The Direct Answer: Every Court Citation Is the Lawyer's Own Work Product
As of September 2026, courts treat a citation in a brief as the filing lawyer's personal representation that the authority exists, supports the proposition stated, and has actually been read. That duty does not transfer to a tool, no matter how sophisticated the tool is. Generative AI can assemble a research section in seconds, but its output remains an unverified draft until a human checks every authority against a primary-source database. An ABA Journal report that the Fifth Circuit instructed lawyers to verify their citations is instructive for a second reason: the court reportedly cited the wrong rule while issuing the warning. Even a court can misstate the law it points to, which is precisely why the verification burden sits with counsel.
Also worth reading: How Do Lawyers Verify AI-Generated Legal Citations in 2026? · Can lawyers be sanctioned for AI-hallucinated case citations, and how do you avoid it? · Can courts sanction a party for AI-related spoliation in eDiscovery, and what do lawyers actually need to do in 2026?
Verification in practice means opening each cited case, statute, or rule in a reliable reporter or official source and confirming the case name, court, year, reporter citation, pinpoint page, and the exact language being relied upon. It also means running a subsequent-history check so that a reversed, vacated, or superseded authority does not slip into a filing. The lawyer who cannot produce the opened source within a few minutes of being asked has not verified anything, regardless of the AI's confidence. By 2026, that expectation is widespread rather than novel, and it rests on ordinary competence and candor duties that long predate AI.
Why Generative Tools Invent Citations That Look Real
Large language models generate text by predicting what typically follows a pattern, not by querying a live legal database. When a user asks for supporting authority on a novel or sparsely documented question, the model fills the gap with the statistical shape of a citation: a real court name, a plausible year, a reporter abbreviation, and a page number that have never existed. As a Bar and Bench commentary puts it, nothing about a fabricated citation looks fake to a reader who does not already know the case. The danger is highest exactly where research is hardest, namely recent decisions, niche jurisdictions, and local rules that appear in few reported sources.
The same mechanism produces fabricated quotations, invented holdings, and misattributed reasoning even when the underlying case is real. A model may place a confident sentence in the mouth of a justice who never wrote it, or attribute a minority view to the majority. The Fifth Circuit episode described by the ABA Journal shows the failure mode at its most embarrassing: a warning about unchecked citations accompanied by an incorrect rule citation. If that can happen in an appellate opinion drafted with assistance, counsel should assume an AI-generated research section contains at least one similar defect until proven otherwise.
None of this makes generative tools useless for legal research. They excel at generating search terms, summarizing the issues in a document set, and suggesting angles that a lawyer can then confirm. The failure lies in treating fluent output as retrieved fact. Courts have noticed the distinction, and reported enforcement figures in 2025 and 2026 show that the gap between what tools produce and what filings require is being policed rather than overlooked.
The 2025-2026 Enforcement Picture in Numbers
Contemporary reporting puts the scale of the problem in concrete terms. One widely repeated count finds at least 167 filings containing AI-hallucinated, fictitious citations across 51 courts and tribunals, a number that almost certainly understates the true total because many defective filings are corrected before anyone measures them. Individual sanctions are material: a lawyer for State Farm was fined $999.99 over seven fake AI case citations, a modest sum that functions as a deterrent rather than a budget item. A JD Supra analysis reported $145,000 in penalties during the first quarter of a recent year, a sign that sanctions are moving from isolated incidents to a pattern.
The consequences are not limited to money. In one reported matter, AI errors left a mortgage trustee without a brief in a foreclosure appeal, showing how a hallucinated citation can cost a party its chance to be heard. Surveys cited in 2026 suggest that roughly 61% of federal judges use AI themselves, which means the bench is familiar with the technology and with its weaknesses. Local rules are also hardening: Miami-Dade and Broward courts have issued unified AI disclosure rules, and standing orders with AI-specific requirements now exist in multiple jurisdictions. A June 9, 2026 item referencing a communication from the Chief Justice on generative AI shows that judicial attention has moved to the institutional level rather than remaining confined to occasional contested filings.
