OpenAI lists 5 recommendations for advancing AI progress

OpenAI has published a wide‑ranging policy note, “AI progress and recommendations,” setting out how frontier systems have advanced and the steps it believes governments, companies and civil society should now take.

Released on November 6, 2025, the paper has argued that today’s models already outperform top humans at certain intellectual competitions and that society needs shared standards, stronger oversight and better measurement to ensure benefits are broadly distributed.

Widening gap between perception and capability

The company has said the public still tends to view AI as chatbots or improved search, while current systems have handled far more complex reasoning. It has also reported that the “cost per unit” of a given level of intelligence has fallen steeply, estimating roughly a 40x annual decline in recent years.

Looking ahead, the company has forecast that AI could make “very small discoveries” by 2026 and more significant ones from 2028, while cautioning that progress remains “spiky” with notable weaknesses.

Five areas for action

• Shared standards and insights: Frontier labs have been urged to agree on common safety principles, share research on new risks and reduce race dynamics.

• Public oversight and accountability: Governance should be commensurate with capability and led by public institutions.

• Building an AI resilience ecosystem: Develop tools, standards and incident response analogous to cybersecurity.

• Ongoing reporting and measurement: Track real‑world impacts, including on jobs and safety, rather than relying on forecasts alone.

• Building for individual empowerment: Ensure adults can use AI on their own terms within democratically set bounds, expanding access much like a core utility.

OpenAI has compared the desired regime to building codes and fire standards—boring but proven protections that have saved lives—arguing that shared control evaluations across labs could be helpful.

Two oversight models for an uncertain trajectory

The note has outlined two schools of thought. In one, AI behaves like a “normal technology”, evolving at a pace society can absorb using conventional regulation. In the other, superintelligence develops and diffuses faster than any past technology, requiring novel coordination with national safety institutes and the executive branches of multiple countries—especially on dual‑use areas such as biodefence and the prevention of bioterrorism. In either case, the “high‑order bit”, the paper has argued, should be accountable to public institutions.

What has OpenAI proposed for resilience?

It has called for an ecosystem—spanning software, standards, monitoring and emergency response teams—akin to how cybersecurity has matured since the early internet. The company has argued that such an ecosystem will not eliminate risk but can reduce it to socially acceptable levels and enable trust in AI infrastructure.

Measurement, empowerment, safety

Because predictions about labour markets and model behaviour are unreliable, OpenAI has urged continuous reporting from labs and governments on concrete impacts. It has also emphasised access: in its view, advanced AI should become a widely available utility that empowers individuals. The paper has reiterated the firm’s stance that no one should deploy superintelligent systems without robust alignment and control, work that it has said still requires significant technical progress.

Leadership context

OpenAI’s CEO and co‑founder Sam Altman has remained at the helm following a March 2024 governance review that reaffirmed his leadership; he has since been re‑appointed to the board.