Litigation: Balancing AI Innovation with Responsibility
The first legal filing of the generative AI era to make international news was unremarkable in every way but one. It supported a routine personal injury claim against an airline; it cited 6 precedents, and none of them existed. When Judge P. Kevin Castel fined the responsible lawyers $5,000 in Mata v. Avianca (2023), the sanction seemed a curiosity. Three years on, it reads as an opening entry. A database maintained by legal researcher Damien Charlotin has now catalogued more than 1,700 court decisions worldwide in which a party relied on hallucinated AI material (Charlotin, 2026).
The profession has settled on a comforting frame for this problem: innovation on one side of the scale, responsibility on the other, with the careful lawyer adjusting the weights. The frame is wrong. Three years of case law, empirical research and regulatory guidance point elsewhere. Responsibility is not the counterweight to innovation in litigation. It is the licence for it.
Start with what the research actually establishes. In the first systematic study of the phenomenon, Dahl et al. (2024) found that general-purpose language models hallucinated on verifiable questions about federal case law between 58 and 88 per cent of the time, depending on the model. The same study found something more troubling: the models could not reliably predict their own fabrications, and they tended to adopt a user’s mistaken legal premises rather than correct them. Purpose-built legal research platforms narrow the gap without closing it. Magesh et al. (2025) tested the flagship tools from LexisNexis and Thomson Reuters and found hallucinated or misgrounded answers between 17 and 33 per cent of the time, despite marketing claims of citation-level reliability. (See Thomson Reuters response below.) The figure that should unsettle litigators, though, is not the error rate but the character of the error. A fabricated citation is not a typo. It arrives in flawless citation format, attached to a plausible court and a confident holding, indistinguishable on its face from the genuine article. The quality controls of legal practice evolved to catch human failure, which tends to manifest as sloppiness. Machine failure is fluent.
Courts noticed quickly, and their response has been more consistent than the headlines suggest. In Zhang v. Chen (2024), the British Columbia Supreme Court ordered counsel to pay costs personally after fake citations surfaced in her filings, holding that submitting invented cases is ‘tantamount to making a false statement’ to the court. Ontario’s turn came in Ko v. Li (2025), where Justice Myers confronted a factum built on authorities that did not exist, initiated contempt proceedings and reminded the bar that it is ‘the lawyer’s duty to ensure human review’ of machine-prepared work. When counsel later admitted she had misled the court about who prepared the document, a second contempt proceeding followed (Ko v. Li, 2025 ONSC 6785). In litigation, it seems, the cover-up compounds faster than the error. Read together, the decisions share one striking feature. No Canadian lawyer has yet been sanctioned for using AI. Every sanction was imposed the moment the lawyer stopped being present: the unread case, the unchecked link and the signature applied to work that no human examined.
Regulators have drawn the same conclusion. The Federal Court now requires a declaration whenever AI-generated content appears in filed materials, while expressly promising that the declaration itself attracts no adverse inference (Federal Court, 2024). That is a rule aimed at transparency, not deterrence. The Law Society of Ontario’s white paper takes a similar posture, mapping generative AI onto the existing Rules of Professional Conduct rather than inventing new ones, and locating the relevant duties in familiar places: competence, confidentiality and supervision of delegated work (Law Society of Ontario, 2024). Ontario went one step further, amending its Rules of Civil Procedure in 2024 so that every factum must carry a signed certificate attesting to the authenticity of each authority cited (Rules of Civil Procedure, 1990, r. 4.06.1(2.1)). Nothing in this architecture is novel doctrine. The profession answered its newest technology with its oldest ideas. The rules did not change; the price of forgetting them did.
None of this argues for abstention, and for plaintiff-side litigation, the stakes cut the other way too. Injured people routinely face institutional defendants with deep reserves and tools that compress document review, assemble medical chronologies or produce a serviceable first draft, lowering the cost of thoroughness itself. Used well, that narrows a resource asymmetry that has always favoured the defence. The governing principle is delegation without abdication, the same standard that has always applied to work handed to a student or clerk. Yet a quieter problem sits underneath it. Verification is a skill, and it is built by precisely the work the machines now perform. The junior lawyer who has never summarised a thousand pages of records is poorly equipped to spot the summary that subtly lies. A profession that automates its apprenticeships may eventually discover it has automated away the supply of people qualified to check the machines. Whatever balance means, it will be measured there.
Which returns us to Ontario’s certificate. Faced with software that can fabricate law, the province did not ban the machine. It re-inscribed the human: a signature, warranting that someone read the cases. It is tempting to file that away as a procedural patch. It is closer to a definition. A lawyer’s signature has always quietly meant that a trained mind stands behind these words and accepts what follows from them. Generative AI has not weakened that warranty. It has revealed that the warranty was the product all along. When everything else in the file can be generated, the one thing that cannot be is the reading. That, it turns out, was always the job.
About the Author
Kanon Clifford is a plaintiff personal injury lawyer with Bergeron Clifford. The ability to make a meaningful change in people’s lives is what attracts Kanon to injury law. For Kanon, clients’ right to fair compensation is the pillar of his deep commitment to improving the lives of injured persons and their families.
References
Charlotin, D. (2026). AI hallucination cases [Data set]. https://www.damiencharlotin.com/hallucinations/
Dahl, M., Magesh, V., Suzgun, M., and Ho, D. E. (2024). Large legal fictions: Profiling legal hallucinations in large language models. Journal of Legal Analysis, 16(1), 64–93. https://doi.org/10.1093/jla/laae003
Federal Court. (2024, May 7). Notice to the parties and the profession: The use of artificial intelligence in court proceedings. https://www.fct-cf.ca/Content/assets/pdf/base/FC-Updated-AI-Notice-EN.pdf
Ko v. Li, 2025 ONSC 2766.
Ko v. Li, 2025 ONSC 6785.
Law Society of Ontario. (2024, April). Licensee use of generative artificial intelligence [White paper].
Magesh, V., Surani, F., Dahl, M., Suzgun, M., Manning, C. D., and Ho, D. E. (2025). Hallucination-free? Assessing the reliability of leading AI legal research tools. Journal of Empirical Legal Studies, 22(2), 216–242. https://doi.org/10.1111/jels.12413
Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023).
Rules of Civil Procedure, R.R.O. 1990, Reg. 194, r. 4.06.1(2.1).
Zhang v. Chen, 2024 BCSC 285.
Thomson Reuters:
‘The Magesh et al. (2025) study referenced in the article was completed in 2024. Much has changed since then, and the AI technology it evaluated is no longer available in Westlaw. Thomson Reuters has continued to invest heavily in its legal AI capabilities, driven by progress in language models, continuous testing, and extensive user feedback from real-world legal workflows. That work produced Deep Research, an agentic AI research solution substantially more powerful and accurate than earlier iterations, and Deep Research Verify, which helps legal professionals confirm that each assertion in AI-generated research is supported, complete, and ready to stand behind. It provides a structured review of cited authorities and relevant law so attorneys can move from AI-generated insight to well-grounded legal judgment. Thomson Reuters remains committed to building AI that meets the standards legal professionals require — and to continued collaboration with academia and industry as these technologies advance.’
Mike Dahn, Head of Product Management, Westlaw Thomson Reuters.