The first time a firm catches a citation that does not exist, the reaction is almost always to interrogate the tool. That is the wrong object. The tool did what these tools do. What the firm discovered is that it had adopted a capability without adopting a control.
The duty is not new, and it is not optional
ABA Formal Opinion 512, issued July 29, 2024, places generative AI inside the same supervisory framework that already governs nonlawyer assistance under Model Rules 5.1 and 5.3. Managerial lawyers are required to establish clear policies on permissible use. Supervisory lawyers are required to make reasonable efforts to ensure compliance. And lawyers are required to verify AI-generated content before submitting it to a tribunal.
That last one is the operative sentence, because it is a duty that scales with output. A tool that triples drafting throughput triples the verification load, and the opinion is indifferent to whether the firm has anyone to absorb it.
The failure mode is now documented, at scale
This is no longer a hypothetical risk that vendors get to wave away.
The public AI Hallucination Cases database maintained by Damien Charlotin at HEC Paris had logged more than 1,500 court decisions worldwide involving AI-fabricated material by mid 2026, and the count continues to rise. The figure is approximate by construction, because the database is updated continuously.
More usefully for anyone choosing tools: Stanford RegLab's peer-reviewed research found that even purpose-built legal research tools produced hallucinated output in roughly 17 to 34 percent of queries, depending on the tool. Not consumer chatbots. Products sold to lawyers, for legal work.
Take the low end of that range and treat it as a defect rate. No production process tolerates a 17 percent defect rate at the last stage before it ships. It survives here only because the inspection step is invisible and unpriced.
Speed without verification is not efficiency
The arithmetic is worth doing plainly. If a tool halves drafting time and the resulting draft requires a full re-read against source to trust, the firm has not saved anything. It has moved the labour from a task lawyers are trained for to one they find tedious, and tedious inspection is where inspection fails.
Firms respond to this in one of three ways, and two of them are bad:
Slow down. Restrict the tool to low-stakes work. Preserves safety, discards the gain.
Trust it. Spot-check, ship, hope. This is the population the Charlotin database is sampled from.
Make checking cheap. Change the output so that verification is a bounded, mechanical act rather than a re-reading of the underlying file.
Only the third one is a real answer, and it is a property of the output, not of the model.
What makes verification cheap
A finding is cheap to check when checking it means opening one thing.
That is the whole design constraint. If a conclusion arrives as a paragraph of confident narrative, verifying it means reconstructing the reasoning against the record — which is the twenty hours the firm was trying to avoid. If the same conclusion arrives as a discrete finding carrying the document, the page, the date, and the excerpt it rests on, verifying it is a click and a glance.
Three properties follow from that, and they are the ones worth insisting on from any system that touches a case file:
Findings stay discrete. A generated narrative is unauditable in principle, because there is no unit to check. A list of findings has units.
Every material finding exposes its source. Not a footnote to a document, a path to a page. If the reviewer has to search for the support, the support does not count.
Nothing that cannot be traced goes out. An unverifiable finding is flagged, not filed. This is the rule that makes the other two mean something.
Log it, because you will be asked
The supervision duty is not discharged by having verified. It is discharged by being able to show that you verify, which means the check has to leave a record: what was verified, against what, by whom, and when.
Firms find this irritating right up until the first time somebody asks. A per-document verification log is the cheapest professional-liability artefact available, and it falls out of the workflow for free if the workflow was built to produce it.
Where this leaves the reviewer
There is a version of this technology story in which the paralegal is competing with the tool. It is the wrong frame and it produces bad decisions.
The tool generates. The rules require that a person verifies. The verification requires somebody who understands the file, has the authority to reject, and is accountable for what they approved. That is a description of a senior reviewer, and demand for it goes up as generation gets cheaper, not down.
What has to fall is the cost of each individual check. That is an infrastructure problem, and infrastructure problems have infrastructure answers.
Field notes
Four of thirty scenarios in this series. All are composites, built from patterns we see repeatedly rather than from single named engagements. Firm details are illustrative.
12 · The firm that adopted a tool and then needed a control
Mixed plaintiff practice · 4 attorneys · review owned by the senior paralegal
The firm adopted generative AI for drafting and research and got a real speed gain. Then a partner caught a citation in a draft brief that did not exist.
What changed. The policy was written first — permitted tools, prohibited inputs, required verification steps — and owned by the senior paralegal rather than by a committee. Verification became a defined stage rather than a habit: every citation and every record reference checked against source before anything reached an attorney. Verification was logged per document, so the firm could demonstrate the supervision ABA Formal Opinion 512 requires rather than assert it. Anything that could not be traced to a source went back and never went out.
Outcome. The firm kept the drafting speed and put a verification layer between the draft and the signature. The policy is reviewed quarterly, because both the tools and the state bar guidance keep moving.
13 · The practice where nothing was written down
Personal injury plus workers' compensation · 3 attorneys · review owned by the lead paralegal
Two distinct case types ran through the same informal, undocumented workflow. New staff learned by shadowing, so onboarding took months and every departure took process knowledge with it.
What changed. An undocumented workflow cannot be delegated, standardised or scaled — and it cannot be supervised, which is now a sharper obligation than it used to be. A firm that cannot describe its own process in writing cannot demonstrate that it supervises that process. Both workflows were mapped end to end and encoded as the firm's criteria. The lead paralegal approved them, which is what made them reflect the practice rather than a vendor's idea of one.
Outcome. New staff ramp time dropped from months to weeks. Writing the process down also surfaced inconsistencies in the firm's own client communication that nobody had noticed.
14 · Rejections across three states
Personal injury, multi-state · 2 attorneys · review owned by the filing clerk, who is also a paralegal
The firm filed in three states and rejections were routine. Each cost a day or two of turnaround, and on files near a limitations deadline that margin was uncomfortable.
What changed. Rejections are rarely substantive — they are formatting, signature blocks, exhibit conventions and fee codes. A filing standards library was built per court capturing the requirements that actually cause rejections, so the paralegal was reviewing rather than reformatting. Every rejection was logged with a root cause and the library updated monthly. Filing authority never moved: she reviews and files under the firm's own credentials.
Outcome. Rejections became rare and traceable, and filings near a deadline stopped being a source of anxiety. A limitations calendar with dual reminders and a hard internal deadline sits well ahead of the statutory one.
15 · Four boards, four sets of rules
Workers' compensation · 2 partners · review owned by the hearing coordinator
The firm practised before four boards and commissions, each with its own forms, filing conventions and procedural quirks. Rejected filings were routine and nobody could say why.
What changed. Venue variation is a documentation problem disguised as a legal one, and the files themselves are heavier than they were: NCCI reports medical and indemnity claim severity each rising roughly 4 percent in 2025 even as lost-time frequency fell about 2 percent. A venue-specific checklist per board captured forms, formatting, service and timing. Medical evidence was assembled on a standing schedule ahead of each hearing date rather than in the week before.
Outcome. Rejections became rare and traceable, and the coordinator stopped relearning four sets of rules from memory.