AI incident communications is the response plan for a public failure caused by an AI system, such as a chatbot that gives customers wrong information or a deepfake that impersonates an executive. Treat the company as responsible for what its AI says, publish a dated statement with confirmed facts, preserve the logs, and tell affected people what to do next.
Last updated November 5, 2026.
Key takeaways
- Courts and tribunals have already held companies responsible for what their chatbots say, so "the AI did it" is not a defense or a statement.
- AI incidents fall into four types: wrong answers, harmful output, impersonation and fraud, and wrong claims about the company inside other AI products.
- The first hours need logs, a pause on the faulty system and one confirmed statement.
- Plan for the incident before launch, with a named owner for every AI system that talks to the public.
What counts as an AI incident?
An AI incident is any event in which an AI system causes public harm, confusion or loss, and the company must explain what happened. Four types cover most cases, and each has a different first audience. The table lists them with the first message each needs.
| Incident type | Example trigger | First audience | First message |
|---|---|---|---|
| Wrong answer | A chatbot misstates a policy, price or deadline | The customer who relied on it | Correct the facts and state the remedy |
| Harmful or off-brand output | A customer-facing assistant produces offensive or unsafe text | Customers and the social team | Pause the system and confirm what happened |
| Impersonation or fraud | A deepfake video or voice poses as an executive | Finance, security and law enforcement | Verify through a known channel and freeze transfers |
| Wrong claims in AI answers | An AI assistant repeats a false claim about the company | Communications and the source publisher | Trace the cited source and correct it |
The first three are incidents the company must handle directly. The fourth is covered in 5W's AI crisis monitoring guide, because it depends on what other systems retrieve.
What did the Air Canada chatbot case show?
The case showed that a company is responsible for what its chatbot tells customers. In February 2024 the British Columbia Civil Resolution Tribunal, in Moffatt v. Air Canada, found the airline liable after its chatbot told a grieving customer he could claim a bereavement fare retroactively. The airline's policy did not allow that, and the airline refused the refund, according to a Dentons summary of the ruling.
The tribunal decided the matter on February 14, 2024. It rejected the argument that the chatbot was a separate source of information, held that the airline did not take reasonable care to ensure the chatbot was accurate, and awarded the customer about C$812 in total, including interest and fees. The money was small. The public lesson was large, because the ruling was widely reported as a test of who answers for an AI system.
What to take from it: a statement that blames the technology invites the same criticism the airline received. State what the system said, correct it, make the customer whole and describe the change.
What did the Arup deepfake fraud show?
The Arup case showed that a convincing deepfake can defeat an ordinary approval habit, and that a company can contain the communications damage by confirming facts early. In February 2024, Hong Kong police described a fraud in which a finance employee joined a video call with people who looked and sounded like the chief financial officer and colleagues, then sent HK$200 million, about US$25.6 million, in 15 transfers. Arup confirmed in May 2024 that it was the company involved, according to CNN's report.
Arup said it had notified Hong Kong police in January and confirmed that fake voices and images were used. Its chief information officer, Rob Greig, said attacks of this kind were rising sharply, and that he hoped the experience would raise awareness. That framing turned a loss into a warning to others.
What to take from it: confirm the facts with police and finance first, say plainly what was used, and tell peers what to check. The fraud itself is a security matter, and communications supports it with clear, calm language.
What should happen in the first 24 hours?
The first 24 hours should stop the harm, preserve evidence, confirm facts and issue one statement. The table orders the work for a chatbot incident, and the same order applies to a deepfake with the security and finance steps added.
| Window | Actions | Owner |
|---|---|---|
| Hours 0 to 2 | Pause or limit the system, save the logs and screenshots, and name one decision owner | Product or security lead |
| Hours 2 to 6 | Confirm what the system said, to whom and for how long. Identify affected customers | Product lead with legal |
| Hours 6 to 12 | Draft the statement and the customer remedy. Brief support and social staff | Communications lead |
| Hours 12 to 24 | Publish a dated statement, contact affected customers directly and set the next update time | Executive lead |
Preserving the logs matters because the question that follows is always how long the error lasted and who saw it. The first-day discipline is the same as in the data breach communications plan, and the sequence sits inside the wider crisis PR guide.
