Brand Chatbot Went Wrong: Contain and Correct
Your brand chatbot went wrong and was screenshotted. How to contain it, correct it, sort out liability and see what search and AI answers then show.
When a brand chatbot goes wrong and the screenshot spreads, the response has four parts, in this order: stop the bot from repeating the answer, state the correction in one place you control, settle what you owe the people who relied on it, and manage what search engines and AI assistants then say about the incident. The company is treated as the speaker of whatever its chatbot says, so the plan starts from ownership, not from blame.
The damage rarely comes from the bad answer alone. It comes from the second story: a slow reaction, a statement that blames the software, or a refusal to honor what the bot promised. A single wrong sentence from a bot becomes a news item when the company sounds evasive, and that news item is what people find months later when they search your name or ask an AI assistant about you.
This guide covers the incidents where your own customer-facing AI assistant produced a wrong, offensive or unauthorized answer. It does not cover what third-party assistants say about you, which is a different problem with different levers. It is written for founders, communications leads and counsel who have a screenshot in front of them and need a sequence, not a lecture on AI risk.
- Contain first: disable the failing behavior or the whole bot within the hour, and keep the evidence before you change anything.
- You will be treated as the author of what your bot says. A tribunal has already rejected the argument that a chatbot is a separate entity.
- Correct once, in plain words, on a page you own, and link every later statement to it.
- Honor reasonable reliance by affected customers. A small payment is cheaper than a second story about refusing it.
- The incident will rank for your brand name for a long time unless your own page answers it better and the facts stop changing.
Four kinds of chatbot incident, and why the type sets the response
The first decision is classification, because the response, the exposure and the fix differ by type. A chatbot incident is almost always one of four things: a wrong statement of fact or policy, an offensive or off-brand output, an unauthorized commitment, or a manipulated output that a user engineered on purpose.
Related readWikipedia Page for My Company: Notability and Disclosure| Type | What it looks like | Main exposure | First move |
|---|---|---|---|
| Wrong fact or policy | The bot states a refund rule, price or feature that does not exist | Customers who relied on it, consumer claims | Publish the correct rule, identify who relied |
| Offensive or off-brand | Swearing, insults, criticism of your own company | Reputation, not money | Pull the feature, short apology, explain the cause |
| Unauthorized commitment | The bot agrees to a price, discount or promise | Contract arguments, goodwill | State what is and is not binding, review each case |
| Engineered manipulation | A user steers the bot with crafted prompts, then posts the result | Mixed: reputation and sometimes a claim | Say what happened without mocking the user, harden the bot |
Public cases show each type. In January 2024 a delivery company's assistant was reported to swear at a customer and describe the company as the worst delivery firm. The company said an error occurred after a system update and that it had disabled the AI element. In December 2023 a car dealership's website bot was reported to agree to sell a vehicle for one dollar and to call the offer legally binding, after a user steered it to agree with any statement the user made. The bot's provider was reported to have shut it down on the dealership's site. In a later case, a software company's support bot was reported to tell a user that a one-device-per-subscription rule applied, and the company's co-founder said publicly that no such policy existed; the company was reported to have apologized and refunded affected users.
These are reported incidents, described here from news coverage and not from company records. The pattern is what matters: the same four types recur, and the companies that recovered fastest were the ones that acted on the type in the first day.
Why the type matters for liability
A wrong fact about policy invites reliance, and reliance is what courts and tribunals look at. An offensive output usually causes no loss to anyone, so the exposure is reputational. A manipulated output sits in between: a user who deliberately engineers a one-dollar "deal" has a weak claim, but a user who innocently follows a bad refund instruction has a strong one. Classify honestly before you decide what to say.
Related readService Outage Brand Reputation: What Stays in SearchThe first hour: contain before you explain
Containment means the bot stops producing the problem, and the evidence is preserved first. The order matters, because changing the bot before saving its logs destroys the record you need for the correction, the customer review and any legal question.
- Save the evidenceExport the full conversation logs, the system prompt, the knowledge base version and the vendor configuration as they stand, with timestamps. Capture the screenshot as posted, with its URL and date.
- Switch off the failing behaviorIf you can restrict the topic (refunds, pricing, commitments) do that. If you cannot isolate it, disable the chat widget and route visitors to a form or a human queue.
- Search for the same failurePull every conversation that touched the same topic or used the same trigger. One screenshot usually means more instances exist, and you want to find them before a journalist does.
- Name one ownerOne person decides on the public statement, one on customer remedies, one on the vendor. A committee in the first hour produces silence.
- Tell the vendor in writingNotify any third-party provider, cite the contract, and ask them to preserve their own logs. Do this before any public statement attributes the cause.
A partial shutdown is acceptable and common, but be accurate about it. If you only turned off one function, say so. A statement that says the bot is offline when it is still answering other questions creates a second screenshot.
Do not argue with the person who posted it
The person who posted the screenshot is not your audience. Your audience is everyone who will see the screenshot later. Do not dispute their motives, do not call a manipulated output a hoax if the bot really said it, and do not ask for the post to be taken down as a first move. A takedown request on a true screenshot is the classic way to turn a small post into a bigger one.
