OpenAI's Hidden Logs and Robots Doing Surgery

For two full years, OpenAI reportedly told a federal court it couldn't search its own ChatGPT logs. Turns out, it already had — on a sample of 78 million of them.

That's the bombshell at the center of a new sanctions motion filed by The New York Times and other news organizations suing OpenAI for copyright infringement. If the allegation holds up, it means OpenAI didn't just drag its feet on discovery. It allegedly looked the court in the eye and described a technical limitation that didn't exist.

The whole thing unraveled during a re-deposition of OpenAI privacy engineer Vincent Monaco in April. He was brought back to the stand after a previous appearance left the court unsatisfied, and his testimony apparently opened a can of worms that two years of careful legal positioning had kept sealed. According to the NYT's filing, Monaco revealed that OpenAI had already conducted searches across large anonymized samples of ChatGPT logs — including datasets of 10 million and 78 million entries — well before the lawsuit even began.

The news plaintiffs argue this matters enormously. Those logs are potentially the most consequential evidence in the entire case. They could show whether real users were regularly prompting ChatGPT to reproduce paywalled articles verbatim — exactly the kind of behavior that would support the infringement claims. Or they could show the opposite, potentially helping OpenAI's fair use defense. Either way, keeping them hidden for two years wasn't a minor procedural hiccup.

The sanctions request argues that OpenAI's concealment inflated legal costs, stretched out discovery, and wasted the court's time. That's the kind of conduct judges tend to take seriously, regardless of which side they sympathize with on the underlying claims.

OpenAI, for its part, framed the sanctions motion as a privacy grab dressed up in legal language. A spokesperson argued that the NYT's recent decision to drop certain claims was a sign its case was falling apart, and accused news plaintiffs of trying to dig through logs belonging to users who have nothing to do with the lawsuit. The company says it's defending both fair use principles and ordinary people's privacy.

The NYT pushed back hard on that characterization. A spokesperson said dropping those claims was a strategic move to streamline the case and add Microsoft as a defendant — not a concession. The core argument, they insist, hasn't changed: that OpenAI and Microsoft used millions of copyrighted works without permission to build a product that competes directly with the outlets that produced them.

What makes this moment particularly awkward for OpenAI is the timing. The company has spent considerable energy positioning itself as a trustworthy, transparency-minded actor in the AI space. Getting caught — by one of its own witnesses, no less — allegedly misrepresenting its technical capabilities to a federal court is not a great look for that narrative.

The sanctions motion is still heavily redacted, so the full picture isn't public yet. But what's visible is enough to suggest this case is about to get significantly messier for OpenAI before it gets any cleaner.

A humanoid robot just removed a gallbladder from a living animal — and the most interesting part isn't the robot. It's how cheap the whole setup was.

Researchers at UC San Diego recently published results in Nature showing that a Unitree G1 humanoid robot, remotely controlled by a surgeon, successfully completed minimally invasive gallbladder removal surgery on live pigs. Two surgeries, two pigs, zero autonomous decision-making from the machine itself. The robot was the hands. The surgeon was still very much the brain.

The Unitree G1 starts at around $13,500 in its baseline configuration, though a version equipped with the dexterous hands needed for something like surgery can run north of $67,000 once you factor in upgrades. That sounds like real money — until you compare it to an Intuitive Surgical da Vinci system, which can cost anywhere from half a million to several million dollars, weighs close to 1,800 pounds, and requires dedicated space that most smaller hospitals simply don't have.

The UC San Diego team had to do some engineering legwork to make this happen. The Unitree G1 wasn't designed for the operating room, so researchers built custom physical adapters to let it hold surgical instruments, and wrote software to translate natural hand movements from a surgeon at a control console into precise tool movements at the robot's wrists. The surgeon wore a stereo headset to see what the robot saw and used a foot pedal to connect or disconnect their hand movements from the tools — a practical safety feature that gives the operator a quick way to pause without dropping anything.

The vision here goes well beyond novelty. Surgical assistant professor Shanglei Liu described the potential deployment scenarios in pretty vivid terms: rural clinics, battlefield medicine, even space. Those aren't throwaway examples. They all share the same underlying problem — a patient who needs surgery and a specialist who is physically somewhere else. Teleoperated humanoid robots are one possible answer to that gap, and a relatively affordable one if the technology matures.

That's a big if, though. The da Vinci system has decades of clinical trial data behind it and regulatory clearance from the FDA and equivalents in other countries. The humanoid robot approach is still in preclinical territory, meaning the bar to human trials — let alone widespread deployment — remains high. A successful pig surgery is genuinely meaningful progress, but it's one data point, not a green light.

Still, the fact that this worked at all with an off-the-shelf commercial humanoid robot and custom-built software is notable. It suggests the barrier to entry for experimental surgical robotics is dropping fast. The specialized, purpose-built surgical robot has dominated this space for years largely because nothing else could realistically compete on precision or reliability. That assumption may need revisiting sooner than the industry expects.