Under the proposal, AI developers would provide confidential reports to a new Commerce Department office every quarter, and within four business days of any major change.
Americans for Responsible Innovation (ARI), a nonprofit AI policy group, proposed that Congress require the largest AI developers to report confidentially to the Commerce Department every quarter on how much of their AI research and development their own AI now does.
ARI first proposed quarterly reports like these in the federal AI plan it released in August. Its new policy paper sets out what the reports would cover.
Developers would report on four kinds of R&D work
A new office in the Commerce Department would receive the reports. The law itself would set four areas each report must cover. The office would then write rules listing the specific details developers must report in each area:
- Training data: AI choosing, filtering, and generating the data models learn from, and AI adjusting a model after its main training to shape how it behaves.
- Model design: AI designing how models are built and trained.
- Experiments: AI designing, running, choosing, and interpreting experiments, including training runs, the long computing jobs that produce a new model.
- Safety work: AI monitoring the work AI systems do on their own, testing models for dangerous capabilities, and investigating safety incidents.
Each report would need specific, auditable numbers. ARI’s two examples are the share of the company’s computing power used for work AI does without people, and the share of AI’s R&D work and decisions that people review.
The proposal applies to “major frontier AI developers” but does not define them. ARI’s August plan covered companies that had trained a model using at least 10^26 computing operations and had spent at least $100 million on AI training in the past year.
Major changes would be reported within four business days
Between quarterly reports, a developer would have four business days to report a major shift in how much of its R&D AI does. The clock would start when the developer found the shift, or when its own monitoring should have caught it.
The law would outline which events are considered important enough to require a report. The office could add more through its own rules. ARI gives two examples: AI running a complete R&D cycle on its own, and a company putting a new type of AI model to work that works independently in one of the four reporting areas.
These reports would include the same numbers as the quarterly report for the affected area, what changed, and what the developer did about it. The office could ask follow-up questions and require more frequent reports in the area where the change happened.
False reports could be referred to the Justice Department
The office could follow up on any report. If a developer knowingly filed a false or incomplete report, the office could refer the developer to the Justice Department. ARI also proposes a whistleblower program as a second check.
The public would see only industry-wide summaries
The reports would be exempt from the Freedom of Information Act, which lets the public request government records. ARI says this would protect companies’ security and intellectual property. The office would instead publish its own industry-wide trend reports.
Why ARI wants the reports
ARI says the reports would give policymakers early evidence of whether AI is improving faster than people can oversee it and test it for safety.
ARI says only the developers know how fast their AI models are improving. The public learns only what the companies choose to share, and each company measures its progress its own way.
Anthropic, which ARI cites, has published its own measure. It says that as of August 2026, its AI model Claude “leads” 26% of the company’s AI R&D work, up from under 1% in February. Anthropic defines “leads” as Claude completing most of a task from a short instruction while an employee supervises.
ARI says AI already writes most of the code at the leading AI companies, but has limited ability to judge on its own which lines of R&D are worth pursuing, and cannot run the training of a new model from start to finish without people. According to ARI, OpenAI aims to build an automated “AI researcher” by March 2028. Anthropic co-founder Jack Clark wrote in his May newsletter that he sees about a 60% chance of automated AI R&D by the end of 2028.
The proposal still needs Congress
The proposal is not a bill, and ARI does not name any member of Congress backing it. ARI is asking Congress to pass a law that would create the program.
