Two measurements, two results
First: finding citations. A pre-registered, peer-reviewed study tested the three market-leading research tools built specifically for legal work on 202 legal questions — not general chatbots, then, but systems wired to their own legal databases.
| Tool | Accurate | Hallucinated | Incomplete or refused |
|---|---|---|---|
| A | 65% | 17% | 18% |
| B | 41% | 33% | 25% |
| C | 19% | 17% | 62% |
It matters what "hallucinated" means here: either the tool states something false, or it cites the wrong authority. The second is the more dangerous one — the citation exists, it just does not say what the text claims it says. Looking the citation up and finding it does not catch that.
Second: drafting contracts. A 2025 comparative study handed 30 real contract-drafting tasks to 13 AI solutions and to a human control group — in-house commercial lawyers with ten years of experience on average. The second round was double-blind.
| Human | AI average | |
|---|---|---|
| Share of reliable output | 56.7% | 57% |
| Usefulness (on a nine-point scale) | 7.53 | 7.25 |
| Time spent | ~13 minutes | under 1 minute |
(This study is not a peer-reviewed academic publication, and several of the vendors measured took part in it — the numbers are worth reading with that caveat.)
Two things stand out. One is that the human baseline is not flawless either: even experienced lawyers produced output the benchmark counted as reliable in only a little over half the cases. The other is that the type of task decides it. Putting something into words when the structure is already known — that works. Working out what can be cited — that does not.
The invented citation: 2,039 documented cases
For this there is no estimate; there is a register.
A continuously maintained database collects the court cases in which an AI-invented citation ended up in a filing. As of 12 September 2026: 2,039 cases. The most recent entry is two days old.
| Country | Cases |
|---|---|
| United States | 1,396 |
| Canada | 217 |
| Australia | 110 |
| Israel | 57 |
| Brazil | 41 |
| India | 15 |
| France | 13 |
| Germany | 11 |
The methodology is stricter than you would expect, and that is what makes the number usable: a decision only gets in if the court expressly found that a party had relied on hallucinated content. An allegation on its own is not enough. So 2,039 is a lower bound.
The sanctions are no longer symbolic. One tally puts the fines imposed for AI-generated false citations in the first quarter of 2026 at $145,000 or more; the largest single item was $30,000 in an appellate court, against two lawyers, topped up with attorney's fees and double costs — there, more than two dozen citations turned out to be wrong, misread or non-existent. For a single day: on 31 March 2026, seventeen US court decisions mentioned suspected AI hallucination.
This number grows every week — if you see 1,200 or 1,600 quoted somewhere, that is an earlier snapshot of the same database.
What the rules ask for
This is not legal advice and does not stand in for it — but the professional-body documents are clear enough that anyone who has a machine draft documents can learn from them.
The October 2025 guidance from the European bar associations states three things:
- the output has to be checked before it goes into the work;
- client or confidential data does not go into the interface without appropriate safeguards — the reasoning being that what you type in can be stored and used by the provider for training;
- responsibility stays where it was: the human answers for the work, for the advice and for the claims.
They also name two risks worth remembering: automation complacency, and the model's tendency to nod along with whatever you want to hear.
And one sentence in their March 2026 technical annex reaches far beyond the legal profession: whether your prompt is used for training is not decided by the architecture of the model but by the contract, the technical configuration and the provider's policy.
The US ethics opinion (2024) adds two further points, and both carry over to any profession:
- boilerplate is not consent. A general clause in the contract that "authorises the use of AI" is not enough — you have to say why you are using it and with what risk;
- on hourly work you may bill the time you actually spent. If the machine does it in one hour instead of three, one hour goes on the invoice. Commercially this cuts deeper than any technical question.
National professional regulation, though, does not move at one pace. On the same question — what a lawyer may do with generative AI — two member states give two completely different answers today. Open the one that applies to you:
Hungary — there is no comprehensive bar rule
In Hungary, as of September 2026, there is no comprehensive bar rule covering lawyers' use of generative AI. There is a single, narrowly scoped professional position, dated 13 July 2026, and even that is specifically about Hungarian-language speech-to-text, not about document drafting.
That is news in itself: the EU-level guidance exists, the national detail does not. A Hungarian lawyer using AI today can lean on the European bar federation's guidance and on the general ethics rules — not on a circular from their own bar.
Spain — a binding bar interpretation, and four court fines
The national council of Spanish bar associations issued a binding interpretation of the code of conduct in April 2026 on the use of generative AI. On that reading the use is lawful in itself, but the model "must not operate as a substitute" for the lawyer, only as an auxiliary function under human supervision. Responsibility is exclusive: a technical failure of the tool is no excuse, and the developer company can never be held responsible in professional ethics.