Who is liable for what the chatbot says
Assume you are liable for your bot's statements to your customers, and then work out whether any of that can be passed on to a vendor. The clearest ruling so far is a small-claims decision from British Columbia, Canada, and it is the one counsel will reach for first.
In Moffatt v. Air Canada, 2024 BCCRT 149, decided on February 14, 2024, a customer asked the airline's website chatbot about bereavement fares. The chatbot said he could apply for the reduced rate within 90 days after the ticket was issued. The airline's actual policy did not allow that after travel. The tribunal found that the airline did not take reasonable care to ensure its chatbot was accurate and that the customer reasonably relied on it. It rejected the airline's argument that it could not be held responsible for what the chatbot said, which it described as a remarkable submission. The airline had also pointed to a correct policy on a separate page of its site. The tribunal still held that the chatbot is part of the airline's website and that the airline is responsible for the information on it, whether it sits on a static page or comes through a chatbot.
Related readDeepfake scam ads using your CEO: the response planSource: the tribunal's published decision in Moffatt v. Air Canada, 2024 BCCRT 149, which awarded damages of $650.88, interest of $36.14 and fees of $125, payable within 14 days.
Read the amount with care. The loss was a fare difference, so the damages were small. The cost to the airline was the headline, not the money. The lesson is that a company cannot treat a chatbot as an independent speaker, and a correct answer elsewhere on the site does not cancel a wrong answer in the chat.
The vendor question
If a third party built or hosts the bot, the allocation sits in your contract, not in the public narrative. Counsel will look at four things: who controls the prompt and the knowledge base, whether the contract has an indemnity for outputs, what the vendor warrants about testing, and what liability cap applies. In most disputes the customer-facing company stays exposed to its own customers whatever the vendor agreed to pay it. The vendor relationship is a recovery question after you have dealt with the customer, not an excuse to give the customer.
The EU transparency rule
For brands serving people in the EU, Article 50 of the AI Act says providers must design AI systems that interact directly with people so the people are informed they are interacting with an AI system, unless that is evident from context. Article 50 applies from August 2, 2026, according to the text on that site, and law-firm commentary on the Digital Omnibus amendments says the chatbot disclosure duty was not postponed; only a transition period for the marking of synthetic content was added. Read the current text and any later amendment on the official site before relying on it, and ask counsel whether you are the provider or the deployer under the Act. A visible disclosure also helps in practice: a customer who knows it is a bot is less able to claim they believed they were reading company policy from a person.
Related readBrand Reputation During a Strike or Labor DisputeCorrect the record once, on a page you own
The correction is a single page on your own site that says what the bot said, what is true, and what you did about it. Every later statement, reply and press response links to that page. A single canonical page is what search engines and AI assistants can lift, and it stops you from issuing five slightly different versions.
A good correction has five elements, in this order.
- What happened, in one sentence, without adjectives.
- What is actually true: the real policy, price or rule.
- What you have changed: the function restricted, the review added, the disclosure added.
- What affected customers should do, with a route that does not need a lawyer.
- A contact route, which can be a form.
Keep the cause section short and factual. If a system update broke a guardrail, say that. If the cause is still being investigated, say that, and say when you will update the page. Do not name the vendor as a way of shifting blame, and do not describe the output as the software acting on its own. Readers hear "the bot did it" as "nobody is responsible."
What a good statement sounds like
The tone is a factual notice, not an apology tour. Say that the answer was wrong, that the company is responsible for it, and that the correct information is on the page. Avoid phrases such as "we take this seriously" and "isolated incident" unless you have the logs that prove it was isolated. The logs from the first hour are what let you say how many conversations were affected, and a specific number is more credible than reassurance.
Handle the customers who relied on the answer
Review every conversation where the bot gave the wrong answer and someone acted on it. The decision rule is reasonable reliance: did a customer do something, or fail to do something, because of what the bot said, and would a reasonable person have trusted it. If yes, put them in the position the bot described or compensate the difference, in writing, without making them prove it twice.
Related readUDRP Complaint for Cybersquatting: A Brand GuideThis is where the money is smaller than the instinct suggests. In the Air Canada decision the award was the gap between what the customer paid and what the bot's answer implied. A similar gap in your case is likely to be modest, and paying it quickly removes the one fact a journalist can add to the story: that the company refused.
For unauthorized commitments the same logic applies with a different outcome. A price that is plainly absurd, obtained by a user who told the bot to agree to everything, is not a reasonable-reliance case. Say so calmly in the correction page: the bot cannot set prices, and a price is only binding when confirmed through the checkout. Do not attack the user. Counsel should confirm that position for your jurisdiction before you publish it.
What shows in search and AI answers afterwards
Expect the incident to attach to your name in three places: news and social results for your brand plus the word chatbot, autocomplete and related searches, and the summaries AI assistants write when asked about you. The incident does not need to be large to persist, because there is little other content about your bot to compete with it.
Searching your brand name together with chatbot, AI, glitch or apology will show you the current picture. Run the same questions in the main AI assistants and answer engines, not in only one, and save the answers with the date. The first week usually shows the original coverage and the social posts. After that, the results that tend to remain are the two or three best-written articles and, for an incident with a legal angle, the legal commentary.
