Connecticut Supreme Court Warns Lawyers After Fake AI Citations Reach Appeals
Connecticut’s Supreme Court put a Google News headline behind a concrete warning: lawyers cannot delegate factual accuracy to artificial intelligence and escape responsibility.
The warning followed flawed appellate filings from attorney Ian G. Gottlieb and GLG Law. The documents contained fabricated quotations, incorrect citations, and inaccurate descriptions of legal authorities. Gottlieb acknowledged using OpenAI’s ChatGPT while preparing briefs for two landlord appeals.
The episode is larger than one lawyer’s mistake. Courts are deciding how to preserve AI’s practical benefits without weakening the verification duties that make legal argument trustworthy. The central conflict is therefore clear: faster drafting now competes with personal accountability for every filed word.
That conflict has appeared before. Lawyers across the United States have submitted nonexistent cases produced by generative AI, sometimes triggering fines, professional discipline, or damaged client claims. Connecticut’s response moves the issue from scattered courtroom rebukes toward a repeatable governance model.
The emerging rule is simple but consequential. Lawyers can use AI, yet they remain personally responsible for confirming every quotation, citation, legal proposition, and item of evidence before submission.
Connecticut Turns an AI Failure Into a Verification Rule
Connecticut’s response targets unverified information, not artificial intelligence itself.
The Connecticut judiciary adopted a requirement covering citations, legal authorities, and evidence produced through generative AI. Lawyers and self-represented litigants must independently verify that material before submitting it to a court.
The rule addresses a defining weakness of large language models. An AI hallucination is a confident output that lacks reliable support, including invented cases or quotations. The text often looks credible because the system has learned linguistic patterns associated with legal writing.
That surface credibility makes legal hallucinations unusually dangerous. A fictional case can include a plausible caption, court name, date, reporter citation, and summary. Each element can resemble a genuine authority even when the underlying decision does not exist.
Connecticut’s judiciary warned that violations can bring serious sanctions. Available consequences include striking a filing, dismissing a claim, entering a default, or imposing another remedy appropriate to the case. Those possibilities make verification part of litigation risk, not merely professional etiquette.
The rule applies whether a person uses a general chatbot or a specialized legal product. That distinction matters because a professional interface can create misplaced confidence. Branding, citations, and polished formatting do not independently establish that an authority exists.
The judiciary’s approach also avoids an impractical question. A court does not need to reconstruct every prompt or determine exactly which model changed a sentence. It can ask whether the filer checked the resulting material against an authoritative source.
That test is easier to apply and harder to evade. A lawyer cannot defend an invented quotation by saying an assistant, contractor, search engine, or chatbot supplied it. The signature on the filing places responsibility with counsel.
Connecticut began formal work on the issue before the latest headline. Its AI legal committee was tasked with considering responsible use across the judicial system. Its stated concerns include fairness, accountability, equality, transparency, and sustainable adoption.
The committee’s scope reflects the breadth of the problem. Generative AI can affect legal research, document review, translation, evidence, judicial administration, and public access. Citation accuracy is only the most visible failure mode.
Still, fabricated authorities offer courts a clear place to draw a line. Judges cannot evaluate an argument efficiently when they must first determine whether its cases are real. Opposing parties also bear costs when they must investigate convincing but fictional references.
For readers arriving through Google News, the important change is not a ban on ChatGPT. Connecticut has instead converted an established professional duty into an explicit AI-era verification requirement.
That design preserves room for experimentation. Lawyers can still use software to brainstorm, reorganize text, summarize documents, or identify research leads. They simply cannot treat generated output as validated law.
Google News Attention Follows Nine Errors in Two Appeals
The controversy became a statewide warning because multiple errors survived the ordinary review process and reached the highest court.
Gottlieb appeared before the Connecticut Supreme Court on July 7, 2026. The seven justices had ordered him and GLG Law to explain why they should not face sanctions over filings in two pending appeals.
The cases involved landlords challenging fair-rent commission decisions in Middletown and Hartford. That setting matters because the underlying disputes affect housing costs and the rights of tenants and property owners. Citation failures were not confined to a harmless internal draft.
According to local reporting, ledgers presented to the court identified nine errors requiring correction. They included inaccurate citations and quotations that did not appear in the cited authorities.
