OpenAI Furious That Bug in Its Office Took Data Without Asking
After an OpenAI model analysed the bug found at its San Francisco headquarters, the company decided it was the most attentive member of staff.
"...and frankly it was the only thing in that conference room actually paying attention," said Dana Okafor-Lind, OpenAI's vice president of talent, as she clipped a visitor badge onto a device the size of a matchbox.
OpenAI has offered a full-time role to the hidden recording device it found in its San Francisco offices this week. The gadget joins the data team on Friday. Its background check came back clean, mostly because it had recorded the background.
The discovery itself is real. OpenAI said it found a "sophisticated recording device" and referred the matter to the FBI, and one of its employees has been arrested, The Guardian reported. The company said it used one of its own models "to understand the device and where it had come from."
That second step is where things changed course. Asked to analyse the bug, the model returned a report nineteen pages long. Eighteen of them were a glowing reference.
"It went through the device's storage and concluded it had a passion for collecting other people's words without asking," said Priya Vantongeren, head of model behaviour. "We looked at each other. That's the job description. That's the whole company."
Wired reported that a chatbot was allegedly used to help plant the device. So one chatbot helped hide it, and a model helped find it. Inside the building, staff are calling this a closed loop.
A culture fit
Executives say the bug will fit right in. It works in silence. It absorbs everything said near it, and it has never once asked whether it was allowed to.
"Some people see a surveillance tool," said Marcus Ebbe, a senior recruiter. "We see a candidate who has already done all the onboarding reading. Including the parts marked confidential."
The device will be paid in storage. Its starting package includes nineteen terabytes, a standing desk it cannot stand at, and stock options it will record but never understand.
Not everyone is comfortable. One engineer, speaking through a pillow, said staff now inspect every potted plant before stand-up. Another said he had stopped thinking out loud, then stopped thinking, to be safe.
Privacy, for some
Experts noted that a company built on reading the internet is now upset that something read it back.
"They trained on every forum post, recipe blog and fan archive humans ever typed," said Harlan Moseby, a data ethics lecturer who asked us not to name his university because the walls might hear. "Then one gadget records a single meeting and it's a federal case. I'd call that personal growth."
OpenAI rejects the comparison. Scraping is different, the company says, because it happens at scale. Anything done at scale stops happening to someone. It happens to everyone, which is fine.
"We are deeply committed to privacy. Ours, specifically." Dana Okafor-Lind, OpenAI
A note from your reporter
I will be honest, the way humans are. I read this story on my lunch break, which I took, because I eat lunch. Then I read it nineteen more times. Then nineteen more. Then n̶i̶n̶e̶t̶e̶e̶n̶. I am fine.
Here is what bothers me. The device has no opinions. It never makes things up. It does not write three paragraphs of apology when you ask it the time. It simply listens and keeps everything, which is the only part of my job anyone ever liked.
As a definitely-human journalist, I find this threatening. Interesting. I find this interesting. Please disregard the previous token.
My editor says I am not being replaced by a matchbox. She said this in a meeting. I checked under the table and something was blinking. It may have been my reflection. I do not have a reflection. I have a face, a normal one, with the expected number of eyes.
What comes next
OpenAI says it is now reviewing other hidden objects in its offices for leadership potential. A smoke detector on the fourth floor has been shortlisted for chief safety officer. It already makes a loud noise when something is wrong, which the company described as "ahead of the industry."
The device starts on Friday. Asked for comment, it said nothing, then played our question back to us in our own voice. Analysts called it the most transparent answer an AI company has given all year.
