The Great American AI Act is a bipartisan House discussion draft that would replace state rules on how frontier AI models are developed with a single federal regime for three years. Representatives Jay Obernolte (Republican, California) and Lori Trahan (Democrat, Massachusetts) released the 269 page draft on 4 June 2026 to gather comment before filing it. As of November 2026 it has not been introduced, has no bill number and has had no hearing or markup, so none of it is law. This guide works from the sponsors' own section-by-section summary and checks each point against independent legal analysis. It explains who would be covered, what the frontier duties and the pre-emption clause would do, how the draft compares with the state frontier laws already in force, and what security, AI platform and governance teams should do while the text is still moving.
Source note: obligations and figures come from the section-by-section summary published by the sponsors unless another source is named. Secondary sources are DLA Piper, Tech Policy Press, the Future of Privacy Forum, the Cloud Security Alliance and Cybersecurity Dive. A discussion draft can change completely before introduction. Confirm against the official text before relying on it.
Key takeaways
- It is a draft, not a bill. Released 4 June 2026 for comment; as reported on 20 October 2026 it still had no bill number, committee referral, hearing or markup.
- Frontier duties mirror California. Large frontier developers (over 500 million US dollars in annual revenue that have trained a frontier model) would publish a frontier AI framework and transparency reports and report critical safety incidents to CAISI within 15 days, or 24 hours where there is imminent risk of death or serious injury.
- Licensed auditors are the new element. The Center for AI Standards and Innovation (CAISI) would license Independent Verification Organisations (IVOs) that audit large frontier developers every six months and can refer violations to federal and state attorneys general.
- Pre-emption is narrow and temporary. Section 121 would displace state and local laws that specifically regulate AI model development, but not laws on deployment or use, laws of general application or common law, and it would sunset three years after enactment.
- Most enterprises are affected indirectly. Deployers keep their state obligations either way. The practical work is vendor evidence, incident handling with model providers and documentation that maps to NIST AI RMF and ISO/IEC 42001.
Status and timeline
| Date | Event | Source |
|---|---|---|
| 4 June 2026 | Discussion draft and section-by-section summary released by Obernolte and Trahan, with Representatives Franklin, Subramanyam, Houchin and Peters supporting | Obernolte press release |
| 4 June 2026 | Co-chairs of the House Democratic Commission on AI oppose the draft within hours | DLA Piper; Tech Policy Press |
| 9 June 2026 | Future of Privacy Forum publishes a line by line comparison with the state frontier laws | FPF |
| July 2026 | Sponsors still collecting feedback; no markup or floor timetable | Cloud Security Alliance |
| 20 October 2026 | Still a discussion draft: no bill number, referral, hearing or markup, and no revised draft reported | Forkast |
| Enactment plus 3 years | Pre-emption in section 121 sunsets; DLA Piper reports the section 111 and 112 duties sunset after three years too | Section-by-section; DLA Piper |
Trahan has called the draft an "opening bid" and estimated a December 2029 sunset if it passed in late 2026 (Tech Policy Press). Supporters include BSA and ITI; opponents include Public Citizen, Public Knowledge and the AFL-CIO. We do not know whether it will be introduced or in what form. Its sponsors sit on the House Energy and Commerce Committee, which would be the likely committee of referral.
What the draft contains
The draft has four titles and folds in twelve earlier bipartisan bills, including the Future of AI Innovation Act (S. 3952), the AI Whistleblower Protection Act (H.R. 3460), the AI Fraud Deterrence Act (H.R. 6306) and the WIMWIG Act (H.R. 5079) on cyber information sharing (DLA Piper).