A Verification Workflow That Survives Scrutiny
The safest way to use AI for legal research is to treat its output as a list of leads rather than a list of authorities. Start by asking the tool for search terms, issue framing, and candidate sources, then confirm each candidate independently in a subscription platform such as Westlaw, Lexis, or Practical Law, or in a free primary-source repository such as CourtListener for public opinions. For every citation that will survive into the brief, open the full text and check four things: that the case exists with the exact name and court, that the reporter citation is correct, that the pinpoint page contains the quoted language, and that the quoted language actually supports the proposition. A citation that fails any of these four checks does not go into the brief, no matter how helpful the point would have been.
The second half of the workflow is history and context. Run a citator such as KeyCite or Shepard's on each case and confirm that it has not been reversed, vacated, distinguished, or superseded, and that a more recent decision has not changed the law. If a quotation will be used verbatim, compare it word for word against the reporter version, because AI often paraphrases while presenting the result as a quote. For statutes and rules, confirm the current version and the effective date, and check for local variations that a national summary will miss.
Finally, document the work and build in a second set of eyes. A practical record includes a short verification memo or spreadsheet listing each citation, the source opened, the date checked, and the reviewer, along with saved copies of the opinions. Budgeting matters: manually confirming a single citation can take three to five minutes once research is complete, so a fifty-citation section can consume several hours. A second lawyer reviewing the citations is cheaper than answering a show-cause order, and the memo becomes exhibit-ready evidence of the reasonable care courts expect.
Comparison: Standalone AI Chatbots Versus Research Platforms With AI Features
Not every source of citations carries the same risk. Standalone chatbots are fast and inexpensive but produce text from model memory, while integrated platforms answer from a licensed corpus and attach citator data to each result. Neither replaces human judgment, but they fail in different ways and suit different stages of the process.
| Feature | Option A: Standalone AI chatbot | Option B: Research platform with integrated AI |
|---|---|---|
| Source of citations | Model memory and training patterns | Licensed case law, statutes, and secondary sources |
| Citation reliability out of the box | Low; fabricated cases and quotes are common | Higher; results link to retrievable primary sources |
| Speed of first draft | Seconds to minutes | Minutes, with search steps exposed |
| Pinpoint and quotation checks | Manual and often skipped | Easier because full text is one click away |
| Subsequent-history signals | Usually absent | Citator status attached to results |
| Typical cost structure | Low or subscription-based chat access | Higher per-seat subscription, often bundling AI features |
| Court acceptance | Depends entirely on lawyer verification | Still requires verification, but easier to document |
| Audit trail | Prompts and outputs only | Search history plus saved reports and alerts |
The Mistakes That Repeatedly Trigger Sanctions
The most common failure is citing a case the lawyer never opened, which is effectively outsourcing professional judgment to a tool. A related error is trusting a summary: AI-generated case summaries, headnotes, and compilations are secondary material, and a court holds the filing lawyer to the actual text. Lawyers also forget subsequent history, relying on a model that pulled a case from a training set predating its reversal, and they overlook local rules that require disclosure of AI use or impose standing-order requirements specific to that court.
Quotation errors form a second cluster. Models frequently present paraphrase as verbatim language, and a misquoted sentence is treated as a misstatement of the record. Firms also assume a paid AI add-on means the tool checked the law, when features such as automatic citation formatting do not verify that a case exists. Teams skip documentation, so when a show-cause order arrives, the lawyer cannot demonstrate that any reasonable check was performed. Non-lawyer staff who prepare research drafts intensify the risk because supervision duties fall on the attorney who signs. None of these mistakes is exotic, which is why courts treat them as avoidable rather than excused by complexity.
