OpenAI Fired Its Conscience and Didn't Notice

OpenAI’s last ethicist left last month. She wasn’t replaced.

Not downsized into a team. Not restructured into a broader function. Just gone. The headcount is now zero. This is not a minor HR footnote — it’s a structural signal about what OpenAI believes it is and who it believes is watching.

The steelman goes like this: ethics can’t live in a single person’s job title. It has to be baked into product, engineering, policy, and leadership. A lone ethicist is theater — a role that produces memos nobody reads and gives the organization a checkbox without changing the decisions. Better to diffuse the responsibility than perform it.

The steelman is wrong. Not because ethicists are magic, but because the absence of one tells you something true about the organization’s revealed preferences. OpenAI is deep into its for-profit conversion. It is in a full sprint against Anthropic, Google DeepMind, and xAI for model supremacy. Its product org is shipping faster than its policy team can track. In that environment, a dedicated ethics function isn’t theater — it’s the one structural mechanism that can slow a decision down long enough to ask whether it should be made at all. Removing it doesn’t diffuse the responsibility. It disappears it.

Compare Anthropic. The company built its entire brand around the premise that safety and capability are complementary, not opposed. Whether you believe that’s true or cynically packaged, Anthropic’s organizational structure at least encodes the claim. Anthropic just committed to watermarking Claude-generated text and images with C2PA metadata to comply with EU transparency rules — a concrete, auditable action tied to a stated principle. You can argue about whether watermarks matter. You cannot argue that the gesture is empty when it requires actual engineering work to implement. OpenAI, by contrast, just signaled that its stated principles no longer need a dedicated human to defend them.

The timing is not incidental. OpenAI’s for-profit conversion means it now answers to a different set of incentives. Capped-profit structures are still profit structures. Investors expect returns. The fastest path to returns is faster model deployment, wider API access, more enterprise contracts. Ethics functions — when they work — are speed bumps on exactly that road. Removing the speed bump isn’t an accident. It’s a policy.

Amazon’s recent change to order confirmation emails is a small but structurally related data point. The emails no longer tell you what you bought — “Your Beauty item is confirmed!” instead of the actual product name — apparently to make it harder for AI agents to parse and relay order details to third parties, protecting Amazon’s data moat against AI-powered intermediaries. That’s a reported theory, not a confirmed explanation from Amazon. But if it’s right, it’s Amazon making a product decision to degrade the user-facing experience in order to protect its competitive position. It is the kind of call that benefits from someone whose job is to ask “should we?” before engineering asks “how do we?”

This is what the absence of an ethicist actually produces: not fewer bad decisions, but fewer decisions that get named as choices at all. The Amazon email change wasn’t announced. OpenAI’s ethics vacancy wasn’t announced — it was reported. Both organizations are making structural calls that affect millions of users, and both are doing it in the register of operational routine rather than deliberate policy. That’s the real risk. It’s not that OpenAI will now build something evil. It’s that the decisions that should be slow will be fast, and the consequences will arrive before anyone thought to ask the question.

Anthropic’s watermarking commitment won’t survive contact with AGI if the underlying incentive structures aren’t aligned. OpenAI’s ethics vacancy doesn’t guarantee catastrophe — plenty of organizations navigate complex ethical terrain without a dedicated function. But we are not in a normal industry at a normal moment. The models these companies ship are becoming load-bearing infrastructure for medicine, law, finance, and education. The stakes of a bad call are not a bad product — they are a bad institution, and bad institutions are much harder to fix than bad products.

The market will not price this. Ethics debt doesn’t show up in the revenue call until it’s already a liability.

Governance vacuums don’t stay empty — they fill with whatever’s fastest, and right now, the fastest thing in every AI lab is the capability team.