AI incident response: what to do when a model gets it wrong in public

KEY TAKEAWAYS Summary by the editors

1. A company is generally expected to answer for what its AI tools say on its own channels: in Moffatt v Air Canada (2024), a Canadian tribunal rejected the airline's argument that its chatbot was a separate entity responsible for its own statements.
2. An AI incident plan should define severity levels, name an incident lead, give a way to switch the system off or restrict it quickly, and set who speaks publicly.
3. If personal data is involved, GDPR Article 33 requires the controller to notify the supervisory authority without undue delay and, where feasible, within 72 hours of becoming aware of the breach, unless it is unlikely to result in a risk to people.
4. For high-risk AI systems, Article 73 of the AI Act sets serious-incident reporting deadlines of 15 days in the general case, 2 days for widespread infringements and 10 days where a person has died, with high-risk obligations now applying from December 2027 and August 2028.
5. The most useful output of an incident is a recorded root cause, a changed control and an updated test, not only a public apology.

When an AI model gets something wrong in public, the first hour matters most: contain the problem, preserve evidence, decide who speaks and check whether a legal clock has started. The steps below suit a fashion brand or retailer whose chatbot, recommendation engine, pricing tool or generated content has produced a harmful, false or embarrassing output on a public channel.

Why do fashion companies need an AI incident plan?

AI outputs are customer-facing in more places than before: sizing assistants, service chatbots, product copy, campaign imagery and personalised offers. When one of them misfires, responsibility usually lands on the company that put it in front of the customer, not on the technology.

A widely cited example is Moffatt v Air Canada. According to an American Bar Association summary, a British Columbia tribunal ruled on 14 February 2024 that the airline was liable for negligent misrepresentation after its website chatbot gave a customer wrong information about bereavement fares. The tribunal rejected the argument that the chatbot was a separate entity responsible for its own actions, and held that the airline is responsible for all information on its website, whether static or chatbot-generated. The amount awarded was small, about 650 Canadian dollars plus interest and fees, but the reasoning is what matters. It is a Canadian decision, not an EU one, so it shows the direction of travel rather than binding law for European brands.

What counts as an AI incident?

Define it broadly in the plan, then grade it. An incident can be a factual error in a customer reply, an assistant that makes promises the company cannot keep, offensive or discriminatory output, a leak of confidential or personal data through a prompt or an output, a pricing error pushed to a storefront, or generated imagery that infringes rights or misrepresents a product. The AI Act has its own, narrower term, the serious incident, which applies to high-risk systems.

Read also

An AI acceptable-use policy for fashion teams: a checklist

What should the first hour look like?

Speed depends on preparation. The following sequence works for most teams, and each step should have a named owner in advance:

1. Contain: use a kill switch or fallback to pause the feature, route to human agents or revert to a previous version.
2. Preserve evidence: save prompts, outputs, logs, model and configuration versions, and screenshots with timestamps before changing anything.
3. Triage: assign a severity level and name an incident lead.
4. Assess scope: work out how many people saw the output, which channels and markets were involved, and whether personal data or minors are affected.
5. Check legal clocks: ask legal and data protection staff whether notification duties have started.
6. Decide communications: agree one internal source of truth and one external spokesperson.
7. Fix and verify: correct the cause, test with the failing case and related cases, and restore service in steps.

Which legal clocks may start?

Two regimes deserve a check, and a third is a matter of contract.

Data protection. If the incident involves a personal data breach, Article 33 of the GDPR requires the controller to notify the competent supervisory authority without undue delay and, where feasible, within 72 hours of becoming aware of it, unless the breach is unlikely to result in a risk to people's rights and freedoms. A processor must inform the controller without undue delay. A late notice must give reasons for the delay.

AI Act. Article 73 covers serious incidents involving high-risk AI systems. The general deadline is immediate reporting once a causal link is established or reasonably likely, and in any event within 15 days of becoming aware. For a widespread infringement the limit is 2 days, and where a person has died it is 10 days. Deployers of high-risk systems must under Article 26 inform the provider and relevant authorities of serious incidents. High-risk obligations for stand-alone systems now apply from 2 December 2027, so this mainly concerns tools in areas such as recruitment and worker management. Most fashion customer-service and merchandising tools are not in scope, but the transparency duties of Article 50 apply since 2 August 2026.

Contracts. Supplier and retailer agreements may contain their own notice periods for incidents affecting shared data or customers. Check them early.

How should a brand communicate in public?

Say what happened in plain words, say what is being done, say what customers should do and avoid blaming the technology or the vendor. A statement such as "our assistant gave incorrect information and we have switched it off while we fix it" is easier to stand behind than one that implies the system acted on its own. Where customers acted on wrong information, decide in advance how the company will make them whole, since the reasoning in the Air Canada case suggests that a tribunal may expect it to.

How do you learn from an AI incident?

Close each incident with a short written review: what happened, which control failed, what changed. Add the failing prompt and similar cases to a regression test set, adjust guardrails or review rules, and record whether the vendor needs to act. The NIST AI Risk Management Framework, which is voluntary, groups such work under its Manage function alongside Govern, Map and Measure, and can give the review a recognised structure.

Read also

What should boards ask about AI? A briefing for fashion directors

What should be prepared before launch?

Before any customer-facing AI goes live, confirm there is a way to disable it quickly, that logs are retained and reviewable, that a human fallback exists, that vendors have committed to notify you of defects and security events, and that the incident lead and spokesperson are named. Preparation of this kind costs little against the cost of an unplanned public correction.

Frequently asked questions

Is a company liable for what its chatbot says?

Courts and tribunals have started to say yes. In Moffatt v Air Canada, a Canadian tribunal held in 2024 that the airline was responsible for information given by its website chatbot and rejected the argument that the chatbot was a separate entity. It is not binding in the EU, but it signals how such cases may be argued.

Do I have to report an AI error to a regulator?

Only in defined cases. If a personal data breach occurred, GDPR Article 33 sets a 72-hour notification target to the supervisory authority. For high-risk AI systems, Article 73 of the AI Act sets serious-incident deadlines of 15, 2 or 10 days depending on the case.

What is the first thing to do when an AI tool misbehaves in public?

Contain it by pausing or restricting the feature, and save the prompts, outputs, logs and configuration before changing anything. Then triage severity, name an incident lead and check legal notification duties.

Who should speak publicly after an AI incident?

One named spokesperson, working from a single internal fact base prepared by the incident lead, with legal and data protection input. Avoid blaming the technology or vendor.

GuideThe complete guide to AI strategy for fashion companiesRead the complete guide