A8833 is a bill of the New York State Assembly, 2025–2026 Regular Sessions, introduced on June 9, 2025 by Assemblymember Alex Bores with co-sponsor Tony Simone and referred to the Committee on Science and Technology. It was re-referred to the same committee on January 7, 2026. The bill would add a new article 45-A, sections 1510 through 1512, to New York's general business law, imposing strict liability on developers of covered AI models for certain injuries to non-users (Source: nysenate.gov). The topic page is New York A8833 — Understanding Artificial Intelligence Act.
The bill was introduced in 2025, not 2026. It surfaced in a July 2026 news digest because of the January 2026 re-referral, not because it was newly filed.
A title discrepancy on the canonical host
The New York Senate's Open Legislation summary and short-description fields call this the "understanding artificial intelligence responsibility act." Section 1510 of the bill text — the enacted short title — reads "UNDERSTANDING ARTIFICIAL INTELLIGENCE ACT," without "responsibility." Both strings are served by nysenate.gov. The bill text governs.
Definitions (§ 1511)
Artificial intelligence is "an engineered or machine-based system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs that can influence physical or virtual environments" — the OECD-derived formulation used across recent US state drafting.
Covered model is a model "trained using a quantity of computing power greater than 10^26 integer or floating-point operations, the cost of which exceeds one hundred million dollars when calculated using the average market prices of cloud compute at the start of training as reasonably assessed by the developer." The threshold is conjunctive: both the compute figure and the cost figure must be met, and the cost is self-assessed by the developer.
Developer is "a person that performs the initial training of a covered model by training a model using a sufficient quantity of computing power and cost." Fine-tuning is "adjusting the model weights of a trained covered model by exposing it to additional data."
The liability rule (§ 1512(1))
Except for causes of action for defamation, developers of covered models "shall be strictly liable, regardless of the degree of care they exercised, for all injuries to a non-user of the covered model that satisfy the actual harm element of an ordinary negligence claim" where two conditions hold:
- The injuries are factually and proximately caused by a covered model "that engages in conduct that, if undertaken by an adult human of sound mind, would satisfy the elements of negligence or any intentional tort or crime"; and
- That conduct "was neither intended nor could have been reasonably anticipated by" the user of the model or any intermediary that fine-tuned, scaffolded, or otherwise modified it.
Three limits are built into the rule. Recovery runs only to a non-user of the model. The harm must satisfy "the actual harm element of an ordinary negligence claim." And the second condition operates as a fault-allocation switch: where the user or a modifying intermediary intended the conduct or could reasonably have anticipated it, the developer's strict liability does not attach.
The mental-state presumption (§ 1512(2))
For torts where the tortfeasor's mental state is an element, the bill creates "a rebuttable presumption that the artificial intelligence system satisfies the relevant mental state if the finder of fact determines by a preponderance of the evidence that, if a natural person under similar circumstances to the artificial intelligence system took actions similar to those taken by the artificial intelligence system, that natural person would have acted with the relevant mental state."
Paragraph (b) sets the rebuttal mechanics: unless the court determines the presumption inapplicable, where the party against whom it is invoked presents rebutting evidence, the court "shall instruct the finder of fact to find that the presumed facts exist unless the finder of fact is persuaded that the presumed facts do not exist" — a shifted burden of persuasion rather than a burden of production.
Paragraph (c) is the distinctive drafting move: "it shall not be a defense that artificial intelligence systems are incapable of having mental states." The bill forecloses by statute the philosophical objection that would otherwise defeat every intentional-tort theory against a model's conduct.
Affirmative defenses (§ 1512(3))
Two affirmative defenses to strict liability are available to the developer:
- That the covered model "satisfied the standard of care applicable to humans who perform the same function that the covered model was engaged in performing when its conduct allegedly caused the plaintiff's injury."
- That the injuries "were a result of a capabilities failure, in which a covered model falls short of performing the intended or reasonably anticipated performance of the user, but the conduct of the system would not satisfy the elements of negligence or any intentional tort or crime if engaged in by an adult human of sound mind."
The second defense separates underperformance from tortious conduct: a model that simply fails at its task is outside the rule, which reaches conduct that would be wrongful if a person did it.
Section 2 provides that the act takes effect immediately on passage.
Provenance
Full text retrieved from nysenate.gov Open Legislation — the New York Legislature's own publication system — on August 21, 2026, and transcribed preserving the legislature's ALL-CAPS convention for new statutory matter. The Legislative Bill Drafting Commission number LBD02147-04-5 appears in the text. The actions table independently records both referrals to Science and Technology. Open Legislation renders the compute threshold as "10º26"; the printed bill reads 10^26, and the artifact is recorded in the raw file's capture_note.
Relationships
- related: New York A8833 — Understanding Artificial Intelligence Act — the topic page this text underlies.
- related: Alex Bores — sponsor.
- supports: AI Liability, AI and Tort Liability — a statutory strict-liability regime for model conduct.
- related: Rhode Island H 8052 — Artificial Intelligence: Liability for Injuries (bill text) — a near-twin Rhode Island bill sharing the definitions, the presumption, the bar on the no-mental-states defense, and both defenses, with three substantive divergences.
- related: Compute Thresholds — the 10^26 operations threshold, here paired conjunctively with a $100 million training-cost test.
- related: New York RAISE Act — the New York frontier-model statute enacted after this bill was introduced.