WeVote
Log In

Bill

Bill

SB 3444

AI MODEL SAFETY

104th Regular Session Introduced by Bill Cunningham

Illinois creates a safety framework for frontier AI models, offering liability protections to developers who meet specified safety, transparency, and governance standards.

0
WeVote Research Nonpartisan
Bill Summary · SB 3444

Summary of SB 3444 (104th Illinois General Assembly) — Artificial Intelligence Safety Act

This bill, introduced February 4, 2026, proposes a framework to regulate frontier artificial intelligence (AI) models in Illinois and provides liability protections for developers under specified conditions. It creates a new Act—the Artificial Intelligence Safety Act—and outlines safety, transparency, and compliance requirements. The Act can be overridden if federal law/regulations create overlapping requirements, or if another state’s framework is deemed substantially similar.

1) Purpose and Intent

  • Establish a state-level safety regime for frontier AI models to reduce the risk of “critical harms” while providing liability protections to developers that meet defined safety, transparency, and governance standards.
  • Encourage alignment with European Union safety standards or U.S. federal government collaboration, thereby facilitating recognition of compliance across jurisdictions.
  • Create a mechanism to sunset the Act if federal regulations establish overlapping frontier-model requirements.

2) Key Provisions and Changes

Definitions

  • Artificial intelligence model: A system that varies in autonomy and can generate outputs influencing physical or virtual environments.
  • Frontier model: An AI model meeting either:
    • More than 10^26 compute operations during training, or
    • Compute cost exceeding $100,000,000.
  • Critical harm: Death or serious injury of 100+ people or at least $1,000,000,000 in damages to rights in property, caused by frontier models, through either weaponization (chemical/biological/radiological/nuclear) or conduct that would be criminal if done by a person (lacking meaningful human intervention, and would constitute a crime if done by humans).

Liability Protections (Frontier Model Transparency)

  • A developer shall not be liable for critical harms if: 1) They did not intentionally or recklessly cause the harm, and 2) They published a safety and security protocol on their website (per Section 15) and a transparency report (per Section 20) on their website at the time of release.
  • Exceptions: If the developer could not foresee any material difference in capability/risks from a previously evaluated model, the protection does not apply.

Compliance Pathways (Subsection 10(b))

  • A developer is deemed to comply with the liability protections if, instead of a bespoke protocol/report, they:
    • (1) Agree to safety and security requirements adopted by the EU’s Artificial Intelligence Act (Article 56),
    • or (2) Enter into an agreement with a U.S. federal agency that:
    • Allows agency access to frontier models for research/evaluations,
    • Enables evaluation of cyber/biological risks,
    • Allows the federal government to release evaluation information about publicly released frontier models.

Certification Requirement (Section 10(c))

  • If a developer enters into a federal agreement (per 10(b)(2)), they must file a certification with the Attorney General verifying compliance with the agreement. False or misleading statements in the certification are prohibited.

Safety and Security Protocols (Section 15)

  • Developers must create a safety and security protocol documenting technical and organizational measures to identify and mitigate risk of critical harm.
  • Protocol must include:
    • Testing procedures for foreseeable risks,
    • Thresholds to identify risk levels and potential actions at each threshold,
    • Mitigations and assessment of their effectiveness,
    • Use of third parties for risk assessment and mitigation effectiveness,
    • Cybersecurity practices for frontier model weights and protection against unauthorized access/modification,
    • Monitoring and response plans for deployed models and actions to take when new material risks arise.
  • Redactions allowed to protect security, trade secrets, and proprietary information.

Transparency Reports (Section 20)

  • Reports must identify the frontier model and summarize the assessment results and steps taken to address identified risks.
  • Redactions allowed to protect sensitive information.

Sunset and Compliance with Other Jurisdictions (Section 25)

  • The Act does not apply if the federal government enacts laws/regulations with overlapping requirements.
  • Compliance with substantially similar frontier-model safety laws in another state can satisfy Illinois requirements.

3) Who/What Is Affected

  • Developers of frontier AI models meeting the thresholds (compute operations > 10^26 or compute costs > $100 million).
  • Potentially a broad set of AI researchers, tech firms, and organizations maintaining frontier models.
  • Entities engaged in federal or EU-aligned safety frameworks to demonstrate compliance.
  • The Illinois Attorney General, for certification related to federal-compliance agreements.

4) Procedural and Timeline Aspects

  • Status: Introduced in February 2026; subject to committee process and readings.
  • Key dates in action history indicate committee deadlines and potential third-reading deadlines in 2026, with the Rule 2-10 thresholds noted (e.g., May 15, 2026).
  • The Act contains a sunset/clause: if federal regulations establish overlapping requirements, Illinois’ Act would cease to apply.
  • Potential cross-state compliance: developers in Illinois may satisfy Illinois requirements by meeting another state’s substantially similar framework.

If you’d like, I can provide a comparison to the EU AI Act and a brief impact assessment for Illinois-based AI developers.

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

Sign in to ask a question.