Courts Hold Companies Liable for Chatbot Statements

PYMNTS distinguishes two failure modes: a hallucination is a false output such as a made-up citation or number, while a confabulation is when a model fills a knowledge gap with an answer delivered with the confidence of a real policy, which, to a customer or a court, carries the same authority as an actual company statement. The clearest precedent is Moffatt v. Air Canada, 2024 BCCRT 149: Air Canada's chatbot told customer Jake Moffatt he could buy a full-fare ticket and claim a retroactive bereavement discount within 90 days, but no such policy existed. Air Canada argued the chatbot was a separate legal actor responsible for its own statements; tribunal member Christopher Rivers rejected that directly, calling it "a remarkable submission," and ordered Air Canada to pay Moffatt 812.02 Canadian dollars (about $570) in damages and fees. PYMNTS reports a second case in April 2025: an AI support bot at coding startup Cursor, named Sam, invented a policy limiting subscriptions to a single device after a session bug caused unexpected logouts; the claim spread on Hacker News and Reddit before co-founder Michael Truell publicly corrected it, saying "Hey! We have no such policy... this is an incorrect response from a front-line AI support bot."

PYMNTS also documents the financial system pricing this risk directly: Lloyd's of London, through startup Armilla, launched an insurance product in May 2025 covering claims tied to AI hallucination losses; FINRA's 2026 Annual Regulatory Oversight Report flagged hallucinations as a specific compliance concern for broker-dealers; and hallucination-mitigation startup Scaled Cognition raised $100 million in June 2026 to build enterprise-grade controls. PYMNTS frames this as evidence that the cost of confabulation, once treated as a product quality problem, is now being priced as a financial risk.

Hallucinations and confabulations call for different mitigations. Factual errors are often addressable with better grounding and retrieval-augmented generation; policy-sounding fabrications need explicit uncertainty signaling, deterministic fallbacks for contractual or pricing questions, and escalation to a human or an authoritative source before a model is allowed to state a policy as fact. Given the Air Canada and Cursor precedents, product and legal teams should treat any customer-facing conversational path that touches pricing, refunds, or contractual terms as a liability surface requiring audit logging and response review, not just a quality metric.

Whether more jurisdictions issue rulings similar to Moffatt v. Air Canada, whether AI-hallucination insurance following Lloyd's/Armilla becomes standard for enterprises deploying chatbots, and whether regulators beyond FINRA issue explicit guidance on AI customer-service liability.