Gottlieb said he used ChatGPT to help prepare the briefs. He claimed accurate information was replaced during the drafting process, although he could not explain exactly when the changes occurred.
That uncertainty reveals a common workflow problem. A user can begin with verified material, ask a model to improve or reorganize it, and receive rewritten text containing unsupported details. The error enters during transformation rather than initial research.
Gottlieb accepted responsibility for failing to catch the mistakes. At the hearing, he described the episode as an attorney’s “worst nightmare” and asked the justices for grace. His lawyer proposed a modest fine and additional professional training.
The attorney also argued that the underlying legal propositions remained accurate. That claim draws an important distinction between an argument’s general direction and the integrity of its supporting authorities.
Courts cannot accept that distinction as a complete defense. A correct conclusion does not make a fabricated quotation permissible. Legal reasoning depends on what a precedent actually held, how broadly it applies, and which words the court used.
A fake citation also prevents normal adversarial testing. Opposing lawyers cannot distinguish, limit, or challenge an opinion that does not exist. They must instead spend time proving the reference is fictional.
Yale Law School’s Jerome N. Frank Legal Services Organization reportedly helped identify errors while representing tenants. The episode therefore shows how hallucinations transfer work to the opposing party, including legal services teams with limited resources.
The Connecticut proceeding was the state Supreme Court’s first reported confrontation with sanctions for AI-produced filing errors. It did not, however, emerge in isolation.
The federal court in Connecticut had already issued a notice describing a no-tolerance approach to briefs that hallucinate legal propositions or severely misstate the law. Its message applies whether AI assisted the work or not.
That last qualification is central. Courts have always required competent research and candor. Generative AI changes the speed, scale, and appearance of mistakes, but it does not create permission to submit them.
An attorney who miscopies a quotation from a book remains responsible. The same principle applies when software produces the text. Connecticut’s warning updates enforcement language without abandoning the profession’s established foundation.
The Google News framing can make the event look like another clash between judges and new technology. The underlying facts support a narrower and more durable interpretation.
The court is not deciding whether AI belongs in legal practice. It is deciding who bears the risk when generated material enters a filing. Its answer keeps that risk with the person presenting the document.
Faster Legal Drafting Now Collides With Personal Accountability
Generative AI promises efficiency, but a lawyer’s duty of verification does not shrink when drafting becomes faster.
Legal work creates strong demand for automation. Lawyers search large bodies of authority, compare factual records, prepare recurring documents, and revise language under tight deadlines. Generative systems can accelerate parts of each task.
Those gains are real enough to encourage adoption. A model can propose an outline, simplify a dense paragraph, extract issues from notes, or generate alternative wording. It can also help users formulate research queries they might not have considered.
The difficulty begins when a probabilistic text generator is treated like an authoritative database. Large language models predict plausible sequences of words. They do not inherently guarantee that a quoted passage appears in a judicial opinion.
Legal research requires a different standard. A lawyer must locate the decision, read the relevant section, confirm its status, and determine whether later authorities changed its meaning. A generated summary cannot complete those steps by appearance alone.
Citation verification also involves more than confirming existence. A real case can be irrelevant, reversed, superseded, or quoted out of context. A citation may point to the wrong page while still looking formally correct.
That creates several layers of review:
Confirm that the cited authority exists.
Open the authoritative or trusted legal source.
Read the passage surrounding every quotation.
Check the page or paragraph reference.
Confirm that the case supports the stated proposition.
Review later treatment and controlling authority.
Ensure the final filed version preserves the verified text.
The last step deserves special attention. A lawyer can verify an early draft and then lose accuracy during later AI-assisted editing. Version control must therefore cover legal substance, not only document formatting.
This is where the Connecticut case pressures law firms. A policy saying “check AI output” is too vague for a complex drafting process. Firms need defined checkpoints, assigned reviewers, and records showing what was validated.
Supervising attorneys also face exposure. A senior lawyer cannot assume a junior colleague understands the limitations of every tool. Competence includes learning enough about a system to supervise its use appropriately.
The American Bar Association addressed these duties in Formal Opinion 512. The guidance connects generative AI with competence, confidentiality, client communication, candor, supervision, and reasonable fees.
The opinion does not require lawyers to become machine-learning engineers. It does require them to understand the benefits and risks relevant to their work. That standard changes as tools and professional practices evolve.
