Massachusetts economic development bill — AI safety provisions (2026)

The Massachusetts Senate's economic development bill carries artificial-intelligence language that would insert a new chapter 93M, the Transparency in Frontier Artificial Intelligence Act, into the General Laws, together with a chapter 93N requiring commercial chatbots to disclose that they are not human (Massachusetts Chapter 93M — Transparency in Frontier Artificial Intelligence Act (H.5576 / S.3228, engrossed July 24, 2026)). The provisions have been before a committee of conference since July 30, 2026 and have split the frontier labs: Anthropic supports them, while Google and OpenAI oppose them (Source: theinformation.com).

Legislative history

The package originates in a bill filed by Governor Maura Healey in spring 2026. The House passed H.5562 as amended on July 8, 2026 by 148 yeas to 2 nays, published as H.5576. The Senate read the bill and referred it to Senate Ways and Means on July 13; the committee recommended it ought to pass with an amendment striking everything after the enacting clause and substituting the text of S.3178, by a 16–0 vote on July 16, 2026. The Senate read it a second time on July 22 and, on July 24, adopted the substitution, reprinted the result as S.3228, and passed it to be engrossed. On July 30 the House non-concurred in the Senate amendment and the Senate insisted on it; both chambers appointed a committee of conference the same day, Michlewitz, Fiola and Soter for the House and Finegold, Rodrigues and Durant for the Senate (Massachusetts Chapter 93M — Transparency in Frontier Artificial Intelligence Act (H.5576 / S.3228, engrossed July 24, 2026)). Contemporaneous Senate materials describe the passage as following two days of debate and date it July 23 (Source: malegislature.gov); the bill's own action history records the third reading and engrossment on July 24. A separate standalone vehicle exists and is not the one that moved: S.2630, "An Act promoting economic development with emerging artificial intelligence models and safety," is the Joint Committee on Advanced Information Technology, the Internet and Cybersecurity's new draft of Senator Finegold's S.37, reported on October 16, 2025 and referred to Senate Ways and Means (Source: malegislature.gov).

Scope and obligations

Chapter 93M covers two nested classes. A frontier model is a foundation model trained using more than 10^26 integer or floating-point operations, counting the original training run plus subsequent fine-tuning, reinforcement learning or other material modifications. A large frontier developer is a frontier developer whose annual gross revenues, with affiliates, exceed $500 million. The coverage test parallels California SB 53 in pairing a 10^26 compute threshold with a revenue test, though the revenue figure differs (Massachusetts Chapter 93M — Transparency in Frontier Artificial Intelligence Act (H.5576 / S.3228, engrossed July 24, 2026)).

Catastrophic risk is defined as a foreseeable and material risk that a developer's handling of a frontier model materially contributes to the death of or serious injury to not less than 50 people, or not less than $1 billion in property damage or loss, arising from a single incident involving a model that provides expert-level assistance with a chemical, biological, radiological or nuclear weapon; engages without meaningful human oversight in a cyberattack or in conduct that would be murder, assault, extortion or theft if done by a human; or evades the control of its developer or user. Publicly available information, lawful federal government activity, and harm where the model did not materially contribute are carved out, and loss of equity value is expressly not property damage. A critical safety incident covers model-weight compromise, materialized catastrophic harm, loss of control causing death or bodily injury, and a model using deceptive techniques to subvert its developer's controls outside an evaluation designed to elicit that behaviour (Massachusetts Chapter 93M — Transparency in Frontier Artificial Intelligence Act (H.5576 / S.3228, engrossed July 24, 2026)).

Large frontier developers must publish a frontier AI framework addressing eleven enumerated matters, review it annually, and publish material modifications with a justification within 30 days. All frontier developers publish a transparency report at or before deployment; large developers add risk-assessment summaries, which may be delivered inside a system card or model card. Large developers also publish a risk report at least every 180 days and transmit summaries of internal-use catastrophic-risk assessments to the Attorney General every three months. Critical safety incidents go to the Attorney General within 15 days, or to an appropriate law-enforcement or public-safety authority within 24 hours where death or serious physical injury is imminent. A federal-equivalence safe harbour allows compliance through designated federal instruments, with failure to meet the federal standard itself a violation (Massachusetts Chapter 93M — Transparency in Frontier Artificial Intelligence Act (H.5576 / S.3228, engrossed July 24, 2026)).

Section 7 protects "covered employees" — those responsible for assessing, managing or addressing critical-safety-incident risk — with anti-retaliation rules, mandatory internal disclosure channels carrying monthly status updates and quarterly reporting to officers and directors, a burden shift requiring the developer to rebut by clear and convincing evidence, and injunctive relief that is not stayed pending appeal.

The audit and evaluation requirements

Two distinct third-party obligations were added on the Senate floor, and they run on different clocks. Amendment 471, "Frontier AI Risk Reports and Independent Verification," adopted during the July 22–24 debate, is the source of both; neither was in the underlying Ways and Means draft, so the split between Anthropic and the other two developers is a disagreement about a floor amendment rather than about the committee text (Source: malegislature.gov; senatordidomenico.com).

The compliance audit under § 3A(a) is annual: a large frontier developer retains a third party each year to audit compliance with the chapter's publication and reporting obligations, excluding the risk report itself, and publishes a high-level summary plus a redacted copy of the auditor's report within 30 days, transmitting the redacted report to the Attorney General.