When to Verify and How Quickly
Verification is not a final polish step; it belongs at the moment research enters the draft, because a fabricated citation that reaches a completed brief is harder to remove and easier to discover. The practical trigger is simple: any citation produced by a generative tool in the last seven days is unverified until a lawyer confirms it, and any AI-assisted research heading into a filing, client memo, or opposing-party communication should be checked before it leaves the firm. Deadlines compress risk. If a filing is due within forty-eight hours, manual verification of every citation is the only defensible approach, and the drafter should expect hours of work rather than minutes.
There are also event-based triggers. A show-cause order about AI use should prompt an immediate file review of every related submission, and an opposing lawyer's citation challenge should be answered with a source, not a denial. Firms handling filings in multiple jurisdictions should track each court's local rules, because Miami-Dade and Broward's unified rules show requirements can change faster than national ethics guidance. The habit worth building is a stop rule: no citation moves from research notes to draft without an opened source and a recorded check.
What Verification Costs, and What Skipping It Costs
The direct cost of verification is time and, in many firms, a research subscription. Commercial legal research platforms commonly run from several hundred dollars to more than a thousand dollars per seat per year, with AI features such as Lexis+ AI or CoCounsel included or priced as add-ons, while standalone chatbot access costs far less. Manual checking of a single citation takes roughly three to five minutes, so a fifty-citation research section is realistically a three-to-five-hour task, and complex authorities with negative treatment take longer. None of that is extravagant next to the alternative.
The price of skipping verification is disclosed in sanctions. The reported $999.99 fine for seven fabricated citations and the $145,000 in first-quarter penalties cited by one legal analysis illustrate a range from irritating to existential for a solo practitioner. Beyond monetary penalties, defective filings can be stricken, deadlines can pass without a brief, and in one reported case a mortgage trustee lost its appellate presentation entirely. The reputational cost lands on the supervising lawyer and the firm, and a single show-cause order can ripple into every later filing. Budgeting for verification is therefore cheaper than budgeting for sanctions, and the comparison is easy to defend to a client.
Disclosure, Supervision, and the Paper Trail
Disclosure rules are converging but not identical across courts. The ABA's Formal Opinion 512 requires lawyers to disclose AI use in court filings, and individual courts layer on their own requirements through standing orders and local rules, such as the unified Miami-Dade and Broward AI disclosure rules. A defensible disclosure identifies the tool, the purpose for which it was used, and the extent of human review, rather than offering a generic statement that AI was used. Where a court has not adopted a specific rule, disclosing use consistent with the jurisdiction's practice reduces the risk that the disclosure itself becomes a disputed issue.
Ethics duties do the real work here. Competence under Model Rule 1.1 now includes a working understanding of AI's limitations, and supervision duties under Rules 5.1 and 5.3 require firms to train staff who use these tools on filings. A March 2026 report of an affidavit admitting that AI-generated expanded legal research used to prepare an order had not been verified shows how quickly a missing paper trail becomes a fact in a sanctions dispute. The paper trail that avoids that outcome is unglamorous: verification logs, saved opinions, prompt records, and reviewer notes kept with the draft. Firms that produce those records can answer a challenge within days, while firms that cannot must reconstruct their research under deadline.
The Practical Bottom Line for Legal Teams in 2026
The law on AI citations is less about new prohibitions than about an old standard applied firmly. Courts expect the lawyer who signs the brief to know every authority in it, and by September 2026, with at least 167 reported defective filings across 51 courts, active sanctions, and local disclosure rules in multiple jurisdictions, that expectation carries enforcement behind it. The sound practice is to let AI generate leads, issues, and drafts, then verify every citation against a primary source, check its history, and record who checked it. A second reviewer on a short research section is inexpensive next to a show-cause order, and a saved verification memo is the best defense available.
For firms evaluating tools, the decisive question is not which model writes the most fluent research; it is which workflow makes verification easiest. Integrated platforms such as Westlaw with CoCounsel or Lexis+ AI generally reduce risk because they link to retrievable sources, while standalone chatbots remain useful for framing and search terms when their output is confirmed. Either way, the duty is unchanged, and the deadline for meeting it is before the brief is filed, not after the first complaint arrives.