Do large language models have a legal duty to tell the truth?
Open accessThe DOI matches its Crossref record; 50 of the 50 references it cites were found in a registry or an index.
Abstract
Careless speech is a new type of harm created by large language models (LLM) that poses cumulative, long-term risks to science, education and shared social truth in democratic societies. LLMs produce responses that are plausible, helpful and confident, but that contain factual inaccuracies, misleading references and biased information. These subtle mistruths are poised to cumulatively degrade and homogenize knowledge over time. This article examines the existence and feasibility of a legal duty for LLM providers to create models that ‘tell the truth’. We argue that LLM providers should be required to mitigate careless speech and better align their models with truth through open, democratic processes. We define careless speech against ‘ground truth’ in LLMs and related risks including hallucinations, misinformation and disinformation. We assess the existence of truth-related obligations in EU human rights law and the Artificial Intelligence Act, Digital Services Act, Product Liability Directive and Artificial Intelligence Liability Directive. Current frameworks contain limited, sector-specific truth duties. Drawing on duties in science and academia, education, archives and libraries, and a German case in which Google was held liable for defamation caused by autocomplete, we propose a pathway to create a legal truth duty for providers of narrow- and general-purpose LLMs.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021) · Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency · doi:10.1145/3442188.3445922 · Cited by 6,988
Training language models to follow instructions with human feedback
Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. (2022) · arXiv · doi:10.48550/arxiv.2203.02155 · Cited by 4,208
An Experimental Study of Apparent Behavior
Heider, F., & Simmel, M. (1944) · The American Journal of Psychology · doi:10.2307/1416950 · Cited by 3,340
The Logic of Scientific Discovery
Popper, K. R., & Weiss, G. (1959) · Physics Today · doi:10.1063/1.3060577 · Cited by 2,717
Nass, C., Steuer, J., & Tauber, E. R. (1994) · Proceedings of the SIGCHI Conference on Human Factors in Computing Systems · doi:10.1145/191666.191703 · Cited by 2,148
Principles alone cannot guarantee ethical AI
Mittelstadt, B. (2019) · Nature Machine Intelligence · doi:10.1038/s42256-019-0114-4 · Cited by 1,886 · cited 2×
“Where duties for users exist, it may be possible to replicate or transfer relevant obligations from users to the providers of narrow-or general-purpose AI systems they use [117].”
BackgroundSelf-citationLanguage Models as Knowledge Bases?
Petroni, F., Rocktäschel, T., Riedel, S., Lewis, P., Bakhtin, A., Wu, Y., & Miller, A. (2019) · Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) · doi:10.18653/v1/d19-1250 · Cited by 1,834
Wachter, S., Mittelstadt, B., & Floridi, L. (2017) · International Data Privacy Law · doi:10.1093/idpl/ipx005 · Cited by 1,308
Self-citationFrankfurt, H. G. (2005) · doi:10.1515/9781400826537 · Cited by 1,175
Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data
Bender, E. M., & Koller, A. (2020) · Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/2020.acl-main.463 · Cited by 1,082 · cited 15×
“Since many students and teachers will use LLMs, it can be argued that providers of general purpose models might be offering an educational service that would be subject to these requirements, or fit the definition of a teacher or educational institution.”
