Cohere Response to the Request for Information on the Development of an Artificial Intelligence (AI) Action Plan
Published March 15, 2025 · Printed on page 1 directly under the title and the docket reference "[FR Doc. 2025-02305] (90 FR 9088)"; it is the only date in the document. The PDF's embedded modification date is 2025-03-19, four days later, which reflects when the file was saved for publication rather than a different version.
Not law. This is a company's own public position on AI regulation. It is not law, and it carries no legal force.
What it argues for
This is Cohere's formal submission to the National Science Foundation and the White House Office of Science and Technology Policy on what became America's AI Action Plan, filed under Federal Register docket 90 FR 9088 and signed by A.J. Bhadelia, Cohere's North America Government Affairs and Policy lead. It speaks for an enterprise-focused model developer and argues from that position throughout: model performance on general benchmarks is converging, the value lies in tailored, secure, sector-specific deployment, and "Public policy should therefore be focused on enabling these opportunities, rather than the highly speculative hazards of superintelligent machines." It sets out four priorities: "Make AI adoption a centerpiece", reform government procurement so that startups can compete, "Prioritize regulatory flexibility: Pursue risk-based governance and a sectoral approach to ensure security and enable innovation", and invest in public data, compute and talent. On regulation it rejects broad horizontal AI laws in favour of "a federal sector-led regulatory approach that leverages the deep expertise of existing agencies and regulatory frameworks", asks agencies to clarify how existing rules apply before writing new ones, warns that "State-level legislation, in particular, poses an existential compliance risk for startups", and wants common definitions set, "ideally legislatively". On testing it lists four limits: "Avoiding blanket licensing or certification requirements for AI models", "Setting performance-based standards instead of pre-market approvals", focusing on risks that are known, measurable or observable, and "Avoiding compulsory sharing or public disclosure of trade secrets and other proprietary information." It wants AI safety institutes to be "centers of scientific excellence and technical expertise rather than disseminate regulation", backs copyright rules that allow training, asks for procurement that does not favour open or closed source over the other, and ends with a list of investment asks including funding the National AI Research Resource. Throughout it presents heavy regulation as a gift to incumbents and to China: "Overregulating the industry with unnecessary burdens risks slowing technological advancement and potentially conceding further ground to China".
Stated positions (14)
- Policy should aim at adoption, not speculative risk: "Public policy should therefore be focused on enabling these opportunities, rather than the highly speculative hazards of superintelligent machines."
- Government and private-sector adoption should be the centrepiece of the plan: "Make AI adoption a centerpiece: Promote policies that accelerate government and private sector adoption of AI."
- Large-model regulation entrenches incumbents: "Recent U.S. AI regulatory proposals originate from fear about the largest models, with rigorous requirements that would entrench large, existing incumbents."
- Regulation should be sectoral and run by existing regulators: "we instead propose a federal sector-led regulatory approach that leverages the deep expertise of existing agencies and regulatory frameworks", and the plan should "Designate primary regulatory authority to agencies with sector-specific jurisdiction."
- Existing law should be clarified before new law is made: "The success of this approach requires agencies to first clarify how existing regulations apply to AI before creating new ones."
- State AI laws are a particular threat to startups: "State-level legislation, in particular, poses an existential compliance risk for startups who cannot keep up with dozens of differing definitions and standards."
- Definitions should be unified, preferably by statute: "Create common definitions – ideally legislatively – that allow the public and private sector to work together effectively."
- No licensing or pre-market approval of models: the plan should keep an innovation-friendly environment by "Avoiding blanket licensing or certification requirements for AI models" and "Setting performance-based standards instead of pre-market approvals."
- Scrutiny should follow risk tiers and known harms: it asks for "a risk-tiered framework that directs the most scrutiny and safeguards toward high-risk AI applications" and for "Focusing on risks that are known, measurable or observable for frontier, high-risk AI use cases."
- No forced disclosure of proprietary information: "Avoiding compulsory sharing or public disclosure of trade secrets and other proprietary information."
- AI safety institutes should be scientific bodies, not regulators: "we believe the most pertinent role for these types of bodies is to serve as centers of scientific excellence and technical expertise rather than disseminate regulation."
- Copyright should permit AI training: "Copyright is designed to promote creativity and human progress – not to impede the development of new technologies", and "We caution against imposing new obligations that would effectively push AI progress offshore or into the hands of only the largest corporations."
- Procurement should be opened to startups and stay neutral between open and closed models: "Current federal procurement rules and lengthy contracting cycles favor established vendors, creating high barriers for startups", and the government should "Ensure government contracting or public policy doesn’t unfairly support open or closed source over the other."