Related readEvent Failure Brand Reputation: The Response Sequence| Phase | What usually shows | What you control | Action |
|---|---|---|---|
| Days 1 to 3 | The post, reposts, first news items | The bot, your channels | Contain, publish the correction page |
| Week 1 to 2 | Trade and general coverage, commentary | Your page, requests for updates to factual errors | Ask publishers to fix errors of fact only |
| Weeks 3 to 8 | Coverage settles, AI answers begin to summarize | Owned pages, structured facts | Test assistants, strengthen your page |
| Months 3 and on | Incident appears as a line in answers about your brand | What you did next | Publish the fix and outcomes, monitor monthly |
These phases are experience, not a schedule. A busy news cycle or a large audience moves them faster, and a legal claim extends them.
What you can and cannot change
You can ask a publisher to correct a factual error, such as a wrong date, a wrong description of the cause or a wrong figure. You can publish your own account and make it the clearest source. You can report an AI answer that states something false, using the assistant's feedback route. You cannot lawfully expect accurate coverage of a real incident to disappear, and an attempt to force it tends to produce a second article about the attempt. We do not suppress accurate reporting. We help brands correct what is wrong and make the accurate, complete account the easiest one to find.
For AI answers, the practical lever is the evidence they draw from: your correction page, your help center and your published policy. If your help center still contains the wrong rule that the bot repeated, fix it, because the bot may have learned the error from your own documents. After that, test the assistants again. When an answer repeats a false detail about the incident, a precise report that cites your correction page is more effective than a general complaint. Our work on what AI assistants say about you starts from this same audit.
Make the fix visible, and keep watching
The best long-term answer to the incident is a short record of what changed. Publish what you restricted, what you now test before release, how a customer reaches a person, and how the bot is labeled as AI. A concrete fix gives journalists a next line to write and gives search engines a more recent page than the screenshot.
Related readPricing Backlash: How a Brand Should RespondBuild a small release routine so the next failure is smaller: a fixed list of high-risk topics (refunds, pricing, legal, medical, safety) that the bot must refuse or hand to a person; a test set of hostile prompts run before each update; a human review of a sample of conversations each week; and an alert when the bot's answers on those topics change. Two of the cases above followed a system update or a user pushing the bot off its task, which is why the test set matters after every update, not only at launch.
Then watch. Run the same brand and chatbot searches and the same assistant questions monthly, record what changed, and keep the correction page current. Monthly monitoring of search and AI answers is the routine we set up for this, and the same log tells you when the incident has faded and when it returns.
Your next step if this is happening now
Do the evidence export and the shutdown of the failing function before anything else, then write the correction page. If the incident is moving quickly, bring in the people who can handle press, counsel and the vendor in parallel.
Send us the screenshot, the logs you have and the links where it is spreading. We will review what shows in search and AI answers for your brand and tell you what to correct first. Start with a private audit at our audit page.
Questions people ask
Can a company be held responsible for what its AI chatbot says?
Yes, in the case that has been decided most clearly. A British Columbia tribunal held that an airline was responsible for a wrong answer its website chatbot gave and rejected the idea that the bot was a separate entity. The decision is not binding elsewhere, but counsel treats it as a strong signal. Ask your own lawyer how it applies where you operate.
Related readIPO Reputation Readiness: The Pre-Listing AuditShould we take the chatbot offline completely?
Not always. If you can restrict the topic that failed, a narrower shutdown keeps service running and shows control. If you cannot reproduce or isolate the failure, take the bot offline and route visitors to a form or a person. Say accurately what you did. A statement that overstates the shutdown creates a second problem.
Does a disclaimer on the chat window protect us?
A disclaimer helps a little and protects less than companies expect. In the Air Canada decision the airline pointed to a correct policy on another page, and the tribunal still held it responsible for the chatbot's answer. A clear notice that the user is talking to an AI is still good practice, and may be required in some places, such as under the EU AI Act.
Can we get the screenshot or the news articles removed?
Usually not, if they are accurate. You can ask for factual errors to be corrected and report content that breaks a platform's own rules, such as harassment or leaked private data. A removal request for a true post tends to draw attention. The stronger route is a clear correction page and a visible fix.
How long will the incident show up in search?
It depends on the coverage, not on a fixed period. A minor incident with little coverage fades in weeks, and one that reached national news can remain in results for years. What changes the picture is newer, better content about what you did next, and steady monitoring so you can see the effect.
What if the vendor caused the problem?
You still answer to your customers first. Handle the customers, publish the correction, then go to the vendor under the contract: preservation of logs, indemnity, testing warranties and the liability cap. Avoid naming the vendor publicly as the cause before counsel has reviewed the contract and the facts.
Will AI assistants repeat the story about our chatbot?
They often will, because news coverage is a main source for answers about a company. The accuracy of those answers depends on what the sources say. A precise, well-linked correction page and a clear record of the fix give assistants a better source to quote, and you can report specific false details through each assistant's feedback route.