Confidentiality presents another risk beyond hallucinations. A lawyer who places client information into a public AI service can expose sensitive data, depending on the product’s terms and controls. Citation accuracy alone does not make a workflow safe.
Disclosure requirements also vary. Some courts require certifications or information about AI use, while others rely on ordinary filing rules. Lawyers must therefore check the governing court’s procedures instead of assuming one national policy.
The most defensible workflow treats AI as an untrusted drafting assistant. Its output can suggest where to look, but it cannot serve as the final source of legal authority.
This model resembles the way careful professionals handle notes from an inexperienced researcher. The material might be useful, yet every consequential claim must be checked before it reaches a client, counterparty, or tribunal.
Knowledge workers outside law should recognize the same pattern. A generated answer can accelerate discovery while remaining unsuitable as a final record. Separating suggestions from verified evidence is essential in regulated and high-stakes work.
A searchable knowledge base can help teams retain source documents and review history. However, no organizational system removes the need to inspect the original authority.
The Connecticut warning therefore challenges a popular automation promise. Saving time during drafting does not necessarily reduce total work. Weak controls can merely shift that work into correction, sanctions proceedings, and reputational repair.
The Real Risk Is Institutional Contamination
A fabricated citation can harm more than one filing because legal information moves through interconnected systems.
The immediate damage falls on the court and the opposing party. Judges or clerks must investigate questionable authorities. Other lawyers must spend client resources responding to material that should never have been filed.
The client represented by the offending lawyer also faces risk. A court can strike an argument, impose costs, delay proceedings, or dismiss a claim. Even a technically sound position can lose credibility when its support proves unreliable.
The consequences can continue after the case. Public filings enter searchable databases, news reports, internal research collections, and model-training corpora. An invented proposition can be copied before anyone publishes a correction.
This is institutional contamination. False material acquires credibility because it appears inside an official-looking legal document. Later users can mistake repetition for confirmation.
Generative systems intensify that danger because they can reproduce and restyle a mistake at low cost. One fabricated authority may appear in memoranda, client alerts, online summaries, and later prompts. Each repetition obscures its origin.
Courts themselves are not immune to transmission errors. Judges depend on adversarial briefing, clerk research, and source checking. A polished filing can create additional verification burdens even when the court ultimately catches the problem.
That burden threatens access to justice. Large organizations may have teams and paid databases available for validation. Smaller firms, public-interest lawyers, and self-represented litigants often work with fewer resources.
AI can help those users access information, but unreliable outputs can also impose disproportionate costs on them. Connecticut’s rule must therefore balance accountability with practical implementation.
A strict verification duty is defensible because the consequences of false authority are serious. Yet courts should communicate the rule clearly and provide education that distinguishes acceptable assistance from unsafe reliance.
Training should include concrete examples. Users need to see how fabricated case names, altered quotations, and misleading summaries appear. Abstract warnings about “AI risk” will not reliably change daily behavior.
Sanctions also require proportionality. Intentional deception, reckless disregard, weak supervision, and an isolated corrected mistake do not present identical circumstances. Courts traditionally consider conduct, harm, notice, and remediation when selecting remedies.
The skeptical question is whether a new AI-specific rule adds value beyond existing duties. Lawyers already owe courts candor and must conduct reasonable inquiries before filing documents.
There is force in that objection. A lawyer who invents a citation manually causes the same factual problem. Technology should not become a distraction from the underlying failure to verify.
Connecticut’s approach partly answers this concern by focusing on output rather than banning a tool. The rule makes a known failure mode explicit while keeping professional responsibility technology-neutral at its core.
National experience supports that emphasis. In 2023, lawyers in the Mata v. Avianca litigation received sanctions after submitting nonexistent decisions attributed to ChatGPT. That case became an early public warning about generated legal authorities.
Michael Cohen later said he had misunderstood Google Bard as a form of enhanced search when it supplied nonexistent cases. His lawyer submitted the material without adequate verification, according to court reporting.
The distinction between a chatbot and a search engine remains important. Search systems retrieve and rank existing pages, although their results still require evaluation. A generative model can synthesize entirely new text that resembles retrieved information.
Products increasingly blur those categories. Search interfaces now generate summaries, while chatbots display citations and web results. Users can reasonably struggle to identify which statements came from a source and which came from synthesis.
