When AI Editing Breaks Checked Citations: Connecticut's Supreme Court Weighs In

TL;DR: On July 31, 2026, the Connecticut Supreme Court sanctioned a lawyer and his firm for briefs with about seven bad citations. The research had been checked. The errors came later, when the drafts were pasted into ChatGPT to improve the writing and the tool quietly added or changed citations. The court found no intent to deceive and still imposed sanctions. The lesson is simple: the check has to run on the final version, after the last edit, not on an earlier draft.

Most stories about AI and fake case law follow the same script. Someone asks a chatbot for cases, the chatbot invents some, and nobody checks. The Connecticut case is different, and more useful, because the research was done right. The failure happened in a step most people would call harmless: asking an AI tool to polish the writing.

What happened in TOV Realty v. Suarez?

The case is TOV Realty, LLC v. Suarez, SC 21183, decided together with Kosel Equity, LLC v. MacGregor, SC 21184. Both were landlord appeals connected to fair rent commission decisions. The court decided the appeals themselves earlier in 2026. The sanctions order came on July 31, 2026, in an unsigned (per curiam) opinion of the court.

The court opened by saying it was confronting a lawyer's use of generative AI that produced hallucinated citations “for the first time.”

According to the order, the lawyer did his research on LexisNexis. He checked the quotations and citations in his rough drafts with Shepard's, the citation-checking service. Then he copied the drafts into ChatGPT to make the writing process faster and improve the final product. Without his knowledge, ChatGPT added new case citations and changed existing ones.

Other lawyers at the firm reviewed the documents before filing. Nobody checked the citations a second time.

Why did no one catch the errors?

Because the arguments still read correctly. The court noted that the AI changes did not affect the legal propositions in the briefs, which remained well established and correct. Since the law looked right, nothing prompted the lawyer or his colleagues to check the citations again.

That is the trap. A reviewer reading for substance sees a sound argument and moves on. The citation under the sentence is a different thing. It can be wrong even when the sentence above it is right.

The errors came to light only when amici, represented by the Jerome N. Frank Legal Services Organization, told the court about what appeared to be hallucinated citations. The lawyer then filed corrections.

Did the lawyer argue the law was correct anyway?

A reader might expect this case to test a “the law was right, so no harm” argument. It did not. At a show cause hearing on July 7, 2026, the lawyer took full responsibility. Through counsel, he admitted filing a total of seven erroneous and unverified citations and admitted that doing so violated Rule 1.1, the duty of competence. The court agreed.

The court noted that other rules might apply, including Rule 3.3 (candor to the court), Rule 1.6 (confidentiality), Rule 5.1 (supervision), and Rule 8.4 (misconduct). Because of its finding under Rule 1.1, it did not decide those questions. So the case does not rule on whether correct law excuses a bad citation. It treats the bad citation itself as the problem.

What sanctions did the court impose?

The court found the errors came from negligence in using a new technology, not an attempt to deceive. It counted the lawyer's candor, cooperation, remorse, and clean record in his favor. Then it imposed sanctions anyway:

• Six hours of continuing legal education in ethics and law office management, on top of the usual twelve, with three of those hours on generative AI.
• A $1,000 donation from the lawyer and another $1,000 from the firm to the CT Bar Institute.
• A report from the lawyer and the firm to the court within six months on whether they complied, with a courtesy copy to the Statewide Grievance Committee.
• Compliance with any reciprocal discipline rules in other places where the lawyer is admitted.

The court also posted the order publicly and sent a copy to the Statewide Grievance Committee. It said that copy is not a referral for further discipline.

The lawyer's own counsel had suggested six hours of education and a $250 payment. The court went higher on the payment and added a separate payment from the firm.

Why was the law firm sanctioned too?

The court said the firm shared responsibility because it had no policies or procedures in place for the responsible use of generative AI. Partners had reviewed the filings. Review without a rule about AI was not enough.

That point reaches beyond one lawyer. A firm that lets people use AI tools without a written rule on when and how citations get rechecked is exposed in the same way.

What did the court say about privilege and public AI tools?

In a footnote, the court added that the duty of competence in using generative AI includes understanding what it means for attorney-client privilege, client confidentiality, and the work product doctrine, especially when a public platform is involved. It cited United States v. Heppner, a 2026 federal decision from New York on AI and privilege. Pasting a client draft into a public chatbot raises more than an accuracy question. For more on that ruling, see the post on AI and attorney-client privilege after Heppner.

What do Connecticut's new rules require?

The court pointed to new rules that took effect this summer:

• Practice Book § 4-9 (b), effective June 23, 2026, for the trial courts. As the court quoted it, anyone who uses generative AI “in the creation or editing” of a court filing must independently verify all citations, legal authorities, or evidence the AI produced.
• Practice Book § 4-2 (b), same date. Signing a filing certifies compliance with § 4-9.
• Practice Book § 62-6 (d) and § 85-2 (11), effective July 14, 2026, for appeals. The first carries the same signing rule into the appellate courts. The second makes failure to comply sanctionable.

Note the word “editing.” Connecticut's rule covers exactly what went wrong in this case. Using AI only to clean up prose still triggers the duty to verify.

Florida takes a similar path by a different route. Since June 15, 2026, Florida Rule 2.515(d)(2) has required the signer of a filing to represent that the cited authorities “exist and are accurately cited.” A companion Florida Supreme Court order, AOSC26-12, bars local courts from adding their own AI disclosure or certification rules.

What does this mean for anyone using AI on legal writing?

Four practical points come out of the order.

Check the final version. A citation check on a rough draft does not carry forward. Any tool that touches the text after the check, including one used only for style, can change a citation. The check that counts is the one run on the document that gets filed.

Check the citation, not just the argument. Reading for whether the law sounds right will not catch a changed case name, volume, or page. Someone has to pull each cited source and confirm it says what the sentence claims.

Put the rule in writing.The firm in this case was sanctioned for having no AI policy. A short written rule, such as “every citation is rechecked against the source after the last edit, by a named person,” is the kind of procedure the court found missing.

Treat style tools as drafting tools. If an AI tool can rewrite a sentence, it can rewrite the citation in it. Some people limit AI editing to text with no citations in it, or strip the citations out first and put them back from the checked version. Either way, the filed document gets checked last.

For a step-by-step way to build that check, see a framework for verifying AI-generated legal content.

The court closed by saying it was not rejecting new technology in law. Its message was narrower. People who sign court filings remain “personally responsible” for making sure what they submit is true and accurate. Checking an earlier draft does not change that.