A Connecticut lawyer checked the citations in his rough drafts. Then an AI editing pass changed them.
That sequence sits at the center of a July 31 Connecticut Supreme Court order. The lawyer had researched the cases through LexisNexis and verified the quotations and citations with Shepard’s. He then pasted the drafts into ChatGPT to improve their organization and writing. The resulting documents added new citations or altered existing ones. About seven erroneous and unverified citations made it into filings before the state’s highest court.
The legal propositions remained sound. The source trail did not.
The court found no intent to deceive. It noted the lawyer’s cooperation, remorse, clean disciplinary history, and later corrections. It still found a violation of the professional duty of competence. The lawyer must complete six extra hours of continuing legal education, including three hours on generative AI. He and his firm must each donate $1,000 to the Connecticut Bar Institute and report compliance within six months.
The durable signal reaches beyond one case. Verification can expire.
The last prose change owns the risk
A source check belongs to the exact document that was checked. Once a later tool rewrites a quotation, citation, factual statement, or the sentence that a source is supposed to support, the earlier review may no longer describe the current draft.
That sounds obvious when written out. It is easy to miss in a normal workflow. Editing tools arrive after research. Their job appears cosmetic: clean the prose, improve structure, tighten repetition. But any system that can rewrite language can also change meaning or disturb the evidence attached to it.
The Connecticut order makes the sequence unusually visible. The rough drafts were checked. The altered documents were reviewed by colleagues. Nobody repeated the citation check after the AI pass. The court said the firm shared responsibility because other partners reviewed the filings and the firm lacked policies and procedures for responsible AI use at the time.
Connecticut had already turned this concern into a court rule. Practice Book section 4-9 took effect on June 23. It requires independent verification of citations, legal authorities, and evidence produced by generative AI in documents created or edited with it. Filing also represents a good-faith, diligent effort to comply.
This is a Connecticut ruling under Connecticut rules. It does not create a national standard, reject AI-assisted legal work, or provide legal advice. It does show what an enforceable workflow can look like: verify after the final machine-assisted change, then bind the review to the version that gets signed and filed.
Verification bottleneck
Verification is becoming the scarce institutional function.
- What moved faster: AI-assisted editing changed already-checked material faster than the firm’s review process detected the difference.
- Who now has to verify: The signing lawyer owns the final filing, while supervisors and firms need policies that connect tool use, source review, final-version approval, and correction.
- Where the bottleneck sits: A valid case name proves very little by itself. Review must reach the quotation, proposition, current authority, jurisdiction, and exact filed version.
- What to watch next: Court rules that define the final check, firm systems that invalidate stale approvals after edits, and evidence that legal tools preserve a source trail instead of returning another confidence score.
Opportunities
Where value may appear is the legal verification layer.
A useful system could retrieve the source text behind every citation, compare quotations with the original, flag invented or altered authority, test whether the cited passage supports the sentence, and record the authority’s status at the time of review. The receipt should preserve the final document version, reviewer identity, timestamp, and any correction history.
The key product behavior is simple: if a later AI pass changes source-bearing prose, the old verification status should close automatically. The new version returns to review.
Small firms, legal clinics, courts, and self-help programs may also need bounded verification services that produce reviewable clerk and lawyer audit trails. Software can reduce the mechanical load. Qualified people still have to read the source, apply the law, protect confidentiality, and accept responsibility for the filing.
This is builder and operator orientation, not legal, compliance, business, financial, or investment advice.
The next generation of trustworthy editing tools will remember which checks their own edits made stale.