Sam Altman recently laid out five principles guiding OpenAI's work toward AGI. They're framed as the moral and strategic bedrock for ensuring artificial general intelligence benefits all of humanity. Reading them, I'm struck by how simultaneously earnest and evasive they feel.
The principles themselves aren't controversial. They talk about broad benefit, safety, cooperation, and building AGI that empowers people. But principles are cheap. Execution is where the rubber meets the road, and OpenAI's track record shows a widening gulf between the ideals on the page and the decisions in the boardroom.
Let's dig into each principle and assess how well OpenAI is actually living up to them.
Principle 1: AGI Should Benefit All of Humanity
The first principle declares that OpenAI exists to ensure AGI benefits everyone, not just shareholders or specific nations. It explicitly commits to using influence to prevent uses that harm humanity or concentrate power inappropriately.
Noble stuff. But here's the tension: OpenAI is now a capped-profit entity with Microsoft as its primary investor and compute provider. Microsoft has exclusive licensing to OpenAI's pre-AGI technologies. That's a structural arrangement that fundamentally concentrates power.
The GPT-4 API pricing puts state-of-the-art language models out of reach for most individuals and nearly all researchers in developing countries. If you're a student in Kenya or a small nonprofit in Peru, you're not accessing this technology on equal footing with Meta or Google. The economic moat is real and growing.
The Deployment Paradox
OpenAI argues that broad deployment helps them learn and improve safety. But "broad" here means "available to those who can pay" or "integrated into products sold by Microsoft." That's not the same as accessible to all of humanity.
The counterargument—that subsidized or free access would be unsustainable—holds water financially. But it also proves the point: the current model structurally cannot deliver on universal benefit. You can't simultaneously optimize for profit at scale and guarantee equitable access.
Principle 2: Long-term Safety Over Short-term Deployment
The second principle commits OpenAI to invest heavily in safety research and slow down deployment if needed. They promise to work with other institutions and cooperate even with competitors on safety.
This principle has aged interestingly. In 2023 alone, we saw the rushed launch of GPT-4 with minimal external red-teaming, the rapid rollout of ChatGPT plugins before most safety researchers could assess risks, and a public spat with former employees over NDAs that allegedly prevented them from discussing safety concerns.
The departure of Ilya Sutskever and the dissolution of the Superalignment team raised serious questions. If your co-founder and chief scientist—someone deeply invested in safety—walks away, that's a signal. When key safety researchers leave citing concerns about the company's direction, that's more than a signal. That's a siren.
The Pressure Cooker
The competitive dynamics with Anthropic, Google, and now Meta have created a race-to-deploy pressure that makes the "long-term safety" commitment feel hollow. Every delay is a market share loss. Every pause is an opportunity for competitors to leapfrog.
OpenAI's own actions suggest they feel this acutely. The o1 models launched with impressive capabilities but also new jailbreak vectors and reasoning opacity that makes alignment verification harder, not easier. That's not what prioritizing long-term safety looks like.
Principle 3: Technical Leadership Ensures Influence
The third principle argues that OpenAI must stay at the frontier to have maximal impact on AGI development. If they fall behind, they can't steer the field toward safety.
This is the most strategically honest principle. It's essentially saying: we need to win the race to have a seat at the table. There's realpolitik wisdom here, but also danger.
The assumption is that technical leadership naturally translates to positive influence over the field. But leadership can mean many things. Google had technical leadership in search and used it to build an ad surveillance empire. Facebook had leadership in social graphs and used it to optimize for engagement over wellbeing.
The Influence Question
OpenAI's current influence manifests mostly through closed models that others can only access via API. That's influence, but it's influence through control and gatekeeping. Compare that to Meta's Llama releases, which—whatever their flaws—have genuinely democratized access to capable foundation models.
Staying at the frontier also means spending billions on compute, which necessitates the profit motive, which creates the misalignment with principle one. The logic is circular: we need money to stay ahead, we need to stay ahead to have influence, we need influence to ensure safety. At what point does the means compromise the end?
