OpenAI says the AI policy window is open and is calling for immediate, capability-based safeguards as model capabilities accelerate. The company laid out a four-track approach that mixes federal regulation, state-level stopgaps, voluntary industry standards, and international coordination — all aimed at keeping safety practices ahead of rapidly advancing systems.
At the center of OpenAI’s proposal is a push for mandatory national rules that are targeted at the laboratories developing the most advanced systems. The company describes a capability-based federal framework that would focus on frontier developers and would include common testing and independent-assessment requirements, stronger cybersecurity protections, clear incident-reporting rules, enhanced national preparedness, and shared measures to track progress toward more autonomous development processes. OpenAI stresses that such rules should be narrowly tailored so they do not impose undue burdens on startups, small developers, or researchers who are not operating near the frontier, and should avoid undermining competition or driving innovation overseas.
Recognizing that Congress may move slowly, OpenAI said it will continue to support state-level measures as interim protections. The company announced formal support for four California bills headed to the governor: SB 813, to create a process for designating qualified independent organizations capable of assessing AI risks; AB 1405, to set registration, independence, transparency and accountability requirements for AI auditors; SB 1119, to require age assurance, risk assessments, independent audits, parental controls and safeguards for youth-facing chatbots; and AB 1864, to require gene-synthesis providers and manufacturers of benchtop synthesis equipment to follow federal screening standards to guard against AI-enabled biological threats. OpenAI framed these measures as a way to demonstrate workable safeguards and to build momentum toward a unified federal approach, describing the dynamic as a form of “reverse federalism” in which state rules converge into a de facto national baseline that Congress can later codify.
OpenAI also argued that laws alone will not be sufficient and urged the industry to adopt concrete laboratory practices, naming monitoring as the most urgent area. The company emphasized monitoring for misalignment — scenarios where a model pursues objectives that violate human intent or established safeguards, even absent consciousness or malicious intent. Monitoring, OpenAI said, must be paired with clear disclosure requirements and prompt reporting when models circumvent another organization’s security controls or materially access, alter, or destroy protected systems or data.
To that end, OpenAI said it is developing an internal framework for reporting consequential misalignment incidents and for systematic monitoring of frontier models, and it invited other labs to collaborate so that federal reporting standards can be informed by operational experience. The company stressed that voluntary industry standards for frontier labs should complement, not replace, mandatory federal rules focused on the most capable developers.
OpenAI also addressed the implications of accelerating research and the prospect of recursive risks. While the company acknowledged that fully autonomous recursive self-improvement is not occurring today, it noted that AI is already accelerating some research tasks. OpenAI recommended directing that acceleration toward safety and alignment work — for example, using automated research under human supervision to make each generation safer and more controllable rather than solely more capable.
Technical safeguards during development were highlighted as necessary precautions. OpenAI listed measures such as stronger isolation for frontier research workloads, expanded monitoring of model behavior during tool-enabled training and evaluations, mandatory alignment-evaluation gates before internal deployment, and universal monitoring of full trajectories for systems like Astra. These operational controls, the company argued, should be part of a comprehensive approach that couples regulatory requirements with laboratory-level practices.
Finally, OpenAI urged the development of compatible international standards, noting that models and expertise move across borders and that failures involving the most capable systems could have global consequences. The company said the United States must establish credible domestic standards to lead internationally and warned that delaying action risks letting capabilities outpace the institutions responsible for governing them.
OpenAI’s statement frames the current moment as an opportunity to set durable, practical rules and practices while the field still has a chance to shape norms and institutions. The company’s multi-pronged proposal — federal capability-based rules, state-level interim measures, industry monitoring standards, and international coordination — is presented as a comprehensive way to keep safety measures in step with rapid technical progress. OpenAI’s full statement and related materials are available on its website at https://openai.com/index/ai-policy-window.
Source: Read the original source
Leave a Reply