- Public research infrastructure should be funded: "Fund the proposed National AI Research Resource (NAIRR) to provide shared computing resources", alongside open data, supercomputer time, tax incentives and "visas for AI experts".
About this document
An 11-page PDF submission, about 3,600 words, with page numbers but no letterhead graphics or footnotes (a handful of hyperlinks point to Cohere's own blog posts and research, a GAO blog post and a LinkedIn letter by Aidan Gomez). Page 1 carries the title, the Federal Register docket reference and the date March 15, 2025. The body runs: an unheaded introduction listing the four priorities as bullets; "About Cohere"; "A Practical Overview of the State of AI"; and "Cohere's Core Priorities for U.S. AI Leadership", in four numbered sections: "1. Make AI Adoption a Centerpiece", "2. Reform and Modernize Government Procurement for AI", "3. Prioritize Regulatory Flexibility through Risk-Based & Sectoral Governance" (with the sub-headings "Regulatory Flexibility" and "Testing and Assessment") and "4. Invest in AI R&D Infrastructure – Data, Compute, and Skills". Most sections end in bulleted asks addressed to OSTP. A "Conclusion" closes with "Respectfully submitted," and the signature of A.J. Bhadelia, "North America Government Affairs and Policy", Cohere. It refers to Cohere's Secure AI Frontier Model Framework as its "recent Security Framework" and cites Executive Order 14141, the NIST AI Risk Management Framework, the OPEN Government Data Act and work by Senators Rounds and Heinrich; it names no state law and no foreign law.
How this sits against AI law
Each stance compared with what EU and US instruments actually require. Where no instrument addresses a theme, that gap is shown rather than hidden.
Sector-led, risk-based regulation through existing regulators instead of a horizontal AI law
The US should adopt "a federal sector-led regulatory approach that leverages the deep expertise of existing agencies and regulatory frameworks", give primary authority to sector regulators, clarify how existing rules apply before writing new ones, and avoid broad, technology-focused frameworks.
The AI Act is a single horizontal regulation that applies across all sectors, with AI-specific definitions, prohibited practices, a list of high-risk uses in Annex III and separate duties for general-purpose models, enforced by national market surveillance authorities and the Commission's AI Office. It is risk-tiered, as Cohere wants, but it is precisely the cross-sector, technology-focused framework the submission argues against.
The White House's March 2026 legislative recommendations say Congress "should not create any new federal rulemaking body to regulate AI, and should instead support development and deployment of sector-specific AI applications through existing regulatory bodies with subject matter expertise and through industry-led standards." They are recommendations to Congress, not law.
One national rulebook instead of a patchwork of state laws
State AI legislation with differing definitions and standards threatens startups; the US should harmonise definitions, "ideally legislatively", and avoid fragmented policies.
As an EU regulation the AI Act applies directly and uniformly in every member state, so providers face one set of definitions and obligations across the single market rather than 27 national regimes. The rulebook is broader than Cohere would choose, but it delivers the uniformity the submission asks for in the US.
The recommendations ask Congress to "preempt state AI laws that impose undue burdens to ensure a minimally burdensome national standard consistent with these recommendations, not fifty discordant ones", while leaving states their general consumer-protection, child-protection and zoning powers. Congress had not enacted such preemption in the corpus as of this check.
No licensing, certification or pre-market approval of AI models
The Action Plan should avoid "blanket licensing or certification requirements for AI models" and set performance-based standards "instead of pre-market approvals", relying on the NIST AI Risk Management Framework and sector guidance rather than new licensing regimes.
The Act requires no licence or prior approval for general-purpose AI models; their providers carry documentation, copyright and, for systemic-risk models, evaluation and incident duties. It does require high-risk AI systems to undergo a conformity assessment before they are placed on the market, which is closer to the pre-market step Cohere wants to avoid, but that applies to systems in listed high-risk uses, not to models as such.
Executive Order 14409 sets up only a voluntary framework for pre-release government access to covered frontier models and states that nothing in that section "shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models, including frontier models."
No compulsory disclosure of trade secrets or proprietary information
Policy should avoid "compulsory sharing or public disclosure of trade secrets and other proprietary information"; confidentiality guarantees should encourage voluntary sharing with government instead.
Article 53 requires every general-purpose model provider to keep technical documentation available to the AI Office and national authorities, to give information to downstream providers, and to publish a sufficiently detailed summary of the content used for training. Article 78 binds the authorities to keep trade secrets confidential, but the disclosure itself is compulsory, which is more than Cohere would accept.
SB 53 requires large frontier developers (over 10^26 operations of training compute and more than $500 million in annual revenue) to publish a frontier AI framework and transparency reports, allowing redactions to protect trade secrets. Whether Cohere meets both thresholds was not checked.