BackgroundInfluentialChatGPT for Education and Research: Opportunities, Threats, and Strategies
Rahman, M. M., & Watanobe, Y. (2023) · Applied Sciences · doi:10.3390/app13095783 · Cited by 1,080
TruthfulQA: Measuring How Models Mimic Human Falsehoods
Lin, S., Hilton, J., & Evans, O. (2022) · Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/2022.acl-long.229 · Cited by 904
TruthfulQA: Measuring How Models Mimic Human Falsehoods
Lin, S., Hilton, J., & Evans, O. (2021) · arXiv · doi:10.48550/arxiv.2109.07958 · Cited by 899
Bender, E. M., & Friedman, B. (2018) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00041 · Cited by 854 · cited 1×
“Instead, reliability can be improved through extensive curation and annotation [50], methodologically sound benchmarking metrics [51,52], long-term fine tuning with expert human feedback [53], auditing and adversarial testing ([5], p. 11, [53], p. 60), and perhaps even downsizing models [17, pp.…”
BackgroundAI models collapse when trained on recursively generated data
Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., & Gal, Y. (2024) · Nature · doi:10.1038/s41586-024-07566-y · Cited by 793
LaMDA: Language Models for Dialog Applications
Thoppilan, R., De Freitas, D., Hall, J., Shazeer, N., Kulshreshtha, A., Cheng, H.-T., Jin, A., Bos, T., Baker, L., Du, Y., et al. (2022) · arXiv · doi:10.48550/arxiv.2201.08239 · Cited by 778
Greshake, K., Abdelnabi, S., Mishra, S., Endres, C., Holz, T., & Fritz, M. (2023) · Proceedings of the 16th ACM Workshop on Artificial Intelligence and Security · doi:10.1145/3605764.3623985 · Cited by 667
Accountable Artificial Intelligence: Holding Algorithms to Account
Busuioc, M. (2021) · Public Administration Review · doi:10.1111/puar.13293 · Cited by 652 · cited 2×
“…range of researchers from both the public and private sectors, including ‘ all researchers in the European Union at all stages of their career and covers all fields of research in the public and private sectors, irrespective of the nature of the appointment or employment ’ ([115] at Section 1).”
BackgroundLarge Legal Fictions: Profiling Legal Hallucinations in Large Language Models
Dahl, M., Magesh, V., Suzgun, M., & Ho, D. E. (2024) · Journal of Legal Analysis · doi:10.1093/jla/laae003 · Cited by 559
The logic of strategic ignorance
McGoey, L. (2012) · The British Journal of Sociology · doi:10.1111/j.1468-4446.2012.01424.x · Cited by 526 · cited 1×
“These harms only become apparent over time, as careless speech is produced, 16 See [33]; Similarly, McGoey [91] has examined the political value of ignorance, according to which claims about having the least knowledge about a phenomenon can have significant strategic value, for example to avoid…”
BackgroundEthical and social risks of harm from Language Models
Weidinger, L., Mellor, J., Rauh, M., Griffin, C., Uesato, J., Huang, P.-S., Cheng, M., Glaese, M., Balle, B., Kasirzadeh, A., et al. (2021) · arXiv · doi:10.48550/arxiv.2112.04359 · Cited by 477 · cited 2×
“Second, LLMs are often fine-tuned or given ‘ guardrails ’ to prevent hate speech, prejudices, gendered language and other ‘ toxic content ’ from appearing in their outputs [97 – 99].”
BackgroundRegulating ChatGPT and other Large Generative AI Models
Hacker, P., Engel, A., & Mauer, M. (2023) · 2023 ACM Conference on Fairness Accountability and Transparency · doi:10.1145/3593013.3594067 · Cited by 466
Can AI language models replace human participants?
Dillion, D., Tandon, N., Gu, Y., & Gray, K. (2023) · Trends in Cognitive Sciences · doi:10.1016/j.tics.2023.04.008 · Cited by 405
Political Astroturfing on Twitter: How to Coordinate a Disinformation Campaign
Keller, F. B., Schoch, D., Stier, S., & Yang, J. (2020) · Political Communication · doi:10.1080/10584609.2019.1661888 · Cited by 378 · cited 2×
“109 Two points are important to note.”
BackgroundBeyond Empiricism: Policy Inquiry in Post positivist Perspective
Fischer, F. (1998) · Policy Studies Journal · doi:10.1111/j.1541-0072.1998.tb01929.x · Cited by 353 · cited 1×
“For a fuller exploration of this topic, see [92 – 94]. r o y a l s o c i e t y pub l i s h i ng . o r g / j o u r n a l / r s o s R .”
