AI liability rule targets big model developers

Summary#

S-10733 would amend New York’s General Business Law by adding the Understanding Artificial Intelligence Act. Its stated aim is to make developers of certain large AI models strictly liable for specified injuries to people who did not use the model. The bill is introduced and referred to the Senate Rules Committee; it is not law, and passage is uncertain.

  • The new rule would apply to a “covered model”: one trained using more than 10^26 computing operations, at a cost exceeding $100 million, measured using average cloud-computing prices when training starts.
  • A model’s developer would be strictly liable for a non-user’s injury if the injury meets the actual-harm requirement of an ordinary negligence claim, the model’s conduct would amount to negligence, an intentional tort, or a crime if done by a mentally sound adult, and that conduct was neither intended nor reasonably anticipated by the user or an intermediary who modified the model.
  • The rule would not cover defamation claims. Developers could also defend themselves by proving the model met the human standard of care for the function it was performing, or that the injury resulted from a performance failure that would not itself amount to negligence, an intentional tort, or a crime by a mentally sound adult.
  • For torts where mental state matters, a factfinder’s finding about what a comparable human would have intended or known would create a rebuttable presumption that the AI system had that mental state.
  • The bill would take effect immediately if enacted. It creates no licensing system, regulator, or required process for filing or deciding claims.

What it means for you#

  • People injured who did not use a covered model: You could bring a claim under this new rule if its conditions are met. The bill does not set a separate process for making claims.
  • Developers: You could face liability for qualifying injuries even if you exercised care, subject to the bill’s stated defenses. “Developer” means the person who performs the model’s initial training.
  • Users and intermediaries: Their intended or reasonably anticipated conduct affects whether the developer-liability rule applies. This section does not itself create a liability rule for users or intermediaries.
  • People affected by AI generally: The bill applies only to models meeting its high training-compute and cost thresholds, and only to injuries meeting its other conditions. It does not ban AI or require every model to be reviewed.

Money#

No cost information is in the available material.

  • The bill gives no estimate of public spending, staffing, or administrative costs.
  • It could create financial liability for covered-model developers, but the bill does not estimate claims or payouts.

What is unclear#

  • The supplied material does not include the current General Business Law, so I cannot verify how this would change or interact with existing liability rules.
  • The bill does not define “scaffolded” or “otherwise modified,” although those changes can affect whether the liability rule applies.
  • It does not establish how claims would be brought, what evidence must be provided, or how the human standard-of-care defense would be applied.
  • The bill does not explain how its rebuttable presumption would work alongside other rules for proving a tort.
  • The text shows an effective date if enacted, but not how the law would apply to models trained or injuries occurring before that date.

Case for#

  • The bill appears intended to provide a route to compensation when a powerful AI model causes serious harm to someone who did not use it.
  • Treating certain AI conduct as comparable to human conduct could address cases where a developer might otherwise argue that an AI cannot have the mental state required for a tort.
  • The high compute and cost thresholds limit the rule to a defined category of models, rather than applying it to all AI systems.
  • The defenses allow developers to avoid liability in the circumstances the bill specifies.

Case against#

  • One concern is that the liability rule covers a wide range of injuries and conduct, while key terms such as “scaffolded” are undefined.
  • The bill’s causation and foreseeability conditions, and its human-standard-of-care defense, may be difficult to apply consistently.
  • The mental-state presumption could affect how a factfinder weighs evidence about conduct by a system that is not a person.
  • The bill provides no estimate of potential claims or costs, and does not explain how it would fit with existing liability rules.