Added an AI chatbot to your website? You’re responsible for what it says

Are you legally responsible for what your chatbot says?

Usually, yes. The leading example is Moffatt v. Air Canada, decided by British Columbia’s Civil Resolution Tribunal in February 2024. Jake Moffatt asked the airline’s website chatbot about bereavement fares after his grandmother died. The bot told him he could book at full price and claim the discount afterwards. The airline’s actual policy said otherwise, and it refused the refund.

Air Canada argued that the chatbot was effectively a separate legal entity responsible for its own actions. The tribunal called this a “remarkable submission” and rejected it. The chatbot was part of Air Canada’s website, the airline had not taken reasonable care to make it accurate, and customers should not have to cross-check one part of a website against another. Air Canada paid about CA$812 in damages, interest and fees.

The legal label was ‘negligent misrepresentation’: a careless, untrue statement someone reasonably relies on to their cost. It is a small-claims decision that does not bind other courts, but it reflects a principle found in most legal systems: a business answers for what it tells customers, whatever the channel.

When chatbots go wrong: what has actually happened

Most chatbot failures so far have been public embarrassments rather than lawsuits. In December 2023, a user reportedly got a chatbot on the website of Chevrolet of Watsonville, a California car dealer, to agree to sell a new Tahoe for $1 and to call it “a legally binding offer”. The user had instructed the bot to agree with anything the customer said. The dealer did not honour it, and no court case is reported to have followed.

In January 2024, the UK parcel firm DPD disabled the AI element of its online chat after a customer got it to swear and to write a poem criticising the company. DPD blamed an error after a system update. In April 2025, a support bot for Cursor, the AI coding tool, told users that being logged out when switching devices was due to a ‘one device per subscription’ policy. Cursor’s co-founder publicly confirmed no such policy existed, and the company reportedly began labelling AI support replies.

None of these cost much in legal terms, but each cost trust and could have become a complaint, refund dispute or regulator’s letter.

Do you have to tell people they’re talking to AI?

In a growing number of places, yes. The EU is the broadest. Article 50(1) of the AI Act has applied since 2 August 2026. It requires providers to design AI systems that interact with people so that users are told they are dealing with AI, unless that is obvious from the context. Fines for breaching Article 50 can reach €15 million or 3% of worldwide annual turnover.

The US has no federal bot-disclosure law, but several states do. Rules aimed at ‘companion’ chatbots, the kind designed for ongoing personal or emotional conversation, are spreading fastest. California’s SB 243 excludes bots used only for customer service, but general consumer disclosure laws such as Maine’s and Utah’s can still reach them.

The UK, India and Australia have no specific bot-disclosure statute. There, consumer law does the work: letting people believe a human is advising them can be a misleading practice if it affects their decisions. The simplest global approach is to label every bot clearly from the first message.

Can a chatbot’s wrong answer break consumer law?

Yes. Consumer protection laws ban misleading statements and omissions, and none of them carve out AI. In the US, section 5 of the FTC Act prohibits unfair or deceptive practices, and the FTC has said there is no AI exemption from existing law. In the EU, the Unfair Commercial Practices Directive covers misleading information given to consumers through any channel.

The UK has become the sharpest risk. Under the Digital Markets, Competition and Consumers Act 2024, the Competition and Markets Authority (CMA) has since 6 April 2025 been able to decide for itself that a business has broken consumer law and fine it up to 10% of global turnover, without first going to court. Australia’s Consumer Law prohibits misleading or deceptive conduct, with large maximum penalties for companies. India’s Consumer Protection Act 2019 covers misleading advertisements and unfair trade practices.

The test in all of these is the effect on the customer, not your intention. If your bot tells someone they have a 30-day refund window when you offer 14, the customer was misled, whoever wrote the sentence.

Are chat transcripts personal data?

Almost always. Chats contain names, emails, order numbers and often far more: health worries, money problems, complaints about colleagues. Under the GDPR in the EU and UK, India’s Digital Personal Data Protection Act and most modern privacy laws, that is personal data, and you are responsible for it as the business that decided to collect it.

In practice that means four things. Your privacy notice must mention the chatbot and what happens to conversations. You need a retention period rather than keeping logs forever. You need a data processing agreement with the vendor and, if the model provider sits in another country, a lawful basis for that transfer. And you should know whether the vendor uses your customers’ chats to train its models, and switch that off if you can.