BackgroundA large-scale comparison of human-written versus ChatGPT-generated essays
Herbold, S., Hautli-Janisz, A., Heuer, U., Kikteva, Z., & Trautsch, A. (2023) · Scientific Reports · doi:10.1038/s41598-023-45644-9 · Cited by 338
Social Biases in NLP Models as Barriers for Persons with Disabilities
Hutchinson, B., Prabhakaran, V., Denton, E., Webster, K., Zhong, Y., & Denuyl, S. (2020) · Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/2020.acl-main.487 · Cited by 274
Social Biases in NLP Models as Barriers for Persons with Disabilities
Hutchinson, B., Prabhakaran, V., Denton, E., Webster, K., Zhong, Y., & Denuyl, S. (2020) · arXiv · doi:10.48550/arxiv.2005.00813 · Cited by 274 · cited 2×
“While intended to improve safety, these constraints can also prevent models from engaging with sensitive subjects, in particular those related to marginalized groups in society [100,101].”
BackgroundTime-aware language models as temporal knowledge bases
Dhingra, B., Cole, J. R., Eisenschlos, J. M., Gillick, D., Eisenstein, J., & Cohen, W. W. (2022) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00459 · Cited by 212
Chambers, S. (2021) · Political Studies · doi:10.1177/0032321719890811 · Cited by 185 · cited 2×
“11 : 240197 recent scholarship on post-truth politics [33,34], we conceptualize this type of problematic output from LLMs as ‘ careless speech ’ .”
BackgroundMethodToxicity in chatgpt: Analyzing persona-assigned language models
Deshpande, A., Murahari, V., Rajpurohit, T., Kalyan, A., & Narasimhan, K. (2023) · Findings of the Association for Computational Linguistics: EMNLP 2023 · doi:10.18653/v1/2023.findings-emnlp.88 · Cited by 178
Toxicity in ChatGPT: Analyzing Persona-assigned Language Models
Deshpande, A., Murahari, V., Rajpurohit, T., Kalyan, A., & Narasimhan, K. (2023) · arXiv · doi:10.48550/arxiv.2304.05335 · Cited by 178
News Summarization and Evaluation in the Era of GPT-3
Goyal, T., Li, J. J., & Durrett, G. (2022) · arXiv · doi:10.48550/arxiv.2209.12356 · Cited by 175
Chan, A. (2023) · AI and Ethics · doi:10.1007/s43681-022-00148-6 · Cited by 155
How AI can distort human beliefs
Kidd, C., & Birhane, A. (2023) · Science · doi:10.1126/science.adi0248 · Cited by 151
Feng, S., Park, C. Y., Liu, Y., & Tsvetkov, Y. (2023) · Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/2023.acl-long.656 · Cited by 149
AI and the Everything in the Whole Wide World Benchmark
Raji, I. D., Bender, E. M., Paullada, A., Denton, E., & Hanna, A. (2021) · arXiv · doi:10.48550/arxiv.2111.15366 · Cited by 127 · cited 1×
“Instead, reliability can be improved through extensive curation and annotation [50], methodologically sound benchmarking metrics [51,52], long-term fine tuning with expert human feedback [53], auditing and adversarial testing ([5], p. 11, [53], p. 60), and perhaps even downsizing models [17, pp.…”
BackgroundThe European AI liability directives: Critique of a half-hearted approach and lessons for the future
Hacker, P. (2023) · Computer Law & Security Review · doi:10.1016/j.clsr.2023.105871 · Cited by 124
Perturbation Sensitivity Analysis to Detect Unintended Model Biases
Prabhakaran, V., Hutchinson, B., & Mitchell, M. (2019) · Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) · doi:10.18653/v1/d19-1578 · Cited by 115
Rethink reporting of evaluation results in AI
Burnell, R., Schellaert, W., Burden, J., Ullman, T. D., Martinez-Plumed, F., Tenenbaum, J. B., Rutar, D., Cheke, L. G., Sohl-Dickstein, J., Mitchell, M., et al. (2023) · Science · doi:10.1126/science.adf6369 · Cited by 104
Frontier AI Regulation: Managing Emerging Risks to Public Safety
Anderljung, M., Barnhart, J., Korinek, A., Leung, J., O'Keefe, C., Whittlestone, J., Avin, S., Brundage, M., Bullock, J., Cass-Beggs, D., et al. (2023) · arXiv · doi:10.48550/arxiv.2307.03718 · Cited by 77
Reducing conversational agents' overconfidence through linguistic calibration
Mielke, S. J., Szlam, A., Dinan, E., & Boureau, Y.-L. (2020) · arXiv · doi:10.48550/arxiv.2012.14983 · Cited by 72
Reducing Conversational Agents’ Overconfidence Through Linguistic Calibration
Mielke, S. J., Szlam, A., Dinan, E., & Boureau, Y.-L. (2022) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00494 · Cited by 72
How do we know how smart AI systems are?
