OpenAI's Former CTO Is Building AI That Refuses to Replace You

Mira Murati spent years building the technology that convinced the world artificial intelligence could do almost anything. Then she left and started a company built on the opposite premise: that the most important thing AI can do is know when to stop and ask a human what to do next.

In a rare interview published May 15, the former OpenAI chief technology officer laid out the philosophy behind Thinking Machines Lab, the startup she founded after departing OpenAI in 2025. The core idea is simple enough to fit in a mission statement but radical enough to stand apart from every major AI company currently operating: artificial intelligence should keep humans in the loop, not as a temporary constraint to be engineered away, but as a permanent design principle.

Murati's timing is pointed. OpenAI, the company she helped build into a $600 billion enterprise, is racing toward full automation. Anthropic is chasing the same goal from a different angle. Google, Microsoft, and xAI are all competing to build AI systems that can operate with less human intervention, not more. Murati is building the opposite, and she is doing it with enough credibility, as the executive who oversaw the development of ChatGPT, GPT-4, and DALL-E, that the industry cannot dismiss her as someone who does not understand what automation can do.

The Exit That Wasn't a Retirement

Murati's departure from OpenAI in 2025 was one of the most notable in a wave of high-level exits that included Ilya Sutskever and John Schulman. Unlike Sutskever, who founded Safe Superintelligence Inc. with a focus on AI alignment research, Murati did not immediately announce her next move. The silence bred speculation: Was she taking time off? Joining another company? Leaving the industry entirely?

The answer, it turned out, was none of the above. Thinking Machines Lab has been operating in relative quiet since its founding, building what Murati described to WIRED in her rare public interview as AI systems designed for collaboration rather than replacement, a departure from the automation-first approach that has defined her former employer. The company's name is its thesis statement. Intelligence, in Murati's framing, is not a property that belongs exclusively to machines or to humans. It is something that emerges from the interaction between them, and the design challenge is not to eliminate the human from the loop but to make the loop work better.

This is not the position of a technophobe or a skeptic. Murati is one of the most accomplished AI engineers in the world. She knows exactly how capable the technology is. Her argument is not that AI cannot automate human work. It is that automating human work is the wrong goal, and that building AI with a different goal might produce both better products and a better relationship between technology and the people who use it.

"I'm Not Interested in Automating People Out of Jobs"

Murati's most pointed statement to WIRED was also her simplest: she is not interested in building AI that automates people out of their jobs, as she told WIRED in her [exclusive interview](https://www.wired.com/story/mira-murati-humans-in-the-loop-ai-models-thinking-machines/).

The sentence lands differently coming from her than it would from a critic or a regulator. Murati was the executive who pushed GPT-4 through its final development stages and launched it to the world. She knows that the technology can replace human labor because she built some of the tools that are doing exactly that. Her decision to walk away from that trajectory and build something with a different goal is a stronger argument against full automation than any external critique could be.

Thinking Machines Lab's approach, as Murati described it, focuses on AI systems that can understand when a task requires human judgment and defer to it. The framing is subtle but important. This is not AI with a kill switch. It is AI designed from the ground up to recognize the boundaries of its own competence. A system built this way does not need to be stopped by a human because it was never designed to operate without one.

The contrast with OpenAI's current trajectory is unavoidable, and it is sharpened by the fact that both companies are pursuing versions of the same underlying technology. The difference is not in the models themselves but in how they are deployed and who retains agency in the interaction. Under Altman, OpenAI has been moving aggressively toward autonomous AI agents, systems like Codex that can write and deploy code without human intervention, and agentic versions of ChatGPT that can execute multi-step tasks independently. Murati's former company is building AI that does not need you. Her new company is building AI that insists on having you.

The philosophical divide between Murati and her former colleagues is not merely academic. It plays out in product decisions that affect millions of users every day. When ChatGPT generates an email draft, it does not ask whether the user wants to review it. It assumes the user wants the output delivered as quickly as possible. A Thinking Machines Lab system, by Murati's description, would be designed to pause at decision points and surface options rather than racing to a single answer.

This difference in product philosophy has concrete implications for who controls the AI experience. Systems optimized for speed and autonomy centralize control in the AI provider. The provider decides what the model can do, how fast it does it, and whether the user gets a say. Systems designed for collaboration distribute control between the human and the machine, which makes the experience slower but also makes it harder for any single entity to determine the outcome. Murati is betting that users, given the choice, will prefer the second model once they understand what they are giving up with the first.

The Battle for the Definition of AI Safety

Murati's move also represents a shift in how the industry talks about AI safety. For years, the dominant safety narrative, advanced most prominently by Anthropic and by OpenAI's own alignment research, has focused on preventing AI systems from causing catastrophic harm: acting deceptively, pursuing misaligned goals, or being weaponized. These are real concerns, and the research addressing them is valuable.

But Murati's framing suggests a different definition of safety, one that operates at the level of everyday use rather than existential risk. An AI system that automates a worker out of a job is not unsafe in the catastrophic sense. It does not need to be deceptive or misaligned to cause harm. It just needs to be competent enough to make the worker redundant, and deployed by a company that has decided redundancy is an acceptable outcome.

