18 October 2026
Most AI citation cases start with research: a lawyer asks a chatbot for authority and files what it invents. In TOV Realty, LLC v. Suarez the research was done properly. The fake citations arrived later, when the finished brief was pasted into ChatGPT to be polished.
The case in brief
According to the court's 31 July decision, as reported by Eugene Volokh, the attorney researched on LexisNexis and checked the quotations and citations in his drafts with Shepard's. He then pasted the drafts into ChatGPT to make them better. The court found that the drafts ChatGPT produced “added new case citations or altered existing case citations, making the final documents filed in this court inaccurate and misleading.”
Nobody caught it before filing, and the order explains why. The legal propositions in the briefs were still correct, so the changes gave neither the attorney nor his colleagues a reason to check the citations again. The problem surfaced only when amici flagged citations that looked invented. The attorney filed errata and corrected the briefs, and he and the firm expressed their “deep regret for the errors.”
The usual advice, “verify what the AI gives you”, assumes the AI is the source. Here the source was verified. The risk sat in a step most people think of as harmless: asking a tool to make finished writing clearer. A language model asked to rewrite a brief rewrites all of it, citations included, and a changed page number or a swapped case name does not make the argument read any worse. That is exactly why it gets through review.
It is not rare. Damien Charlotin's AI Hallucination Cases database listed 2,199 decisions worldwide involving AI-generated errors in filings, 1,501 of them in the United States, as of 18 October 2026.
Cite-check last, after the final edit
Any AI pass after your cite check produces a new draft. Treat it as unchecked: run Shepard's or KeyCite again on the version you will actually file, not the one before it was polished.
Compare the filed version with the checked one
Put the cite-checked draft and the final side by side in a diff tool and read every change. A changed reporter page or a new case name stands out in seconds. Our free Text Compare runs in your browser and uploads nothing. Open Text Compare.
Ask for suggestions, not a rewrite
Have the tool point out what could be clearer and make the changes yourself. You decide every word, and nothing arrives that you did not type.
Keep citations out of the AI pass
If you want a tool to tighten the argument, give it the prose without the citations, and put the checked citations back afterwards.
Write it into firm policy
Which tools may touch a filing, and the rule that a full cite check follows the last edit by anyone or anything. Check your court's standing orders on AI use and disclosure too.
Keep the checked version
Save the cite-checked draft as its own file. If a question comes up later, you can show exactly what changed and when.
Where ABCcount fits, exactly
ABCcount's Legal mode is built for this risk. Every rewrite is checked against your original before you see it: if a number from your draft is dropped or changed, including the volume, page or year in a citation, the rewrite fails the check. It also refuses deadlines, sums and threats your draft did not contain.
It does not yet catch a whole citation added from nowhere, so the checklist above still applies. Try it free: 7 days of Pro for one person, or a 14-day pilot for your firm. No card.
An attorney researched on LexisNexis and checked his citations with Shepard's, then pasted the drafts into ChatGPT to improve them. The court found the ChatGPT drafts "added new case citations or altered existing case citations", making the filings inaccurate and misleading. On 31 July 2026 the Connecticut Supreme Court ordered extra legal education, including generative AI, and a $1,000 donation each from the attorney and his firm.
Yes. That is what this case shows. A request to make a brief clearer or more concise is still a request to rewrite it, and a language model can change or add a citation while doing so, with the legal argument still reading correctly.
Run the full cite check again on the final version, and compare the final against the version you checked before the edit, with a diff tool, so every change is visible.
Damien Charlotin's AI Hallucination Cases database listed 2,199 decisions worldwide, 1,501 of them in the United States, as of 18 October 2026. Most involve citations an AI tool invented during research; TOV Realty is unusual because the research was sound and the error came in editing.
Partly, and it is worth being exact. In Legal mode, every rewrite is checked against your original: if a number from your draft is dropped or changed, including the volume, page or year in a citation, the rewrite fails the check. It does not yet catch a whole new citation added from nowhere. Cite-check the final version whatever tool you use.
Sources: TOV Realty, LLC v. Suarez, SC 21183 and SC 21184, Connecticut Supreme Court, filed 31 July 2026; Eugene Volokh, “ChatGPT Editing Adding Hallucinations to Otherwise Accurate Material?”, Reason, 4 August 2026; Damien Charlotin, AI Hallucination Cases database, accessed 18 October 2026. This page is general information, not legal advice.
Also free: a demand letter template and an incident report template.