In California, a wave of lawsuits has targeted website chat tools under the California Invasion of Privacy Act (CIPA), an old wiretap law, arguing that the chat vendor ‘eavesdrops’ on customers. Results have been mixed; the Ninth Circuit appeals court upheld a win for Converse in one such case. California’s SB 690, reported to have been signed in September 2026, curbs some CIPA claims from 2027 but reportedly leaves the wiretap provisions untouched. Clear notice that a vendor processes chats is the best defence.

What if your bot gives advice, talks to children or acts as a companion?

Risk rises sharply when a bot moves from ‘where is my parcel?’ to advice people rely on. Financial, medical and legal advice are regulated almost everywhere. A clinic’s bot that suggests a dose, or a bank’s bot that recommends a product, may breach professional rules whatever the disclaimer says.

Utah requires regulated professionals to disclose AI prominently, at the start of the conversation, when it is used for ‘high-risk’ interactions such as financial, legal, medical or mental health advice. Maine now bars offering therapy or psychotherapy through AI to the public unless a licensed professional provides it.

Companion and child-facing bots are under the heaviest scrutiny. In September 2025 the FTC ordered seven companies, including Alphabet, Character Technologies, Meta, OpenAI, Snap and xAI, to explain how they test and limit harms to children and teens from AI companion chatbots. Families have sued Character.AI and OpenAI over teenagers’ deaths; Character.AI and Google agreed to settle several of these cases in January 2026 without admitting liability, while others continue. California’s SB 243 requires companion bots to have suicide and self-harm protocols and gives users a right to sue for at least $1,000 per violation.

Prompt injection, data leakage and accessibility

‘Prompt injection’ means tricking a model with crafted text into ignoring its instructions. That is how the Chevrolet bot was talked into a $1 car. More serious versions hide instructions in a web page, email or uploaded file that the bot reads, and use them to pull out data or trigger actions such as issuing refunds.

A successful attack usually means a data breach or a loss you bear. The GDPR requires ‘appropriate technical and organisational measures’ for security, and a bot connected to your customer database is part of that. Give it the least access it needs.

Accessibility is easy to forget. Since 28 June 2025 the European Accessibility Act has applied to many EU e-commerce services, and US courts have applied the Americans with Disabilities Act to websites. A chat widget screen readers cannot use, offered as the only support route, can be a problem.

What regulators and courts are doing now

AI support bots are now standard in retail, travel, banking and software, across web chat, WhatsApp and voice. The direction of travel is clear: ‘the AI said it’ will not work as a defence.

Utah wrote this directly into law in 2025: it is not a defence to a consumer protection violation that generative AI made the statement or took the action. The FTC’s ‘Operation AI Comply’ sweep in September 2024 included DoNotPay, the self-styled ‘robot lawyer’, which agreed to pay $193,000 over claims about its chatbot. Italy’s data protection authority fined the company behind the Replika companion bot €5 million in 2025 over its legal basis and age checks.

Liability for defective AI is also widening. The EU’s new Product Liability Directive, which member states must apply from 9 December 2026, treats software, including AI, as a product. In the US, a federal judge in the Character.AI litigation allowed product liability claims to proceed in 2025. Expect more cases asking whether a harmful chatbot is a defective product.

Vendor contracts and guardrails that actually work

Read your vendor’s terms before launch. Standard terms often cap the vendor’s liability at the fees you paid and exclude responsibility for outputs. Ask for an indemnity, a promise to cover your losses, at least for intellectual property claims and the vendor’s own security failures. Check data use, storage location and breach notice times.

Disclaimers help less than people hope. Air Canada argued the customer could have found the correct policy elsewhere on its site; that did not save it. Consumer laws in the EU, UK and Australia limit how far you can exclude liability to consumers. A disclaimer works best as a signpost to the binding source: “For refunds, our policy here is final.” It is not a licence to be wrong.

The guardrails that matter are design choices. Test them before launch and after every model or prompt change.

• Limit scope: the bot answers only about your products, orders and published policies, and declines everything else.
• Ground answers in your own documents and link to the policy page it relied on.
• Block commitments: no prices, refunds, discounts or exceptions unless pulled from your systems.
• Offer a human handoff that is easy to find, and route complaints, vulnerable customers and regulated questions to people.
• Log conversations securely for a set period, so you can investigate complaints and fix patterns.
• Label the bot as AI at the start, and do not give it a human name and photo that suggest otherwise.