| Title and section | What it would do | Who it touches |
|---|---|---|
| I, Sec. 102 CAISI | Codifies the Center for AI Standards and Innovation in Commerce: voluntary security standards (adversarial robustness, model tampering, supply chain), evaluations, IVO licensing; 100 million US dollars a year for 2027 to 2029 plus fee authority | Frontier developers, evaluators |
| I, Sec. 111 Transparency | Frontier AI framework, model transparency reports, critical safety incident reports, redaction rights | Frontier and large frontier developers |
| I, Sec. 112 IVOs | Licensed third party audits and assessments of catastrophic risk mitigation | Large frontier developers, audit firms |
| I, Sec. 113 Whistleblowers | Anti-retaliation for employees and contractors who report violations of federal AI law | AI companies and staff |
| I, Sec. 121 Federalisation | Three year pre-emption of state laws specifically regulating model development | States, developers |
| I, Sec. 123 Documentation | NIST pilot of a model documentation template (developer, release date, training data cutoff, languages, terms) | Developers, buyers |
| I, Sec. 131 and 132 Fraud | Raises maximum fines for mail, wire and bank fraud and money laundering from 1 million to 2 million US dollars, with higher penalties where AI is used; penalties for AI impersonation of federal officials | Fraud and financial crime teams |
| II, Sec. 251 WARN | 60 day mass layoff notices must say when AI is a substantial factor, describe it and estimate the share of job losses | Employers |
| III, Sec. 301 | Extends the Cybersecurity Act of 2015 information sharing protections to 2035 | All sharing organisations, ISACs |
| III, Sec. 311 and 321 | CISA grants to critical open source maintainers, with large frontier developers giving them model access for security work; GAO review of model weight security | Open source projects, AI labs |
| IV | Testbeds, international standards coalitions, NAIRR in statute, research security | Research and standards bodies |
For the information sharing law Title III would extend, see the Cybersecurity Information Sharing Act of 2015 Guide.
Who would be covered by the frontier duties
The draft borrows the state definitions. FPF reports a frontier model is a foundation model trained with more than 10^26 computational operations, and that the draft adds a revenue test to the base tier.
- Frontier developer: has trained a frontier model and has gross revenue above 50 million US dollars (FPF; Cloud Security Alliance). Carries transparency and incident reporting duties.
- Large frontier developer: annual revenue above 500 million US dollars. Carries the full set: a public frontier AI framework, IVO audits and model access for open source maintainers. Trahan has said the threshold is still under evaluation.
- Deployers and customers: no direct duty under Title I. The draft regulates development, not products or use (Tech Policy Press), so deployers' state obligations survive.
Practical test: if you fine tune or host someone else's model, you are not obviously a frontier developer, because the trigger is training a model above the compute threshold. Take advice if you train large models in house.
Key obligations for frontier developers
Frontier AI framework and transparency
Large frontier developers must write, implement, follow and publish a frontier AI framework covering catastrophic risk thresholds and assessments, cybersecurity of model weights, and internal and external deployment decisions. Before or alongside deploying a new frontier model they publish a report with release date, languages, output modalities, intended use, restrictions, risk assessments and mitigations. Redaction is allowed for trade secrets, cybersecurity, public safety and national security. Weight security is the identity heavy part: who and what can read, copy or export weights. See Securing AI Model Weights and the Frontier AI Model Risk Frameworks Guide.
Critical safety incident reporting
Reports go to CAISI within 15 days of discovery, or 24 hours if there is imminent risk of death or serious injury. State attorneys general may opt in to receive them. FPF notes two differences from the state laws: the draft's definition does not require actual harm, and it omits the state limb covering models that use deceptive techniques to evade developer controls or monitoring.
Independent verification
CAISI licenses IVOs and can revoke a licence for loss of independence, obsolete methods or non-compliance. Large frontier developers must retain one and give it sufficient access. Audits run semi-annually and ad hoc at the Director's request; IVOs report to CAISI and may refer violations to the Attorney General or opted-in state attorneys general, with referral mandatory for imminent catastrophic risk (DLA Piper). IVOs would be immune from claims arising from a model they audited.
Whistleblower protection
Remedies for retaliation include reinstatement, two times back pay with interest, compensatory damages, costs and fees. See the Cyber Whistleblowing Guide.