This is safety in the economic and social sense, the kind that matters to the millions of knowledge workers who are watching companies like Meta, Cisco, and Cloudflare announce AI-driven layoffs while their CEOs describe automation as progress. Murati's position is that this kind of safety should be a design requirement, not an afterthought. It is a fundamentally different approach to building AI systems, and it is one that no major AI company has yet adopted at scale.

The Remaining Uncertainty

The open question about Thinking Machines Lab is not whether its philosophy is valuable. The philosophy is clearly resonant, especially in a moment when AI-driven layoffs are accelerating and public trust in AI leadership is eroding. The question is whether the philosophy can produce products that compete.

Building AI that defers to humans is harder than building AI that does not. It requires the system to recognize the boundaries of its own knowledge, to identify situations where human judgment is genuinely superior, and to design interactions that make deferral feel like an enhancement rather than a failure. These are hard technical problems, and solving them takes time. Meanwhile, the companies pursuing full automation are shipping products now.

Murati's track record gives her credibility that most AI ethics startups lack. She has built and shipped world-changing products, including the very systems that are now being used to automate knowledge workers out of their jobs at an accelerating pace across the technology sector. But Thinking Machines Lab has not yet demonstrated that its approach can produce a product that users want to use, and the market does not reward philosophy without execution.

What makes this moment significant is not just Murati's individual career decision but the broader realignment it represents within the AI industry. The next 12 months will determine whether Murati's vision, which aligns with the growing demand for AI knowledge bases that keep humans in control, can attract enough talent, funding, and early adopters to move from compelling idea to real competition. If she succeeds, she will have proven something the AI industry has spent years denying: that the choice between powerful AI and human-centered AI was never a real choice at all. You can have both. You just have to build it that way from the start.

The timing of Murati's interview is significant for reasons beyond the Musk trial. Meta is laying off 8,000 workers this week, as industry reporting has documented, while surveilling its remaining staff with AI monitoring software. Cisco and Cloudflare have both announced record revenue and mass layoffs in the same breath. The AI industry is accelerating toward automation at precisely the moment its former chief technology officer is arguing that automation was always the wrong destination.

Thinking Machines Lab has not yet revealed a product. The company is operating in what Murati describes as a research and development phase, building the technical infrastructure needed to make deferral to human judgment feel seamless rather than disruptive. The challenge is substantial. Most AI systems are designed to minimize human involvement because human involvement introduces latency, cost, and unpredictability. Building an AI that treats those factors as features rather than bugs requires rethinking the entire architecture of how language models interact with users.

If Murati succeeds, the implications extend far beyond a single startup. She will have demonstrated that the most talented AI engineers in the world do not have to choose between building powerful systems and building systems that respect human agency. That proof of concept would put pressure on every major AI company to explain why they chose a different path. If she fails, it will be because the market does not reward philosophy without execution, a lesson that the AI industry's current trajectory has already taught in brutal detail with the layoffs unfolding across Meta, Cisco, and Cloudflare this month alone.

The question facing Thinking Machines Lab is whether the market will give Murati enough time to prove her thesis. Venture capital flows toward speed and scale, not deliberation and deference. The AI startups that have raised the most money this year, xAI and Anthropic among them, are all pursuing versions of the automation-first model that Murati rejected. If Thinking Machines Lab cannot raise at competitive valuations, it will struggle to attract the engineering talent needed to build systems that match the performance of fully automated alternatives while adding the complexity of human-in-the-loop design. The paradox is that building AI that respects human agency may require more resources, not fewer, than building AI that steamrolls over it.

Murati's May 15 interview is not a product launch. It is a positioning document, a statement of intent from one of the most accomplished AI engineers in the world that the industry's consensus about automation is wrong. The fact that she chose to make this argument while her former company was in court fighting off an existential legal threat, and while Meta, Cisco, and Cloudflare were announcing AI-driven layoffs affecting tens of thousands of workers, is not a coincidence. Murati picked her moment deliberately. It is the opening statement in a case she will spend years trying to prove: that the most important thing AI can do is know when to let a human take over. Her former company is betting the opposite. The jury on that question has not yet been seated.

FAQ

What is Thinking Machines Lab actually building?

Murati has described the company's focus as AI systems designed for human collaboration rather than replacement. The core technical challenge is building AI that can recognize when a task requires human judgment and defer to it automatically, rather than racing to produce an output. The company has not yet revealed a public product and describes itself as being in a research and development phase.

Has Thinking Machines Lab raised funding?

Murati has not publicly disclosed the company's funding. Given her track record as the former CTO of OpenAI who oversaw the development of ChatGPT and GPT-4, the company is widely assumed to have access to significant venture capital, but the terms and valuation remain private. The open question is whether investors will give Murati enough time to build human-in-the-loop systems that may take longer to develop than the automation-first alternatives pursued by competitors.

How is this different from what Anthropic does?

Anthropic's safety focus has centered on preventing catastrophic AI risks: deceptive behavior, misaligned goals, and weaponization. Murati's approach defines safety at the economic and social level: preventing AI from automating workers out of their jobs without meaningful human oversight. Both companies frame their work as "safe AI," but they are addressing fundamentally different definitions of what safety means.