What should the statements say?
Each AI incident needs a short statement that follows the four-part holding statement: what happened, what the company is doing, what people should do and when the next update arrives. The templates below are starting points for review by counsel. See the holding statement examples for the full structure.
Chatbot error
Between [date] and [date], our [assistant name] gave some customers incorrect information about [policy or price]. The correct information is [facts]. We have paused the assistant while we review it. Customers who relied on the incorrect answer will receive [remedy], and anyone affected can contact us at [contact]. We will update this page at [time].
Executive deepfake
On [date], we learned that fraudulent video or audio impersonating [role] was used to request [action]. This was not a request from our company. We have reported the matter to [authorities] and are reviewing our approval procedures. We will not act on payment or access requests that arrive only by video, voice or message. Please contact [verified channel] with any questions.
Leave out the cause of the model's error and any claim about how rare it is until the review is complete. A statement that explains the model before it fixes the customer's problem reads as an excuse.
Who should own AI incidents?
An AI incident needs a named owner for each AI system that talks to the public, plus a standing link between product, legal and communications. Without that link, the first hour is lost deciding whose problem it is.
- System owner. Knows how the system works, can pause it and can pull the logs.
- Legal counsel. Judges liability, notification duties and what can be said about cause.
- Communications lead. Writes the statement and manages media and social replies.
- Executive lead. Approves the statement and decides whether to take the system offline.
Test the link in a tabletop exercise. The crisis simulation guide includes a scenario format, and 5W's guide to how AI is transforming crisis communications explains why a human with the right knowledge must approve every statement an AI helps draft.
What if AI products say the wrong thing about your company?
When an AI assistant repeats a false claim about a company, the fix is usually at the source it cites, not inside the assistant. Run the same prompts each month, record the answer and the cited sources, and correct the source where possible.
This is a different problem from a company's own AI failing. The GEO measurement guide explains what the data can and cannot show, and the second crisis analysis shows how outside documents shape what search and AI summaries repeat. No agency can guarantee how an AI product answers.
What should a company do before an AI incident?
Before launching any public-facing AI system, a company should decide who owns it, what it must never say, how to pause it and what the customer remedy will be. Four steps are enough to start:
- List every AI system that speaks to customers, with an owner and a pause procedure.
- Write a short list of topics the system must route to a human, such as refunds, safety and legal questions.
- Keep logs long enough to answer how long an error lasted.
- Add the two templates above to the crisis plan and run one exercise a year.
Companies that want help building the plan can work with 5W's crisis PR practice, and the crisis PR cost guide explains how the work is priced.
Frequently asked questions
Is a company responsible for what its chatbot says?
In the Air Canada case, a British Columbia tribunal held the airline liable for its chatbot's incorrect answer and rejected the idea that the chatbot was separately responsible. Companies should assume customers and courts will treat chatbot statements as the company's own.
What should a company do first when its chatbot gives a wrong answer?
Pause or restrict the system, save the logs, confirm what it said and to whom, and correct the information for affected customers. Then publish a short dated statement and set a time for the next update.
How should a company respond to a deepfake of an executive?
Report it to law enforcement, tell finance and security teams, verify requests through a known channel and publish a short statement that the request was not from the company. Do not act on payment requests that arrive only by video or voice.
Should a company blame the AI vendor in its statement?
Not at first. Customers hold the company they dealt with responsible. Describe the fix and the remedy, and leave vendor disputes to counsel and private channels.
Can a PR agency control what AI assistants say about a company?
No. An agency can publish accurate, dated facts and correct the sources that assistants cite, but it cannot guarantee how an AI product answers.