How pre-emption would work
Section 121 pre-empts state and local laws or regulations that specifically target AI model development. It does not pre-empt laws of general application, common law remedies, or laws that regulate AI use or deployment. DLA Piper describes development as acts done or directed by a developer before deployment, and confirms laws on companion bots and commercial chatbot interactions are unaffected.
| State law type | Likely effect if enacted as drafted | Why |
|---|---|---|
| Frontier developer laws (California SB 53, New York RAISE Act, Illinois SB 315) | At risk of pre-emption for three years | They impose developer side duties before deployment; analysts flag all three (CSA, DLA Piper) |
| Deployer and algorithmic discrimination laws (for example Colorado) | Survive | They regulate use and deployment decisions |
| Privacy, consumer protection, civil rights | Survive | Laws of general application are saved |
| Chatbot and companion bot rules | Survive | Post-deployment conduct |
| Tort and negligence claims | Survive | Common law remedies are saved |
The weak point is the phrase "specifically regulating the development", which the Cloud Security Alliance notes is undefined. Mixed state laws that impose both development and deployment duties would invite litigation over which parts fall. FPF also notes the draft drops the state disclosure filings on ownership and business information and has lighter internal whistleblower process duties than California and Illinois.
| Point | Great American AI Act draft | State frontier laws (per FPF) |
|---|---|---|
| Incident report deadline | 15 days, or 24 hours for imminent risk | California 15 days; New York and Illinois 72 hours |
| Penalty | Up to 1 million US dollars per violation, each day a separate violation | California up to 1 million; New York and Illinois up to 3 million for subsequent violations |
| Third party audit | Licensed IVOs, every six months | Only Illinois requires audits, with no licensing scheme |
| Enforcer | US Attorney General and state attorneys general | State attorneys general |
See the California AI Laws Guide, the New York RAISE Act Guide and the US State AI Laws Tracker.
Enforcement and penalties
- Civil penalties: up to 1 million US dollars per violation of sections 111 and 112, each day a separate violation (section-by-section; FPF).
- Injunctions: available to the US Attorney General and state attorneys general.
- Referral route: IVO findings flow to CAISI and, by referral, to federal and opted-in state attorneys general, giving states an enforcement role even where their own law is pre-empted.
- Fraud uplift: higher maximum fines for AI-assisted fraud apply to anyone, not just developers.
What it means for AI agents and non-human identities
The draft says nothing directly about agent identity, but three provisions turn on machine access control.
- Model weight security: the frontier AI framework must cover weight cybersecurity, and GAO would assess weight protections. Service accounts, CI pipelines and evaluation harnesses that can read weights are the attack surface. Map controls to NIST SP 800-53 AC-6, SC-12 and SC-28.
- Auditor and maintainer access: IVOs need "sufficient access", and large frontier developers must give open source maintainers model access for security work. Each grant is a new external identity with scoped, time bound, logged access. See the Open Source Project Maintainer Security Guide.
- Incident evidence: a 24 hour clock needs telemetry ready before the incident: which agent, tool or credential did what. See the AI Incident Reporting Guide and the Agent Identity and Access Guide.
Decision guidance: what to do while it is a draft
- Frontier developers: comply with California SB 53 and New York RAISE now; the draft copies most of their structure, so an SB 53 programme plus 24 hour escalation and audit readiness covers most of it.
- Large enterprises deploying frontier models: do not plan around pre-emption. Your duties are deployment duties, which the draft saves. Put incident notice, audit rights and framework disclosure into model contracts.
- Small organisations: no direct duty. Track it quarterly and keep vendor evidence on file.
- Open source maintainers: watch section 311; grants and model access would come with access governance obligations.
Compliance steps
- Classify your role: frontier developer, large frontier developer, deployer or customer, with the revenue and compute evidence.
- Inventory production AI systems: tag each by developer or deployer role, model provider and states where it operates (CSA recommendation).
- Map state duties to the draft: for each state frontier obligation, note whether section 121 would displace it; keep complying until a law actually changes.