Principle 4: Cooperative Orientation with Other Institutions
The fourth principle commits to cooperation with other research and safety institutions. OpenAI pledges not to compete on safety and to share research when it helps.
The irony is thick here. OpenAI was founded on radical transparency—the name itself signaled open-source AI. That mission evaporated quickly. GPT-2 was initially held back out of safety concerns, but that precedent became a blanket justification for closed development.
Today, OpenAI publishes system cards and safety documentation, but the models themselves are black boxes. The training data, the architecture details, the fine-tuning procedures—all proprietary. You can't independently verify their safety claims because you can't see under the hood.
The Collaboration Mirage
OpenAI does participate in industry groups like the Frontier Model Forum and shares some safety research. But cooperation is limited to areas that don't threaten competitive advantage. That's not cooperative orientation—that's strategic selective sharing.
Meanwhile, researchers at universities can't reproduce OpenAI's work, can't verify their safety claims, and can't build on their advances without going through a commercial API that can change terms or pricing at will.
Principle 5: Building AGI That Empowers Humanity
The fifth principle envisions AGI as a tool that empowers people to achieve more, rather than replacing human judgment or concentrating decision-making power.
This is the most vibes-based principle. What does "empowerment" mean in practice? If ChatGPT helps a student cheat on an essay, is that empowerment? If it helps a doctor draft notes faster, is that empowerment or deskilling? If it automates a customer service job, who's being empowered?
The framing assumes a neutral tool that amplifies human agency. But tools aren't neutral. They encode assumptions, they shape workflows, they redistribute power. GPT-4 integrated into Microsoft Office doesn't empower everyone equally—it empowers those who already have Microsoft licenses and high-end compute.
The Displacement Question
OpenAI has been notably silent on labor displacement. The principles don't address what happens to the millions of people whose jobs become automatable. There's no discussion of retraining, safety nets, or economic redistribution.
Empowerment for some can mean disempowerment for others. Until OpenAI grapples with that tension explicitly, this principle reads more like marketing copy than serious ethical commitment.
The Pattern: Aspirational Words, Pragmatic Choices
Looking across all five principles, a pattern emerges. The language is idealistic, but the decisions are pragmatic to the point of contradiction. OpenAI wants to benefit all of humanity while maximizing profit. They want long-term safety while racing to ship. They want cooperation while staying closed.
None of this makes OpenAI uniquely bad. Anthropic has its own tensions. Google talks about responsible AI while integrating half-baked multimodal features into Search. Meta released Llama open-weights but still hoovers up user data for training.
The AI industry as a whole is navigating genuinely hard tradeoffs between safety and capability, between access and security, between speed and caution. But OpenAI's principles set a bar that their actions increasingly fail to clear.
What Would Accountability Look Like?
If OpenAI were serious about these principles, we'd see different choices:
- Truly independent safety oversight with veto power over deployment, not just advisory boards that can be dissolved when inconvenient
- Transparent reporting on alignment metrics, capability benchmarks, and failure modes—not just curated demos and cherry-picked examples
- Economic models that decouple frontier research from profit maximization, perhaps through differential pricing or open-weights releases after safety validation
- Labor impact plans that address displacement explicitly, with commitments to retraining and transition support
None of these are easy. Some might be financially impossible under current incentive structures. But that's exactly the point: the principles claim a moral high ground that the business model can't sustain.
The Uncomfortable Truth
Principles matter, but only when they constrain behavior. When principles become aspirational statements that don't shape actual decisions, they're worse than useless—they provide moral cover for business as usual.
OpenAI's principles are well-intentioned. I don't doubt that Altman and others genuinely care about AGI safety and broad benefit. But good intentions without structural accountability tend to erode under competitive pressure.
The question isn't whether OpenAI's principles are good. They are. The question is whether the company has the institutional courage to let those principles limit its growth, slow its deployment, or constrain its business model when they conflict.
So far, the evidence suggests the answer is no. And that gap between stated principles and revealed preferences is the story worth watching as we move closer to systems that might actually deserve the AGI label.