Third-party validation and shut-down capability of artificial intelligence models

Summary#

Int. 1100-2026, sponsored by Council Speaker Julie Menin, would amend the New York City Administrative Code by adding rules for artificial intelligence models. It is not law: the latest listed step is referral to committee. Its stated goal is to require third-party validation and a human-operated shut-down capability before covered models are marketed, sold, offered for sale, or deployed in New York City.

  • A person could not market, offer for sale, sell, or deploy a model unless a third-party validator has assessed it and it has a technical capability for a human operator to stop it temporarily or permanently.
  • The validator must assess task performance, determinism, speed and volume of output, training-data origins, bias against protected classes, lawful and secure data handling, informed consent, and safety risks. The Office of Cyber Command may add assessment topics.
  • The validator must give the model’s developer and Cyber Command a certification stating whether the model has been validated and is appropriately positioned for deployment, and disclose any financial or other interest in the model or developer.
  • Cyber Command’s director must make rules on validator qualifications and certification submissions, and may designate an agency to issue summonses.
  • Violations can lead to civil penalties: $25,000 per instance for marketing, offering, selling, or deploying a model without validation or shut-down capability; $25,000 per instance for falsifying a validation; and up to $25,000 for other violations. The City may also seek court orders and other remedies.

What it means for you#

  • AI model developers and people or businesses marketing, offering, selling, or deploying models in New York City: They would need to ensure models meet both conditions. The bill defines a third-party validator as someone engaged by a developer who is not the developer’s affiliate.
  • Third-party validators: They would have to conduct the listed assessments and disclose their interests. Cyber Command’s rules would set their qualifications.
  • Residents and other people affected by models: The bill adds assessments of bias, privacy, consent, and safety, but does not create a right to apply for a review or a process for individuals to challenge a model.
  • Public agencies: Cyber Command must make rules; it or an agency designated by its director may issue summonses. The City’s Corporation Counsel may bring a court case to enforce the law.
  • The bill does not ban AI models outright or require every model to be used. It sets conditions for the listed commercial activities and deployment in the city.

Money#

No cost information is in the available material.

  • The bill sets civil penalties but gives no estimate of penalty revenue, enforcement costs, or staffing needs.
  • Developers would need a third-party validation for covered models; the bill does not say who pays the validator or how much validation would cost.
  • The law would take effect 180 days after it becomes law.

What is unclear#

  • The bill does not explain how Cyber Command will decide whether a model is “appropriately positioned for deployment,” or what evidence validators must use.
  • The bill does not define “per instance” for calculating penalties.
  • It does not set a process, deadline, or appeal route for validation decisions or enforcement.
  • The scope of the definition of an AI model is broad; the bill does not state which kinds of models or uses would be excluded.
  • Cyber Command’s director may add assessment topics, but the bill does not limit what those topics may be.

Case for#

  • The bill appears intended to address risks from models that perform poorly, treat protected groups unfairly, mishandle data, or create safety and security risks.
  • A possible argument for requiring an independent validator is that it adds scrutiny beyond the model developer’s own assessment.
  • Requiring a human-operated shut-down capability could give an operator a way to stop a model if it causes harm.

Case against#

  • One concern is that the bill requires validation but leaves important standards and procedures to future Cyber Command rules.
  • The bill covers several kinds of activity and imposes fixed penalties per instance; unclear counting rules could make potential liability difficult to predict.
  • Validation could add costs and delays for developers, but the bill provides no cost estimate or detail on how the process would work.
  • Cyber Command may add assessment topics, while the bill does not describe limits or a review process for those choices.