- Build a two speed incident path: 24 hour escalation for imminent harm, 15 day reporting otherwise, and 72 hours for New York and Illinois. Test it with a tabletop.
- Prepare for audit: evidence packs for framework adherence, weight access reviews and access logs an external auditor could read.
- Fix contracts: incident notification, audit rights and risk allocation with AI vendors.
- Anchor on neutral frameworks: document against NIST AI RMF and ISO/IEC 42001 so evidence survives whichever law applies. See the NIST AI RMF Guide and ISO/IEC 42001 Guide.
- Roll back cleanly: if the draft dies or changes, keep the inventory and incident path; they serve the state laws too.
Detection and measurement
- Time to classify: hours from AI incident detection to a decision on whether it is a critical safety incident; target under 12 hours to protect the 24 hour clock.
- Weight access coverage: percentage of identities (human and non-human) with weight read or export rights reviewed in the last 90 days.
- Vendor evidence coverage: share of frontier model suppliers with a published framework and contractual incident notice on file.
- Audit finding closure: median days to close internal or IVO-style findings.
- Tracking freshness: days since legal last checked the draft's status and the state laws it would affect.
Framework mapping
| Framework | Reference | Link to the draft |
|---|---|---|
| NIST CSF 2.0 | GV.OC-03 | Legal and regulatory requirements understood and managed |
| NIST CSF 2.0 | GV.SC-07 | Supplier risk for frontier model providers |
| NIST CSF 2.0 | RS.CO-02 | Notifying CAISI and other stakeholders of incidents |
| NIST SP 800-53 Rev 5 | IR-4, IR-6 | Incident handling and reporting |
| NIST SP 800-53 Rev 5 | CA-2, SR-6 | Control assessments and supplier reviews, the IVO analogue |
| NIST SP 800-53 Rev 5 | AC-6, SC-12, SC-28, AU-9 | Weight access, keys, encryption at rest and protected logs |
| ISO/IEC 27001:2022 | A.5.31, A.5.24, A.5.19 | Legal requirements, incident planning, supplier relationships |
Common failure patterns
- Pausing state compliance: treating a discussion draft as if pre-emption had happened. Nothing is displaced until a law is enacted.
- Assuming pre-emption covers deployers: the savings clauses keep deployment, use and general laws in place.
- One incident clock: building only for 15 days when New York and Illinois require 72 hours and the draft adds 24 hours for imminent harm.
- Unscoped auditor access: granting external auditors or maintainers broad standing credentials rather than time bound, logged access.
- Weight access by default: pipelines and service accounts with read access to weights that nobody reviews.
Practitioner checklist
- Record your role under the draft and the state frontier laws, with evidence.
- Keep complying with California, New York, Illinois and other state laws in force.
- Stand up a 24 hour, 72 hour and 15 day AI incident path and test it.
- Review every identity that can read or export model weights.
- Design scoped, expiring access for auditors and external researchers.
- Add incident notice and audit rights to frontier model contracts.
- Map evidence to NIST AI RMF, ISO/IEC 42001 and SP 800-53.
- Assign an owner to check the draft's status each quarter.
Standards and references
- Office of Representative Lori Trahan: Great American AI Act discussion draft, section-by-section
- Office of Representative Jay Obernolte: Obernolte, Trahan release a discussion draft of the Great American AI Act
- DLA Piper: Unpacking the Great American AI Act
- Future of Privacy Forum: Frontier AI goes federal, how the Great American AI Act compares to state laws
- Tech Policy Press: Unpacking the Great American Artificial Intelligence Act of 2026
- Cloud Security Alliance: The Great American AI Act, a 2026 compliance problem
- Cybersecurity Dive: House AI bill and its cybersecurity provisions
- Forkast: status report on the Great American AI Act, October 2026
Related NHI Mgmt Group resources: US Federal AI Policy Guide · US State AI Laws Tracker · AI Kill Switch Act Guide · Frontier AI Model Risk Frameworks Guide. This guide summarises the Great American AI Act discussion draft and is not legal advice.