AI shopping assistant hallucinations: who is liable for wrong product claims?
An assistant that invents a fabric, a discount or a returns rule creates real liability. How hallucinations happen in fashion e-commerce, what the law says and how to reduce the risk.
KEY TAKEAWAYS Summary by the editors
- A hallucination is a confident but false answer from a generative AI system, such as an invented material composition, size availability, discount or returns rule.
- In Moffatt v Air Canada (2024), a British Columbia tribunal held the airline responsible for incorrect information from its chatbot and rejected the argument that the chatbot was a separate legal entity.
- The EU Unfair Commercial Practices Directive covers untruthful information before, during and after a consumer transaction, and its amendment by Directive (EU) 2024/825 applies from 27 September 2026.
- The revised EU Product Liability Directive explicitly covers software including AI systems and applies to products placed on the market from 9 December 2026.
- Grounding answers in verified catalogue data, restricting claim types, logging conversations and clear escalation are the main practical controls.
When an AI shopping assistant states something false about a product, the retailer that deploys it should assume it is responsible, just as it would be for a wrong statement on a product page. Courts have already rejected the argument that a chatbot is separate from the company, and EU consumer law treats untruthful information as unfair regardless of how it was produced. The practical answer is to ground the assistant in verified data and restrict what it may claim.
What is an AI hallucination in fashion e-commerce?
A hallucination is a confident answer that is not supported by facts. Large language models generate plausible text, and when relevant data is missing or ambiguous they may fill gaps. In fashion, typical examples are easy to imagine and costly in practice.
| Claim type | Example of a wrong answer | Possible consequence | Control |
|---|---|---|---|
| Material and composition | Stating 100% cotton for a cotton blend | Misleading claim, returns, allergy complaints | Answer only from structured composition data |
| Sustainability | Calling a product 'eco-friendly' without evidence | Greenwashing exposure under consumer law | Allow only certified, documented claims |
| Size and stock | Confirming a size is available when sold out | Cancelled orders, lost trust | Live stock lookup per size |
| Price and promotions | Quoting an expired discount | Disputes, pressure to honour the price | Read prices from checkout system only |
| Policies | Inventing a longer returns window | Customer claims based on chatbot statement | Retrieve current policy text verbatim |
| Care and safety | Wrong washing instructions | Damaged products, complaints | Use care labels from product data |
Is a retailer liable for what its AI assistant says?
The clearest precedent so far comes from Canada. In February 2024, the British Columbia Civil Resolution Tribunal decided Moffatt v Air Canada. The airline's chatbot had told a customer he could apply for a bereavement fare reduction within 90 days after travel, which contradicted the actual policy. Air Canada argued that correct information was available elsewhere on its website and that the chatbot was a separate legal entity responsible for its own actions. The tribunal rejected this, held that the company was responsible for all information on its website whether from a static page or a chatbot, and awarded around 650 Canadian dollars plus interest and fees.
The sums were small, but the reasoning matters for retail: a company cannot disclaim its own assistant. Laws differ by country, and the case is not binding in Europe, but EU consumer law reaches a similar practical outcome through different routes.
Which EU rules apply to wrong product claims by AI?
- Unfair Commercial Practices Directive: the EU's overarching law on unfair business practices towards consumers, covering untruthful information before, during and after a transaction. Its amendment by Directive (EU) 2024/825 applies from 27 September 2026. Unsupported environmental wording from an assistant deserves particular care.
- EU AI Act, Article 50: from 2 August 2026, AI systems intended to interact with people must be designed so users know they are dealing with an AI, unless obvious. Disclosure does not remove responsibility for accuracy.
- Product Liability Directive (EU) 2024/2853: entered into force on 8 December 2024 and applies to products placed on the market from 9 December 2026. It explicitly covers software, including AI systems, and can make manufacturers liable for defects arising from updates or machine-learning features. It concerns defective products causing damage, so it is more relevant to AI software providers than to a simple wrong answer about fabric.
How can fashion retailers reduce hallucination risk?
- Ground every product fact: configure the assistant to answer product questions only from retrieved catalogue data, and to say it does not know when data is missing.
- Restrict claim categories: block or tightly template sustainability, health and safety statements, allowing only claims documented in product data.
- Use live systems for price and stock: never let the model infer availability or prices from text.
- Quote policies, do not paraphrase: retrieve current returns and delivery texts and link to the source page.
- Test with adversarial questions: include leading questions such as 'this is organic, right?' in pre-launch testing.
- Log and review: keep conversation logs in line with data protection rules and review samples weekly.
- Define a correction process: decide in advance how to handle customers who relied on a wrong answer, for example honouring a quoted price within limits.
What about errors made by external assistants such as ChatGPT?
When a third-party assistant misdescribes your product, your direct control is limited, but the same data discipline helps. External platforms rely heavily on merchant feeds and structured data; complete, accurate feeds give them less room to guess. Monitor how major assistants describe your best-selling products, correct the underlying data, and use platform feedback channels where available.
Who inside the company should own this risk?
Ownership is usually shared: e-commerce owns the assistant, product data teams own the facts, legal sets claim rules, and customer service handles escalations and corrections. Name one accountable owner, agree an accuracy threshold before launch, and give that person the authority to restrict or switch off the assistant if errors spike.
It also helps to document the controls. If a regulator or court ever examines a wrong statement, evidence that the company grounded answers in verified data, restricted sensitive claims, tested before launch and corrected errors promptly will carry weight. A short internal policy covering these points, reviewed whenever the model, prompts or data sources change, is a modest effort compared with the reputational cost of a widely shared screenshot of an assistant making false product claims. Review the policy at least once a year, and after any significant incident, so that it reflects how the assistant is actually used.
Frequently asked questions
Can a company be held liable for its chatbot's mistakes?
Yes, in at least one notable case. In Moffatt v Air Canada (2024), a British Columbia tribunal held the airline responsible for wrong information from its chatbot and rejected the claim that the chatbot was a separate legal entity. Outcomes depend on local law.
Does the EU AI Act make retailers liable for AI hallucinations?
The AI Act's Article 50 mainly requires transparency that users are interacting with AI, applying from 2 August 2026. Liability for misleading product statements comes mainly from consumer law, such as the Unfair Commercial Practices Directive.
How do you stop an AI shopping assistant from making things up?
Ground answers in retrieved catalogue data, use live systems for price and stock, restrict sensitive claim types such as sustainability, and instruct the assistant to say it does not know when data is missing. Test with adversarial questions and review logs regularly.
Does the new EU Product Liability Directive cover AI?
Yes. Directive (EU) 2024/2853 explicitly covers software including AI systems and applies to products placed on the market from 9 December 2026. It concerns defective products causing damage rather than every incorrect answer.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
