Sam Altman outlines five principles that double as justification for OpenAI's business decisions

OpenAI's CEO has laid out five guiding principles for the company's future work. They also serve as a rationale for some of OpenAI's more unconventional business moves.

In a post titled "Our principles," OpenAI CEO Sam Altman lays out five guiding principles. The central premise: power over superintelligence will either be concentrated among a few companies or distributed among the people. OpenAI says it's aiming for the latter.

What's notable is what Altman doesn't present as an alternative: a pluralistic landscape of competing providers. His focus is on access - end users equipped with AGI. Under this logic, OpenAI can remain a central provider while still framing itself as a vehicle for decentralization.

The first principle, democratization, combines two components: AI access for everyone, and decisions about AI made through democratic processes rather than by AI labs alone. How that's supposed to work in practice when AI labs are funneling millions into lobbying through Super PACs remains unclear. It also reads as an indirect jab at competitor Anthropic, whose red lines were a thorn in the Pentagon's side.

The second principle, empowerment, grants users broad autonomy, paired with a commitment to minimizing catastrophic harm, local damage, and "corrosive societal effects." When in doubt, OpenAI says it will err on the side of caution and only loosen restrictions as evidence grows.

With the third principle, universal prosperity, Altman provides a justification for business decisions that he admits look "weird": massive compute purchases despite relatively modest revenue, vertical integration, and building data centers around the world. The argument is that the cost of AI infrastructure needs to drop dramatically. Altman also hints that governments may need to explore new economic models to distribute the value AI creates more broadly.

Iterative deployment is just one piece of the safety puzzle

The fourth principle, resilience, calls for a society-wide approach. For pathogen risks, for example, countermeasures need to work regardless of the specific threat. For cybersecurity, it means rapidly deploying models to secure open-source software and critical infrastructure.

Iterative deployment - gradually rolling out new capabilities - is just one element of this approach. Technical alignment and secure systems remain separate challenges.

Altman acknowledges that there may be phases where OpenAI needs to collaborate with governments, international agencies, and other AGI projects to solve alignment, safety, or societal problems before moving forward. Funding from the OpenAI Foundation is meant to support this work.

The fifth principle, adaptability, explicitly reserves the right to change course. As one possible example, Altman suggests that empowerment may eventually need to be weighed against greater resilience. He points to the debate over releasing the GPT-2 weights as a case study: the concerns at the time turned out to be overblown in hindsight, but the episode gave rise to the iterative deployment strategy.

The principles also likely serve as an indirect response to the criticism OpenAI faced around its Pentagon deal. Altman closes the post by inviting criticism, saying OpenAI will make mistakes and correct them.