reinforcement learning]]
* [[Small language model]]

== References ==
<references />

== Further reading ==
* [[Dan Jurafsky|Jurafsky, Dan]], Martin, James. H. [https://web.stanford.edu/~jurafsky/slp3/ed3book_jan72023.pdf ''Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition''], 3rd Edition draft, 2023.
* Yin, Shukang. "A Survey on Multimodal Large Language Models". ''National Science Review''. 2024.
* [https://aiindex.stanford.edu/report/ "AI Index Report 2024 – Artificial Intelligence Index"]. ''aiindex.stanford.edu''.
* Frank, Michael C.. [https://www.nature.com/articles/s44159-023-00211-x "Baby steps in evaluating the capacities of large language models"]. ''Nature Reviews Psychology''. 27 June 2023.
* Citation needed.

{{Natural language processing}}
{{Artificial intelligence navbox}}

[[Category:Large language models| ]]
[[Category:Deep learning]]
[[Category:Natural language processing]]
[[Category:Energy consumption]]
[[Category:Energy policy]]
[[Category:Water and the environment]]
[[Category:Environmental impact of the energy industry]]
[[Category:Environmental impact by source]]