Incident response when wrong AI output reaches a court or client
By the ChamberOne team 5 minute read
Someone on the team notices that a case cited in yesterday's submissions does not say what the submissions claim, or perhaps does not exist at all. The temptation is to fix it quietly and hope nobody noticed. That is almost always the wrong instinct. The firms that come through these episodes best are the ones that have already decided, before anything goes wrong, who does what in the first hour.
Key points
- Name an incident lead in advance and start a written incident note straight away.
- Establish the full scope before you speak, including other matters produced the same way.
- Tell the court and the client promptly and plainly, and do not blame the tool.
- Notify your insurer early and check regulator guidance for your own jurisdiction.
- End the review with specific, enforceable changes, not a reminder to be careful.
Decide who is in charge in the first hour
Incidents go badly when everyone is responsible and nobody is accountable. Name an incident lead in your policy now, usually a partner with authority over the matter, and a deputy for when the lead is away. The lead does not have to fix everything. They have to make sure the steps below happen, in order, and that someone writes down what was decided and why.
In a small firm the lead may well be the person who signed the document. That is a conflict of interest in all but name. If it cannot be avoided, ask another partner, or an outside adviser, to take the decisions about what to disclose, so that nobody is judging their own work.
Start a short incident note immediately. Record who found the problem, when, what is wrong, and which documents went where. Memory degrades quickly under stress, and the note will matter to the court, the client and any insurer.
Establish exactly what went wrong
Before you tell anyone anything, work out the scope. One bad citation is a different problem from a document drafted largely by a tool and filed unread. Pull the draft history if you have it, and talk to the people involved without blame. You want the facts, and people stop giving them if they expect to be punished for honesty.
- Which documents contain AI-generated material, and which parts?
- Who reviewed them, and what did that review actually involve?
- What was sent, to whom, and when: court, opponent, client, third party?
- Is any other live matter produced the same way and therefore at risk?
Tell the court promptly
If the error is in something filed or handed up, the court needs to hear it from you, quickly and plainly. Courts in several countries have shown more patience with lawyers who correct an error themselves than with those who are found out. A short, factual letter or application is usually better than a long defence.
What you say should be accurate and unembellished: what the error is, how it came about, what has been done to correct it, and what you propose to do about the document. Avoid minimising language. Do not suggest the tool was to blame, as the responsibility for what is filed rests with the lawyer. Take advice from your regulator's guidance or your professional indemnity broker on your obligations, because they differ by country and by court.
- State what the error is and how it came about
- Say what has been done to correct it and what you propose for the document
- Use minimising language
- Suggest the tool was to blame for what was filed
Inform the client
Tell the client yourself, before they hear it from the court or the other side. Explain what happened in plain words, what it might mean for their case, and what you are doing. Be ready for difficult questions about cost, and consider whether the client should bear any of the time spent on the correction. Many firms decide they should not.
Then record the conversation on the file. If the error could affect the client's position, discuss whether they want to take independent advice. This is also the moment to think about whether the matter should continue with the same team.
Correct the record and the work product
Withdraw or replace the faulty document through whatever process the court allows. Check every other authority and quotation in the same document, not just the one that was found, because errors from AI tools tend to cluster. If the same material went to the other side, tell them too.
Then look at anything derived from the document: a skeleton argument built on a draft, a client advice note that repeated a claim, a template that now contains the flawed text. Fix the source, not just the copy.
Notify your insurer and take advice
Professional indemnity policies usually require prompt notice of circumstances that might give rise to a claim, and late notice can cause problems. Read your policy or call your broker the same day, even if you think the matter is minor. Ask whether your policy treats AI-related errors in any special way, and keep records of everything you send.
If a regulator may need to hear about the matter, find out through its published guidance or a call to its helpline, rather than guessing. Do this before you say anything public.
Looking after the person involved
Behind most incidents is someone who was tired, rushed or trusted the tool too much. Handle the conversation privately and fairly, and separate the facts from any employment question. A culture in which people hide mistakes is far more dangerous than one in which they report them early, because the next incident will be reported late, when options have narrowed.
Offer support to the person, and make sure they are not left to deal with the court or the client alone. Supervisors share responsibility for what was filed, and the review should say so honestly.
Review the controls, then change something
An incident review that ends with 'be more careful' has achieved nothing. Look for the specific gap that let the error through: no rule about checking authorities against an official source, a junior working alone on a tight deadline, a tool used that the firm had not approved, or a supervisor who signed without reading.
Pick two or three changes you can actually enforce. Examples include a mandatory step where every authority is opened and read in an official or trusted source, a log of which tools were used on which documents, and a clear rule on who may sign off a filing. Then brief the whole firm on what happened, with the individual's name removed, so the lesson is shared and the person who found the problem is thanked rather than shamed.
This article is general information about the law, legal practice and technology, current when it was published. It is not legal advice. Laws and guidance change: check the sources linked above and take advice on your own circumstances from a lawyer qualified where you are.
Read next
Client confidentiality and AI tools: a checklist before you paste anything in
The questions to answer about retention, training use, access, hosting and sub-processors before client information goes into any AI tool.
5 minute readShadow AI in law firms: staff are already using it, so what next?
Why bans fail, how to find out what is being used, and how to offer an approved alternative with clear rules.
5 minute readWriting a law firm AI policy: an outline you can adapt
Six sections that a short, usable AI policy needs, with the questions each one should answer.
4 minute read