Can Claude Check If a Case Is Still Good Law? No, Not Alone

The Short Answer

No. Claude cannot reliably tell you whether a case is still good law, and you should not use it, or any general-purpose AI model, as a substitute for Shepard's, KeyCite, or another citator. A language model does not maintain a current, structured record of subsequent history; it predicts plausible text, and "this case was overruled in 2023" and "this case remains good law" are both plausible text. What Claude can do well is the work around the citator: extracting every citation from a brief, flagging authorities that look suspicious, summarizing what a case you have already pulled actually holds, and checking whether your brief's characterization of that holding is fair. This post separates those two jobs clearly and gives you a workflow that keeps the citator in the loop.

What a Citator Does That a Language Model Does Not

Shepard's (Lexis) and KeyCite (Westlaw) are editorial systems. Trained staff and, increasingly, software read every new opinion, identify which earlier cases it cites, and classify the treatment: followed, distinguished, questioned, overruled, reversed on other grounds, superseded by statute. That classification is stored as structured data attached to the cited case, and it is updated continuously. When you see a red flag on a case, you are seeing the output of a database that was built specifically to answer the question "what has happened to this case since it was decided."

A language model has none of that. Its knowledge of case law is whatever appeared in its training data, frozen at a cutoff date, with no structured index of citing references, no editorial treatment codes, and no mechanism to know that a decision issued last Tuesday narrowed the case you are citing. Ask Claude whether a case is good law and it will answer confidently from pattern-matching on the case name and whatever it absorbed about it. Sometimes that answer is right, and that is precisely the problem: it is right often enough to feel trustworthy and wrong often enough to end a career.

The record of what happens when lawyers skip this step is public. Damien Charlotin's AI Hallucination Cases database tracks court decisions worldwide involving AI-fabricated citations or content; it had identified 1,598 cases as of June 9, 2026 and is updated daily. The majority involve self-represented litigants, but the cases involving licensed attorneys are the ones that end in fines, referrals, and suspensions, as a 2026 sanctions roundup catalogs. Note that "hallucination" in these cases covers more than invented cases: real cases cited for propositions they do not support, and real cases cited after they were overruled, appear throughout the database.

Even the Legal Research Platforms Get This Wrong Sometimes

It is worth being precise about how hard the problem is. Stanford's RegLab tested the AI research assistants built into the major legal platforms, tools that are grounded in the real case-law corpus and sit next to a citator, and found that Lexis+ AI, Westlaw AI-Assisted Research, and Ask Practical Law AI hallucinated between 17 and 33 percent of the time in its preregistered evaluation, with general-purpose GPT-4 doing worse. Those tools have improved since the study's 2024 data, and they are still far better positioned than a general chatbot because they can retrieve real opinions and surface citator flags inline. But if grounded, purpose-built research tools miss a third of the time on hard questions, a general model answering from memory is not a citator substitute. It is not close.

What Web Search Changes, and What It Does Not

Claude can search the web, and this changes the picture in a limited way. With search on, Claude can find the opinion itself on a court website or a free repository, find later opinions that mention it by name, and read a law-review note or blog post discussing whether the case was overruled. That is meaningfully better than answering from memory. It still is not a citator, for three reasons.

• Coverage. Web search finds what is indexed and phrased in a way the search engine matches. A citing opinion that limits your case without naming it prominently, an unpublished disposition, or a state intermediate appellate decision buried in a PDF may not surface. Absence of a search result is not evidence of good law.
• Treatment classification. Finding a later case that cites yours tells you it was cited. It does not tell you whether it was followed or gutted. Claude can read the later opinion and give you an opinion about the treatment, and that opinion is worth reading, but it is a first-year associate's read, not an editorial classification.
• Currency. Search indexes lag, and Claude's summary of a search result can itself be wrong. Every fact it surfaces still has to be confirmed at the source.

Free sources narrow the gap further, but only so far. CourtListener (run by the nonprofit Free Law Project) shows citing opinions for a case and, since May 2026, is available to Claude as a free connector with citation-network and citation-verification tools. The Free Law Project is explicit that its citator, meaning treatment classification (overruled, affirmed, and so on), is still under development; its 2025 progress report describes promising accuracy on Supreme Court overrulings and says the work is a first step, not a product. Google Scholar's "How cited" view is similar: useful for a feel for how a case has been treated, not authoritative. On the commercial side, Thomson Reuters ships a CoCounsel Legal connector for Claude, but its own connector guide states that KeyCite is not currently exposed through it. So as of September 2026, there is no path that puts a real citator's treatment flags inside a Claude conversation; you go to the citator.

What Claude Is Genuinely Good At in a Cite-Check

Having drawn the line, here is the other side of it. Used for the right jobs, Claude makes a cite-check faster and more thorough than most lawyers do by hand:

• Extracting every citation. Paste a brief and ask for a table of every authority cited, with pinpoint, the proposition it is cited for, and the page it appears on. This is tedious for a human and trivial for the model, and it produces the checklist your citator run works from.
• Flagging the suspicious ones. Ask Claude to mark citations that look anomalous: a reporter volume that does not match the year, a case name with a party that appears nowhere else in the brief, a pinpoint outside a plausible page range, a quotation that reads like a paraphrase. It will not catch everything, but it catches the pattern errors that fabricated citations tend to share.
• Holding-versus-characterization checks. Once you have pulled the real opinion, give Claude the opinion and the sentence in your brief that relies on it, and ask whether the brief's characterization is fair, overstated, or dicta dressed as holding. This is where the model's reading ability is a real asset, and it addresses the second most common failure in the sanctions cases: real case, wrong proposition.
• Quotation verification. Give it the opinion and ask it to locate each quoted passage verbatim and report any that do not appear. Misquotation is its own sanctions category.

Our legal research and cite-check workflow packages these steps with prompts, and the prompt guide for lawyers covers how to ask for structured output you can work from.

A Verification Workflow You Can Actually Run

1. Extract. Have Claude produce the citation table from your draft: authority, pinpoint, proposition, brief page.
2. Triage. Ask Claude to flag anomalies. Treat flagged items as "verify first," not "verify only." Every citation goes through the next step regardless.
3. Retrieve. Pull each opinion from Westlaw, Lexis, the court's site, or CourtListener. If it does not exist, it comes out of the brief. This step catches fabrication.
4. Shepardize or KeyCite. Run every authority through the citator and read the negative treatment, not just the flag color. A yellow flag for "distinguished" on an unrelated point is different from "overruled in part" on your point. This step catches bad law, and nothing in Claude replaces it.
5. Characterization check. Feed the real opinion back to Claude with your brief's sentence and ask whether the characterization holds. Fix overstatements.
6. Record it. Keep the citation table with a column for "citator run, date, result." If a court asks, as many now do under standing orders, you have a contemporaneous record of human verification.

The ethics framework behind this is not new. Rule 1.1 competence, Rule 3.3 candor to the tribunal, and Rule 11 all predate AI, and ABA Formal Opinion 512 simply applies them: the lawyer who signs the filing is responsible for every authority in it. Our companion piece on AI disclosure requirements in 2026 covers the growing list of courts that also want to know whether AI was used.