New York Assembly Bill A8833, the Understanding Artificial Intelligence Act, is a pending bill of the 2025–2026 Regular Sessions that would add article 45-A to the state's general business law. It imposes strict liability on developers of "covered models" for injuries to non-users caused by a model engaging in conduct that would be negligent, an intentional tort, or a crime if an adult human of sound mind had engaged in it. The bill was introduced on June 9, 2025 by Assemblymember Alex Bores with co-sponsor Tony Simone, referred to the Assembly Committee on Science and Technology, and re-referred to the same committee on January 7, 2026 (New York A8833 — Understanding Artificial Intelligence Act (bill text)).
The bill predates the signing of the RAISE Act, New York's frontier-model transparency statute, which Bores also sponsored. It is earlier drafting on a different governance mode — downstream liability rather than upstream disclosure — and not a follow-on to it.
Status and timeline
| Date | Event |
|---|---|
| 2025-06-09 | Introduced by M. of A. Bores; referred to Science and Technology |
| 2026-01-07 | Re-referred to Science and Technology |
The bill has not advanced out of committee. It would take effect immediately on passage.
Naming
The New York Senate's Open Legislation summary calls this the "understanding artificial intelligence responsibility act," while § 1510 of the bill text — the enacted short title — reads "UNDERSTANDING ARTIFICIAL INTELLIGENCE ACT," without "responsibility." Both strings come from the legislature's own publication system; the bill text governs (New York A8833 — Understanding Artificial Intelligence Act (bill text)).
Scope
A covered model is one trained using more than 10^26 integer or floating-point operations and costing more than $100 million, calculated using average market cloud-compute prices at the start of training "as reasonably assessed by the developer." Both limbs must be met, and the cost figure is self-assessed. A developer is a person who performs the initial training of a covered model; fine-tuning an existing model does not make a party a developer under this bill. See Compute Thresholds.
Key provisions
Strict liability (§ 1512(1)). Except for defamation claims, developers of covered models are strictly liable "regardless of the degree of care they exercised" for injuries to a non-user that satisfy the actual-harm element of an ordinary negligence claim, where the injuries are factually and proximately caused by model conduct that would be negligent, an intentional tort, or a crime if undertaken by an adult human of sound mind, and where that conduct was neither intended nor reasonably anticipatable by the user or by an intermediary that fine-tuned, scaffolded, or otherwise modified the model.
Mental-state presumption (§ 1512(2)). For torts where mental state is an element, a rebuttable presumption arises that the system held the relevant mental state where the finder of fact concludes by a preponderance that a natural person acting similarly in similar circumstances would have. Once rebutting evidence is presented, the court instructs the finder of fact to find the presumed facts exist unless persuaded otherwise — shifting the burden of persuasion rather than the burden of production. Section 1512(2)(c) provides that "it shall not be a defense that artificial intelligence systems are incapable of having mental states."
Affirmative defenses (§ 1512(3)). A developer may establish that the model satisfied the standard of care applicable to humans performing the same function, or that the injuries resulted from a "capabilities failure" — the model falling short of intended or reasonably anticipated performance without engaging in conduct that would be tortious or criminal if a person did it (New York A8833 — Understanding Artificial Intelligence Act (bill text)).
Comparison with Rhode Island H 8052
Rhode Island H 8052, introduced February 27, 2026, shares this bill's AI definition, its rebuttable-presumption mechanism, its bar on the no-mental-states defense, and both of its defenses, in closely parallel language. Three divergences change the regime materially. Whether the two descend from a common model text is not established by either document.
| Dimension | New York A8833 | Rhode Island H 8052 |
|---|---|---|
| Covered systems | "Covered models" only — above 10^26 operations and $100 million training cost | Any "artificial intelligence system"; no threshold |
| Who may recover | Non-users only | No user/non-user distinction |
| Who is a developer | Performs initial training | Performs initial training or fine-tunes an existing model |
| Fault-allocation switch | Conduct neither intended nor "reasonably anticipated" by user or intermediary | Conduct neither intended nor "a consequence of negligence" by user or intermediary |
| Strength of defenses | Affirmative defenses to this section's strict liability | Absolute defenses to liability "including liability under the common law or existing statutes" |
| Harm threshold | Must satisfy "the actual harm element of an ordinary negligence claim" | No stated harm-element requirement |
The defense-strength difference runs in the opposite direction from the scope difference: Rhode Island reaches far more systems, but a developer who establishes either defense is insulated from pre-existing common-law and statutory claims as well, which the New York bill does not provide.
The mental-state drafting move
Statutorily presuming that an AI system holds a mental state, and then barring the argument that it cannot, is a distinctive response to a recurring doctrinal problem: intentional torts and most crimes have a mens rea element that a model cannot satisfy on any conventional account, which would leave conduct outside the reach of those causes of action however harmful its effects. Both bills resolve it by legislative fiat rather than by analogy to vicarious liability or to the abnormally-dangerous-activity doctrine, the routes canvassed at AI and Tort Liability. Neither has advanced out of committee, so no court has construed the mechanism.
Relationships
- depends-on: New York A8833 — Understanding Artificial Intelligence Act (bill text) — the bill text.
- related: Alex Bores — sponsor.
- related: Rhode Island H 8052 — Artificial Intelligence: Liability for Injuries — near-twin bill in Rhode Island.
- instance-of: AI Liability — a statutory strict-liability regime for model conduct.
- supports: AI and Tort Liability — a legislative alternative to developing the theory through common-law litigation.
- related: Compute Thresholds, New York RAISE Act.