India group DSCI makes a proposal on liability for AI agent injury

The Data Security Council of India (DSCI), an industry group, made a proposal for a framework for liability when an autonomous AI agent causes injury. The paper names 5 groups in a deployment and 12 areas of injury. It applies existing law and contracts, not a new law. DSCI wants feedback from AI developers and other companies.

How it works

DSCI gives an authority map. It shows who controlled a decision and who could prevent the injury.

• Companies that use agents select the uses and the permissions.
• Integrators assemble the system and the approval steps.
• Tool providers supply the memory and safety features.
• Cloud providers supply system isolation, identity control and logs.

The proof

The paper gives no test results. It uses duty of care, product liability, causation and comparative negligence from existing law. It gives 6 safeguards, such as approval by a person for steps with a permanent effect and kill switches that have an owner.

What is not known

A company that uses a closed model can get recourse from the model provider only if the contract lets it. DSCI will give more guidance, and the feedback period is open.

Insight: Control of permissions may decide who pays

The draft law of Representative Lori Trahan in the US puts liability on an AI developer for an injury, also when it was careful. DSCI tells that the liability of a model provider can decrease when other companies get control. Lead Bank tells that a customer has no responsibility when an agent goes more than the set limits. Each reads as a proposal that puts the most risk on the party that sets agent permissions.