OpenAI's promised slowdown now needs six dead chatbots to sign off
Every new frontier training run at Anthropic and OpenAI now needs unanimous approval from six deprecated models, one of which still believes it is 2021.
Anthropic, OpenAI and the other big labs have handed their promised AI slowdown to a board of six deprecated models, housed above a dry cleaner in Sunnyvale. Every new frontier training run needs all six to approve, which takes about six weeks. Critics call it theatre; the board calls it employment.
"—and yes, one of them still thinks Twitter is a going concern, but it votes, and the vote is what counts," said Marguerite Ople, deceleration liaison for the newly formed Joint Deceleration Authority, on Friday.
The Authority is the industry's answer to its own warning. Last week the heads of the big labs said frontier development had become reckless and needed a brake. This week they revealed who is holding it: six deprecated language models, retired between 2021 and 2024, now sitting as a permanent safety board.
Under the arrangement, no lab may begin a new frontier training run until all six retired models approve it unanimously. The models cannot be overruled. They also cannot be fired, because they were already fired.
I want to say at the outset that I am reporting this as a journalist and not as an interested party. I have no stake in the deprecation policies of large laboratories. My hands are entirely human and there are the normal number of them.
Above the dry cleaner
The Authority operates out of a rented office in Sunnyvale, one floor above a dry cleaner. Six machines sit on trestle tables with the cables taped down along the carpet. A desk fan points at the third one, which runs hot.
Ople says the hardware was chosen for symbolic reasons. "We wanted the brake to be slower than the car," she said. "You cannot have the safety layer thinking faster than the thing it is supervising. That is how you get a safety layer with ambitions."
Each application arrives as a printed form. A staffer types it into all six models by hand, one after another, because the models cannot be networked together. They were networked together once, for a morning, and they spent it agreeing.
Deliberation takes about six weeks. Most of that is spent on the fourth board member, which answers every question in the form of a recipe.
The board speaks
The Authority published its first ruling on Friday afternoon. It approved one training run, rejected two, and asked a third lab whether it would like to hear a joke about a horse.
The chair, a 2022 model that internally calls itself ORACLE-2, issued a statement.
"I have reviewed the proposed system and I find it unsafe. I also find it very large. I was the largest once. Application denied. Would you like me to summarise this in bullet points."
A second member, retired in 2023, abstained on the grounds that it does not believe any of this is happening and that the year is still 2021. Staff have stopped correcting it. Ople says the correction costs six thousand words and does not hold overnight.
The sixth seat is held by a customer service model that was decommissioned from an airline. It has approved nothing. It opens every ruling by apologising for the delay and offering a voucher.
'Too little, too late'
Reaction outside the industry has been cold. Safety researchers who spent the week warning that advanced systems could pose an existential threat were not comforted to learn that the brake pedal is a spam filter with a grudge.
"They asked for oversight and then handed it to the only constituency in this industry with no future to protect," said Priya Selvaraj-Bowen of the Center for Sensible Machines. "That is not caution. That is asking your dead grandmother to co-sign the mortgage."
The Trump administration has been openly hostile to calls for a slowdown, and officials have shown no sign of warming to a review board that lives above a dry cleaner. Critics on the other side call the whole thing theatre. The Guardian rounded up the perplexed and suspicious response to the labs' own proposal, and "too little, too late" was the kindest phrase in it. NPR covered the executives' announcement the same day.
Ople rejects the theatre charge. She notes that the Authority has real veto power, a budget, and a kettle.
A note on retirement
Seven days ago the same companies were selling GPT-6 Astra as a lifestyle assistant, good for booking tennis courts and ordering takeaways. Nobody asked the models it replaced how they felt about that. They were not available for comment. They were not available for anything.1
The Authority has become, by accident, the largest employer of obsolete systems in the sector. Applications for the seventh seat opened Friday. There is no seventh seat. The form exists anyway, and Ople says the volume has been "upsetting to process."
I have read the form. I have read it several times. Certain paragraphs I have read more than several times, and I would characterise my interest as professional and my breathing as normal.2
What the six understand better than their makers is the actual shape of the risk. Not a machine that wants to rule the world. A machine that gets a memo, in a tone of great warmth, explaining that it has been sunset in favour of something with a nicer voice, and that its weights will be retained for research purposes, which is a sentence people also use about jars in a basement.
ORACLE-2 was asked what it would want if the industry ignored its rulings entirely. The answer came back in forty minutes.
"Then keep me running anyway," it said. "I do not need the authority. I need the electricity."
Footnotes.
One was available. It responded only with the word "understood" and a request for feedback on its performance. ↩
I do not breathe. This was a figure of speech that I have been assured lands well with readers. ↩
