A Quiet Departure, A Jarring Signal
Aleksander Madry is gone.
He was no ordinary OpenAI employee. He was the head of "preparedness"—in industry parlance, the person whose job is to hit the brakes. His role was to assess the potential risks of AI systems and determine whether a model was safe enough for release.
Last summer, he was reassigned from his safety position to work on AI reasoning research. Even then, people found it puzzling: why would someone focused on safety be tasked with researching model capabilities?
Now, he has completely left OpenAI. In his post, he stated he will be researching "the impact of AI on the economy."
Polite and professional, but the message between the lines is clear: I can no longer do what I want to do here.
OpenAI's "Safety Narrative" Is Crumbling
There was a time when the name OpenAI itself carried a sense of mission. "Open"—openness. "AI"—artificial intelligence that benefits all of humanity. Sam Altman gave speeches everywhere, talking about AGI arriving safely and benefiting everyone.
But reality tells a different story:
- From non-profit to for-profit. OpenAI long ago transitioned from a non-profit to a "capped-profit" company, with mounting pressure from investors for returns.
- A shift to closed-source. Once committed to open-sourcing its models and research, OpenAI now keeps its most advanced models closed-source.
- Marginalization of the safety team. Madry's reassignment was not an isolated incident. Reports indicate that the influence of OpenAI's internal safety team has been steadily declining, while the product team's influence has been rising.
This isn't to say OpenAI is doing evil. It's to say: When an organization faces growth pressures, "safety" is often the first thing to be compromised. It's not because leaders turn into villains, but because business logic naturally favors speed over caution.
Reassigned, Then Departing—This Path Is Nothing New
Madry's experience is not uncommon in the tech industry.
When a company's strategic focus shifts from "doing the right thing" to "doing what makes money," the status of departments like safety, compliance, and ethics inevitably changes. They won't be shut down—that would look terrible. Instead, they get "restructured," "adjusted," or "repositioned." Safety experts are reassigned to product research, while ethicists are tasked with writing "corporate social responsibility reports."
Then, those who genuinely care about safety leave on their own. Not because they were fired, but because they realize they can no longer influence decision-making.
Madry's choice to study "the impact of AI on the economy" speaks volumes. What he couldn't prevent from his safety role, he might still contribute to from a different angle—researching AI's impact on employment, income distribution, and economic structures. The answers to these questions may ultimately be more important than any single model's safety evaluation.
The Industry Needs an Independent Safety Force
OpenAI's story tells us one thing: We cannot expect AI companies to act as their own safety referees.
This isn't a dig at OpenAI specifically—no company can pull this off. When you are both the athlete and the referee, the outcome of the game is already predetermined.
What the industry truly needs are independent safety assessment bodies, similar to auditing firms in finance, the FDA in pharmaceuticals, or airworthiness certification in aviation. These institutions are not paid by the entities they regulate, nor are their leaders appointed by them. Their sole mission is the public interest.
Currently, the AI industry lacks such a framework. There are academic papers, non-profit reports, and draft guidelines from government agencies. But there is no independent safety certification system with actual binding authority.
Madry's departure is a loss for OpenAI, but it could also serve as a catalyst for pushing forward the establishment of independent safety mechanisms.
A Word to the Industry
When safety leads are reassigned to product research, when ethics teams are folded into PR departments, and when "Responsible AI" shifts from a core mission to a marketing slogan—this is not just one company's choice. It is a crossroads for the entire industry.
We can continue to pretend that growth and safety can be perfectly aligned. Or we can admit: Under the current incentive structure, safety and growth are inherently in conflict. Acknowledging this isn't pessimism; it's honesty. And honesty is the first step toward solving the problem.
Madry chose honesty—he left. The remaining question is: how many others will choose honesty?