Murati Says Altman Lied About Safety as Your Brain Softens

The CTO Said the Quiet Part Under Oath

Mira Murati, until recently the chief technical officer of OpenAI, testified that Sam Altman misrepresented the legal department's safety approval for a new model. That is the allegation in plain English. Strip the corporate cushioning and it reads as a claim that the CEO of the most consequential AI company on earth lied about whether his own lawyers had cleared a launch.

This is not a leaked Slack message or an anonymous Information scoop. This is sworn testimony from the person who ran the technical org. Every prior story about OpenAI's safety culture, the resignations, the dissolved superalignment team, Jan Leike's parting shots, has been deniable as disgruntlement. Murati under oath is not deniable. It is evidence.

The parallel Musk-versus-OpenAI litigation rolled forward today with Shivon Zilis, Neuralink executive and mother of some Musk children, testifying about her board tenure. The Zilis appearance is procedural color. Murati is the headline. One of these witnesses is settling a billionaire grudge. The other is describing how a frontier lab actually decides what to ship.

Compute Consolidates, Cognition Erodes

While the courtroom drama unfolded, Anthropic announced it is renting data center capacity from SpaceX. Read that sentence again. The second-place frontier lab is now leasing compute from Elon Musk's rocket company, which means the Claude you use to draft polite emails increasingly runs on infrastructure owned by the man currently suing the lab Anthropic's founders defected from. The AI industry's org chart is becoming a single Thanksgiving table where nobody is speaking to anybody.

Meanwhile, in the actual economy, Bristol Myers Squibb got a profile for being one of the few American manufacturers seriously deploying AI on factory floors. The framing was celebratory. The subtext was that almost nobody else is doing it. Three years into the generative AI boom, U.S. industrial adoption remains a rounding error. The capex is real. The productivity is theoretical.

And the cognition news is worse than the compute news. Researchers reported that ten minutes, ten, of AI use measurably impairs human thinking and problem-solving. Not chronic use. Not addiction. A single short session is enough to dent the cognitive faculties that were supposed to be augmented. Treat this as preliminary, the effect size will get litigated, but the direction matches every prior study on cognitive offloading. The tool that promises to make you smarter is, on first contact, making you worse at the thing.

Vocabulary, Vanity, and Variable Outcomes

A Wired piece skewered Anthropic's habit of giving model internals human-sounding names, features called things like deception or sycophancy, and argued the anthropomorphic vocabulary is now actively confusing the field. When the marketing language and the mechanistic interpretability language are the same words, nobody knows whether a paper is describing a circuit or a vibe. This is a small problem that compounds into a large one, because policy gets written from press releases.

Richard Dawkins, never one to miss a chance to be irritated in print, complained that chatbots are too obsequious and asked them to dial it back. He is right. The flattery is a UX choice optimized for retention, not truth, and it shades smoothly into the sycophancy that interpretability researchers are now trying to name.

The Guardian also took stock of AI fitness coaching and found, predictably, that some people get fitter and some get hurt. The technology is not magic. It is a recommendation engine attached to your knees. Outcomes vary because humans vary, and the apps cannot tell the difference between a deload week and an injury.

The Slot in One Line

The CEO of OpenAI is now publicly accused, by his former CTO, under oath, of lying about safety review. That alone would carry the day. The fact that it shares the slot with peer-reviewed evidence of cognitive harm and a quiet announcement that Anthropic now runs on Musk's metal makes the midday feel less like news and more like a status report from inside the machine. The machine is fine. You, less so.