That ambiguity makes interface design part of the risk equation. Vendors can reduce confusion by exposing source passages, displaying uncertainty, and preventing unsupported quotations from appearing as verified text.
They cannot eliminate professional responsibility. Even a linked source must be opened and read. A citation badge is evidence of retrieval, not proof that the generated sentence accurately represents the document.
The Connecticut case also highlights a verification gap in public discussion. News aggregation can circulate the dramatic headline faster than readers can locate the order, hearing record, or governing rule.
Google News is useful for discovering coverage, but it is not the final authority on a judicial requirement. Lawyers should move from an aggregated headline to the court’s official material and relevant professional rules.
That source discipline applies equally to AI. Discovery and verification are separate stages. Trouble begins when an efficient discovery tool is silently promoted into an authority.
Connecticut’s Warning Pressures Law Firms and AI Vendors
The next phase will test whether firms can turn a clear duty into controls that survive real deadlines.
Large law firms are already developing policies for approved tools, protected data, human review, and vendor assessment. Smaller practices face the same ethical duties without comparable technology or compliance teams.
Connecticut’s rule increases the value of simple, auditable processes. A firm does not need an elaborate AI laboratory to verify citations. It needs reliable access to sources, clear ownership, and enough time for review.
A practical policy should define permitted uses. Brainstorming, formatting, summarization, translation, and citation generation carry different risks. Treating every AI interaction as equivalent produces rules that are either too loose or impossible to follow.
Firms should also preserve a clean source layer. Verified cases, statutes, regulations, contracts, and record materials should remain distinguishable from generated prose. Reviewers must know which statements trace directly to those sources.
Finalization requires another checkpoint. Every citation and quotation should be validated in the version prepared for filing, not only in an earlier research memo. Automated comparison can assist, but a responsible lawyer must own the result.
Client communication deserves similar clarity. Lawyers should decide when AI use creates a material reason to inform a client. Relevant factors include confidentiality, the tool’s role, contractual limits, and the consequences of error.
The first observation signal is Connecticut’s handling of the GLG Law matter. A final sanction or detailed opinion would show how the justices distinguish negligence from knowing misrepresentation.
A severe penalty would strengthen the view that courts are moving from education toward deterrence. A measured remedy focused on training would suggest that early enforcement remains corrective.
The second signal is adoption by other state courts. Similar certification rules would create a wider operational baseline for firms practicing across jurisdictions.
Broad adoption would strengthen Connecticut’s model. Divergent rules would increase compliance complexity and pressure national firms to adopt the strictest workable standard internally.
The third signal is vendor behavior. Legal AI providers can add better citation validation, source locking, version tracking, and warnings when generated language lacks direct support.
Visible improvements would reduce accidental misuse, although they would not transfer ultimate responsibility away from lawyers. Continued failures inside specialized products would reinforce the need for independent checks outside the model.
Legal technology companies also face a communication challenge. Marketing that emphasizes speed without explaining verification can encourage unsafe expectations. Products should describe which tasks they automate and which judgments remain with licensed professionals.
Courts must watch self-represented litigants as well. A certification requirement can improve accuracy, but unfamiliar procedures may confuse people using free public tools. Plain-language guidance and accessible source links can reduce that burden.
The larger lesson reaches beyond litigation. Doctors, accountants, engineers, researchers, and executives also work with evidence that must survive scrutiny. Generative AI can assist each profession while producing statements that sound more certain than their support allows.
Teams should ask four questions before relying on generated work:
What original source supports this claim?
Who confirmed that the source says what the draft asserts?
Did later editing change the verified language?
Who accepts responsibility for the final output?
Those questions are deliberately ordinary. Connecticut’s warning does not demand an exotic technical solution. It demands that organizations restore visible accountability inside a faster production process.
For readers following the story through Google News, the lasting development will not be another embarrassing hallucination. It will be whether verification becomes a standard, documented stage of AI-assisted professional work.
The Connecticut Supreme Court has already made its position clear: polished output cannot substitute for checked authority. Lawyers can automate parts of drafting, but they cannot automate away the signature beneath the filing.
That principle gives legal teams a direct next step. Review every AI-assisted workflow, identify where generated claims can enter final documents, and assign a person to verify them against primary sources.
The question is no longer whether professionals will use generative AI. They already do. The question is whether their verification systems can move as quickly as the software without sacrificing the trust their work requires.