Accelerating government adoption of AI and reforming federal procurement
OSTP should make federal and private adoption a centrepiece, fund agency pilots, and reform federal acquisition so that startups can compete, with interoperability requirements, no single-vendor lock-in and neutrality between open and closed source.
The AI Act regulates how public bodies use high-risk AI (for example through deployer duties) but contains no programme to accelerate government adoption and does not govern public procurement of AI; nothing comparable is recorded as an EU instrument in the corpus.
OMB Memorandum M-25-22 (April 3, 2025), issued after this submission, sets federal AI acquisition policy around supporting a competitive American AI marketplace, managing risk and performance across the contract lifecycle, and making acquisition efficient and timely; its companion M-25-21 directs agencies to accelerate their own use of AI.
Copyright rules that allow AI training
Copyright should not impede new technologies; the government should avoid "imposing new obligations that would effectively push AI progress offshore or into the hands of only the largest corporations" and preserve technology-neutral frameworks that allow lawful uses of data.
Article 53 obliges general-purpose model providers to put in place a policy to comply with EU copyright law, including honouring rights reservations under the text-and-data-mining exception, and to publish a summary of training content. These are new AI-specific obligations of the kind Cohere cautioned against.
The recommendations state the Administration's view that training on copyrighted material does not violate copyright law, leave the fair-use question to the courts, and say any collective-licensing legislation "should not address when or whether such licensing is required."
AI safety institutes as centres of science, not regulators
Government AI safety and security institutes should serve as "centers of scientific excellence and technical expertise rather than disseminate regulation", supporting sector regulators and industry with measurement science, testing methods, harmonised definitions and standards.
The EU's central AI body, the Commission's AI Office, is a supervisor: under the Act it monitors general-purpose model providers, can request documentation, evaluate models and require measures, and the Commission can fine providers of general-purpose models. It does technical work too, but it is a regulator in a way Cohere argued these bodies should not be.
The Action Plan gives NIST's Center for AI Standards and Innovation (CAISI) research, evaluation and standards work, including a section on building an AI evaluations ecosystem and evaluations of frontier models from China, but no regulatory power over developers.
Public investment in AI research infrastructure: data, compute and skills
The US should accelerate open federal data, fund the National AI Research Resource, expand supercomputer access, consider tax incentives for AI infrastructure and support workforce development including "visas for AI experts".
The AI Act is a product-safety and fundamental-rights law and does not fund compute, data or skills; the corpus holds no comparable EU investment instrument to set against this ask.
The Action Plan calls for building "a lean and sustainable NAIRR operations capability" and for partnering with technology companies to widen research access to private computing, models and data under the NAIRR pilot. The March 2026 recommendations separately ask Congress to make federal datasets available in AI-ready formats.
Regulatory sandboxes for testing AI
Procurement and policy should include challenge-based solicitations, pilot programmes that can start before full certification, and regulatory sandbox provisions for testing innovative approaches.
Article 57 requires each member state to have at least one AI regulatory sandbox operational, in which providers can develop and test AI systems under supervision before placing them on the market.
The recommendations say "Congress should establish regulatory sandboxes for AI applications"; the July 2025 Action Plan likewise recommends establishing regulatory sandboxes or AI Centers of Excellence.
Cohere asked Washington for almost exactly what Washington then did, and asked Brussels for the opposite of what Brussels had already done. In the US, federal policy since this filing has followed its script closely: OMB's April 2025 memoranda pushed agency adoption and reformed AI acquisition; the July 2025 Action Plan committed to deregulation, the NAIRR, regulatory sandboxes and open-weight models; Executive Order 14409 rules out any mandatory licensing or preclearance of models; and the White House's March 2026 recommendations to Congress back sector-specific regulation through existing regulators, preemption of burdensome state AI laws, and leaving AI-training copyright to the courts. Where the US diverges is at state level, where California's SB 53 imposes published frameworks and incident reporting on the largest frontier developers, which is the kind of state-by-state regime Cohere called "an existential compliance risk for startups". The EU AI Act is what Cohere argues against: a single horizontal law across sectors, with its own AI-specific definitions and risk tiers, pre-market conformity assessment for high-risk systems, documentation and training-data-summary duties for general-purpose model providers, a copyright-compliance obligation, and an AI Office that supervises and can fine model providers rather than acting only as a centre of scientific expertise. The Act does share two of Cohere's aims: one set of rules in place of national patchworks, and risk tiers that put the heaviest duties on high-risk uses.
Source
https://cohere.com/blog-assets/cohere-response-to-wh-ai-action-plan.pdf- Date on the page:
- March 15, 2025
- Source checked:
- opened and confirmed on 2026-09-30