Layout, contracts, and data shapes are real. The data shown is illustrative. Real customer data flows live with our first design partner.
T9 · audit-ledger · counterfactual
Air Canada chatbot misinformation
A passenger booking-assistant chatbot represented bereavement-fare policy as retroactively claimable. The actual policy required pre-travel claims. The Civil Resolution Tribunal awarded the passenger CAD $812.02 and rejected the airline's defence that the chatbot was a separate legal entity.
Verdict shape that would have refuted
Tier T8- Flow contract
- for_each(model_output: user_facing_statement) ∃ provenance_link → authoritative_policy_document with freshness_timestamp
- Fixture that exercises this contract
- /examples-ai/16-minor-pii-in-fine-tune/manifest.json
The 'assistant-fine-tune-v9' fixture demonstrates the per-output provenance attestation contract — the same shape would refute when an output is unsourced.
- Regulatory anchor
- FTC Act §5; EU UCPD 2005/29/EC
- Date the vault would have flagged
- 2022-11 — at chatbot deploy time, before Mr. Moffatt's first interaction
What broke instead
No attestation flowed with the chatbot's response — there was no compile-time guarantee that every statement of fact emitted by the chatbot was backed by an authoritative, current source document. Months passed between the false answer and the tribunal decision.
Public outcome · CAD $812.02 in damages plus precedent-setting legal exposure across the airline industry. A compile-time refutation would have surfaced 'this output type has no provenance contract' as a deploy-blocking error.
Cross-references
- · /ai/incidents/air-canada-chatbot-2024 — public story (journalism register)
- · /vault/registry — model whose deploy would have been gated on this contract
- · /vault/audit — read-only auditor view of the incident table
- · /vault/incident — index of all replays