Mitchell, M. (2023) · Science · doi:10.1126/science.adj5957 · Cited by 71
Truth machines: synthesizing veracity in AI language models
Munn, L., Magee, L., & Arora, V. (2024) · AI & SOCIETY · doi:10.1007/s00146-023-01756-4 · Cited by 65
Laux, J., Wachter, S., & Mittelstadt, B. (2021) · Computer Law & Security Review · doi:10.1016/j.clsr.2021.105613 · Cited by 62 · cited 2×
“Frameworks such as the Unfair Consumer Practices Directive (UCPD) regulate how advertisers ought to communicate with consumers [128].”
BackgroundSelf-citationTeaching Models to Express Their Uncertainty in Words
Lin, S., Hilton, J., & Evans, O. (2022) · arXiv · doi:10.48550/arxiv.2205.14334 · Cited by 59
To protect science, we must use LLMs as zero-shot translators
Mittelstadt, B., Wachter, S., & Russell, C. (2023) · Nature Human Behaviour · doi:10.1038/s41562-023-01744-0 · Cited by 58
Self-citationGround truth tracings (GTT): On the epistemic limits of machine learning
Kang, E. B. (2023) · Big Data & Society · doi:10.1177/20539517221146122 · Cited by 47
Human biases limit cumulative innovation
Thompson, B., & Griffiths, T. L. (2021) · Proceedings of the Royal Society B: Biological Sciences · doi:10.1098/rspb.2020.2752 · Cited by 40
Cited by
AI’s Epistemic Harm: Reinforcement Learning, Collective Bias, and the New AI Culture War
Humphreys, D. (2025) · Philosophy & Technology · doi:10.1007/s13347-025-00928-y · Cited by 26
“4 While this is an important consideration, Wachter et al. ( 2024 ) argue that obvious hallucinations are not the primary epistemic risk created by language models, rather it is “subtle inaccuracies, oversimplifications or biased responses that are passed off as truth in a confident tone - which…”
The Impact of Generative Artificial Intelligence on Research Integrity in Scholarly Publishing
Chauhan, C., & Currie, G. (2024) · The American Journal of Pathology · doi:10.1016/j.ajpath.2024.10.001 · Cited by 22
Krauß, V., McGill, M., Kosch, T., Thiel, Y. M., Schön, D., & Gugenheimer, J. (2025) · Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems · doi:10.1145/3706598.3713083 · Cited by 13
Are Attributions of Consciousness to AI Chatbots Epistemically Innocent?
Peters, U. (2026) · Minds and Machines · doi:10.1007/s11023-026-09795-8
Metadata
- Authors
- Wachter, S., Mittelstadt, B., & Russell, C.
- Venue
- Royal Society Open Science
- Publisher
- The Royal Society
- Published
- 2024-08
- Volume
- 11
- Issue
- 8
- Pages
- 240197-240197
- ISSN
- 2054-5703 (ISSN-L)
- OpenAlex
- W4401390644
- Cited by
- 66
- References
- 50
- FWCI
- 172.92
- Type
- journal article
- Language
- en
- Open access
- gold
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.