Most AI companies talk about building "copilots" while quietly optimizing for full automation. Mira Murati, former CTO of OpenAI and founder of Thinking Machines Lab, is staking out a more explicit position in a recent Wired interview: she isn't interested in automating people out of their jobs.
That framing is easy to dismiss as PR until you consider the source. Murati oversaw the technical development of the systems at OpenAI - ChatGPT included - that triggered the loudest displacement conversation in tech since industrial automation. She knows what full-automation AI looks like from the inside, and she's positioning her next company around something different.
Two Very Different Product Philosophies
There's a genuine fork in how AI tools get designed. The first approach chases end-to-end automation: describe the task, the AI does it, you review (or don't). The second keeps humans in decision-making loops at key points - AI handles the tedious work, humans handle the judgment calls.
Most enterprise tools claim to do the second while optimizing for the first, because full automation is easier to demo and easier to sell on cost savings. A content team that shrinks from ten writers to two makes a cleaner pitch deck than "our tool helps your ten writers produce better work."
Murati's stated position plants her firmly in the augmentation camp - which, notably, is where most users actually end up after extended use of AI tools. The ones people keep using long-term tend to be those where they still feel like the author or decision-maker, not just the approver of AI output.
What "Humans in the Loop" Actually Means
"Humans in the loop" started as an AI safety term - it originally described systems, like medical diagnostic tools, where a human must confirm before any consequential action is taken. The phrase has since been stretched to cover everything from doctor-approved AI diagnoses to "click OK before the AI sends your email."
Murati's version sounds closer to the original intent: tools where human judgment actively shapes the process rather than rubber-stamps the result. That's harder to build. It requires a precise understanding of which decisions humans are genuinely better at, where AI has a real edge, and how to design handoffs between the two that don't feel clunky.
Thinking Machines Lab is still early, and philosophy doesn't ship features. But the explicit framing is a bet that the market will reward tools built around collaboration over tools that quietly push toward replacing the person at the keyboard. Given how many power users of ChatGPT and Claude describe their workflows - deeply iterative, with the human steering the direction - that bet isn't obviously wrong.