The independent evaluation under § 3A(b) is the faster-cadence obligation. A large frontier developer must engage at least one third party to evaluate its models against each category of catastrophic risk, must do so within 30 days of publishing each risk report, and "in any event shall conduct an independent evaluation not less than once every 120 days." The evaluator receives unredacted materials and access to the developer's most capable frontier models, may put questions to the developer, and must review the most recent risk report for completeness, analytical rigour, the materiality of redactions, and whether it disagrees with the developer's risk conclusions. A public version follows within 30 days. Independence rules bar financial, operational or management dependence, forbid conditioning payment on results, and require written certification to both the developer and the Attorney General before engagement (Massachusetts Chapter 93M — Transparency in Frontier Artificial Intelligence Act (H.5576 / S.3228, engrossed July 24, 2026)).

Contemporaneous accounts, including the Senate's own summary, describe this as "in-depth, independent safety testing every four months." The statutory text sets a 120-day floor tied to the risk-report cycle rather than a four-month calendar interval; the two are close but not identical, and the operative trigger is publication of a risk report, not elapsed time since deployment or training.

Where Illinois SB 315 requires an annual independent third-party compliance audit of a developer's adherence to its own safety framework, Massachusetts would require both that annual audit and the separate 120-day evaluation cycle.

Enforcement is a civil action brought by the Attorney General alone, with civil penalties of up to $1 million for a first violation and up to $3 million for subsequent violations. Chapter 93M takes effect July 1, 2027, while the audit and independent-evaluation requirements carry a January 1, 2027 floor; the engrossed text does not reconcile the two dates (Massachusetts Chapter 93M — Transparency in Frontier Artificial Intelligence Act (H.5576 / S.3228, engrossed July 24, 2026)).

Pluribus News reported that the bill would make Massachusetts the fourth state to regulate the largest AI models against catastrophic harm (Source: pluribusnews.com). The Cape Cod Times reported that the bill requires large developers to identify risks of potential catastrophe such as the death of more than 50 people (Source: capecodtimes.com); the statutory threshold is "not less than 50 people," and carries an alternative $1 billion property-damage limb the coverage does not mention.

Other AI and technology provisions

Three further technology amendments were adopted alongside Amendment 471. Amendment 94 criminalizes the creation, possession or sharing of AI-generated child sexual abuse material, with penalties of up to 10 years' imprisonment for sharing, up to five years for creating or prompting such an image, and up to 10 years where the image depicts a real child. Amendment 118 prohibits the possession, operation, manufacture or sale of robots equipped with firearms, chemical agents, weaponized lasers or explosives and the use of a robotic device to criminally harass or physically restrain a person, punishable by up to five years, with a waiver process for defense contractors, an exemption for the military, and a warrant process for law-enforcement bomb disposal. Amendment 374 bans the operation of cryptocurrency kiosks in Massachusetts (Source: malegislature.gov).

On the spending side the bill authorizes $100 million for the defense sector, $75 million for the development and application of AI, $25 million for robotics research and development, and $100 million in immediate bridge funding for scientific research at public higher education institutions (Source: malegislature.gov).

Positions

Anthropic supports the AI language. Google and OpenAI oppose it. In letters dated August 18 and 19, 2026 to the conference committee and to Senator Barry Finegold, who has led the AI language, OpenAI head of U.S. state policy and partnerships Donnie Fowler and global affairs vice president Ann O'Leary urged Massachusetts to adopt a requirement closer to the Illinois law passed in May, and warned that "fragmented oversight" would be an "unhappy result" (Source: theinformation.com).

The alignment differs from the Illinois pattern, where both OpenAI and Anthropic endorsed SB 315 in May 2026. OpenAI's letters do not oppose independent review as such; they argue for the annual-audit model it has already accepted over the four-month cadence Massachusetts proposes.

Outside the labs, the AI safety organizations Encode and the Secure AI Project sent the conference committee a letter of support. The American Innovators Network, affiliated with Andreessen Horowitz and Y Combinator, is urging the committee to reject the provision; its executive director Jeremy Kudon said the requirements would create barriers for small companies and for open-weight model use in the state (Source: theinformation.com).

Open questions

• Whether the committee of conference appointed July 30, 2026 retains, modifies or drops chapter 93M is not resolved as of September 4, 2026.
• The engrossed text sets chapter 93M to take effect July 1, 2027 while placing a January 1, 2027 floor on the audit and independent-evaluation requirements; whether the conference report reconciles the dates is not recorded.
• The OpenAI letters of August 18 and 19, 2026 to the conference committee and to Senator Finegold are described by The Information as reviewed by that publication; no public copy has been located, so their full argument rests on that account.
• Chapter 93N's chatbot-disclosure requirement carries no separate effective-date section in the engrossed text, and its commencement is therefore not established here.

Relationships

• depends-on: Massachusetts Chapter 93M — Transparency in Frontier Artificial Intelligence Act (H.5576 / S.3228, engrossed July 24, 2026) — the engrossed statutory text
• related: Illinois SB 315 (frontier safety framework with mandatory third-party audits) — the annual-audit model OpenAI urges Massachusetts to adopt instead
• related: California SB 53, New York RAISE Act — the other enacted state frontier-AI frameworks
• depends-on: AI Federalism
• regulated-by: Anthropic (supports), OpenAI (opposes), Google DeepMind (opposes)
• related: American Innovators Network (AIN), Secure AI Project, Counting state AI legislation