The sanction is not theoretical: the interpretation classes taking over citations without verification as a serious disciplinary offence against the dignity of the profession, and as very serious where the conduct is gross or repeated. The breach is committed when the document is filed — voluntary correction in time can still prevent it. Madrid's bar has a separate good-practice guide on top of that, from October 2025.
And there is something to learn from: Spanish courts are fining lawyers one after another. In February 2026 a regional high court imposed a fine of 420 euros for citing 48 non-existent judgments — setting the amount expressly at "half the approximate cost of an annual subscription to a legal AI tool". In July the same court set 840 euros for 25 invented citations, and another high court 1,800 euros for 24. All three were referred to the local bar.
Judges were not left out either: under an instruction of 28 January 2026 from the national judicial council, a judge may use only those AI systems in judicial activity that are provided by the justice administration or by the council itself — public, general-purpose tools are therefore off limits. Draft decisions too may only come from those, with a complete and critical personal review.
Six things a human always has to review
This is not opinion — behind each point stands one of the sources cited in this article.
- Every citation. A statutory provision, a decision, a standard, an earlier contract. And it is not enough that it exists — you also have to check that it says so.
- Every number, date, deadline and amount. These are the easiest mistakes to check and the most expensive ones to leave in.
- Everything that calls for professional judgement. The ethics opinion expressly forbids pushing that responsibility onto the machine.
- What is allowed in at all. Feeding client data in is not a technical step but a decision taken in advance.
- The presence of the mandatory elements. AI does not flag what is missing — it only shapes what is there.
- Tone and deliberate ambiguity. In the experiment on patent drafts these were two separate review dimensions, not a question of style.
The commonest error is not the lie but the omission
Look again at the third data row of the very first table: 62% incomplete or refused answers.
That is the best evidence for the thing we expect least in an office setting. AI does not fill the document with lies. It leaves things out. A section, a clause, a mandatory element — and the output still comes out rounded, readable and confident.
This is why reading through what it wrote is not enough. It has to be compared against a checklist that comes from your template, not from its output. What the machine never wrote will not catch your eye while you are reading its text.
What happens to expertise
A three-month randomised field experiment published in September 2026, run across eleven firms with 133 patent attorneys, measured draft quality under blind review along five dimensions.
Quality improved significantly. But the authors put it more precisely: most of the improvement came from avoiding very bad work, not from reaching the best. AI raised the floor, not the ceiling.
And the most interesting finding: on certain tasks the experienced lawyers kept the improvement even without AI once the experiment was over. Among the juniors there was no consistent skill gain — some got better, some got worse.
That is the strongest argument for having a human review the draft: not only because of the mistake, but so that someone is still able to spot the mistake five years from now.
The recipe
1. THE TEMPLATE IS YOURS, NOT THE MACHINE'S
Template first, generation second. The list of
mandatory elements lives in a file, not in your head.
2. DATA, NOT A REQUEST
Hand over the facts: parties, amounts, dates,
service. Do not ask it to "write me a contract".
3. ASK FOR A QUOTE WITH EVERY CITATION
If it cannot attach the wording word for word,
the citation goes out.
4. COMPARE IT TO A CHECKLIST, DO NOT JUST READ IT
An omission does not show up on a read. The list does.
5. NUMBERS IN A SEPARATE PASS
Date, amount, deadline, percentage. One by one,
from the source, not from the draft.
6. WHAT IS CONFIDENTIAL DOES NOT GO IN
Client name, personal data, trade secret — only where
the contract expressly provides for it.
Point two makes the biggest difference to quality. Ask it to "write me a service contract" and you get generalities; give it a template loaded with the facts and you get a usable draft.
Common misconceptions
- "A tool built for legal work does not hallucinate." Those are exactly the tools that were measured: they hallucinated between 17% and 33% of the time, wired to their own legal databases.
- "I looked the citation up, it exists." The more dangerous kind of error is not the source that does not exist but the one that does and says something else.
- "If it were wrong, I would see it." The commonest error is omission. Without a checklist, the missing section does not stand out.
- "Human work would be flawless." It is not. In the same measurement the human baseline was 56.7%. The question is not flawlessness but who answers for it.
- "Faster, therefore cheaper." At task level, yes: under one minute instead of thirteen. At working-time level the strictest measurement shows 2.8% of total hours — because the checking takes part of the gain back.
What to take away
In document drafting the machine matches a human at wording and does not at citation. In practice those two jobs are mixed into the same text, which is exactly why they have to be pulled apart: the sentences you can hand over, the support behind them you cannot.
And responsibility does not move. Whoever files the document answers for it — the European guidance and the US ethics opinion both say so in the same terms, and 2,039 court cases show that it is being taken seriously.
Which step of reading and research is worth handing to a machine is measured out in the faster reading post, with the numbers; the daily office time saving, broken down by type of task, belongs to the real time savings post.