WeVote
Log In

Bill

Bill

HB 4668

Trade: business regulation; requirements and safety standards for developers of certain artificial intelligence models; provide for. Creates new act.

2025-2026 Regular Session Introduced by Sarah Lightner

Requires large AI developers to publish safety, security, and transparency protocols and quarterly risk-based reports for critical risks.

re-referred to Committee on Communications and Technology
#communications:artificialintelligence #trade:businesspractices #trade:businessregulation #trade:consumergoodsandservices
0
WeVote Research Nonpartisan
Bill Summary · HB 4668

Summary — HB 4668: Artificial Intelligence Safety and Security Transparency Act

Status: Enacted (signed by Governor 6/20/2025; effective immediately)
Introduced: Filed 3/12/2025 (Rep. Sarah Lightner). Companion: SB 1876.

Purpose

To require large developers of “foundation models” to adopt and publish documented safety, security, and transparency practices designed to identify, mitigate, and respond to high‑consequence (“critical”) risks from those models. The act also creates employee protections, reporting obligations, and civil remedies/sanctions tied to compliance.

Key definitions

  • Foundation model: an AI model trained on a broad dataset, designed for general outputs and adaptable to many tasks.
  • Artificial intelligence model: a machine-based system that infers from input to generate outputs that can influence physical or virtual environments.
  • Deploy: use a foundation model or make it foreseeably available to third parties for use/modification/copying/combination (with limited exceptions for development and evaluation).
  • Large developer: a person/entity that has developed (a) a foundation model whose compute costs at U.S. cloud prices are ≥ $5,000,000, and (b) within the previous 12 months, one or more foundation models with combined compute costs ≥ $100,000,000 (measured at the time compute was used).
  • Critical risk: a foreseeable, material risk that a foundation model’s development, storage, or deployment will cause either (a) death or serious injury to more than 100 people, or (b) more than $1,000,000,000 in monetary/property damages — arising from specified incidents such as creation/release of CBRN agents, cyberattacks enabled by a model, model conduct that (with limited human intervention) would be a criminal act, or harms made substantially more likely by the developer’s activities.

Major requirements

  • Safety & security protocol (required of large developers, effective Jan 1, 2026):

    • Must be produced, implemented, followed, and conspicuously published.
    • Must document technical and organizational measures addressing critical risks, including:
    • Scope/exclusions, intolerable‑risk thresholds and responses, testing/assessment procedures (including misuse/evasion scenarios), deployment decision procedures, physical/digital/organizational access controls, safeguards and their evaluated efficacy, incident response procedures, conditions for reporting incidents, and modification procedures.
    • Identification of portions suitable for independent scientific assessment and provision mechanisms for experts (including access to unredacted material where appropriate).
    • Role(s) of financially disinterested third parties, if any.
    • Material modifications to the protocol must be published within 30 days.

  • Transparency reporting (beginning Jan 1, 2026):

    • At least once every 90 days, publish a transparency report covering the period from 120 days before publication to 30 days before publication.
    • Reports must include conclusions of any risk assessments completed in the reporting period; per‑risk assessments of the model(s) that would pose the highest level of each critical risk if deployed without safeguards; and, when a newly deployed/modified model increases critical risk, the decision process and safeguards applied.
    • Test records and assessment results must be retained for five years, with sufficient detail for qualified third parties to replicate testing.

Other provisions & effects

  • Employee protections: the act defines “employee” broadly (including contractors, subcontractors, unpaid advisors, and corporate officers) and provides protections (summary indicates protections for certain employees involved with assessing/managing critical risks).
  • Enforcement & remedies: the act establishes civil causes of action and sanctions for violations and assigns powers/duties to state and local governmental officers/entities (text provides for civil remedies but specific penalty amounts/structures are in the full text).
  • Who’s affected: primarily “large developers” as defined (entities using high levels of compute to train foundation models). Indirectly affects cloud providers, downstream deployers who obtain models, and expert third parties who may be engaged for assessments.
  • Timing: compliance obligations kick in Jan 1, 2026; transparency reports are quarterly (every 90 days) with a prescribed reporting window; protocol changes must be published within 30 days.

Procedural history (summary)

  • Filed 3/12/2025; referred to committees and considered in hearings. Passed both chambers (May 2025), sent to Governor 5/28/2025 and signed 6/20/2025; effective immediately.

For full statutory text, definitions, and the detailed enforcement/penalty language, consult the enrolled bill or the state’s legislative website.

Compiled from official sources — confirm details with the bill’s official record.

Sign in to ask a question.