AI Liability

Picture a customer-service agent that promises a refund the company never offered, or a coding agent that ships a bug that takes down a client’s system. Someone bears the cost. Existing law (contract, negligence, product liability, consumer protection) already answers many of these cases, but AI adds wrinkles: the developer may not know how the model will be used, the deployer may not know how it was built, and the chain of cause is hard to prove.

The law is still moving. The European Commission proposed an AI Liability Directive in 2022 to make it easier for victims to bring claims involving AI; the European Parliament’s legislative tracker shows the proposal was withdrawn, with the withdrawal announced in the Commission’s 2025 work programme and made official in October 2025. Liability questions therefore still run largely through general rules that differ by country.

Contracts fill part of the gap. In episode 17 at 18:31, IBM’s Maryam Ashoori describes IBM’s approach for its Granite models: govern and document the training data, “be transparent with the market about how they are trained and provide client indemnification,” meaning IBM takes on certain legal claims for customers. That is a vendor’s account of its own offer; episode 17 was sponsored by Galileo. Editorially, indemnities shift who pays without settling who is at fault, which is why data provenance, an audit trail and clear algorithmic accountability matter when something goes wrong. For the training-data side, see copyright and AI training.