Turing's eyes follow you

At OpenAI's headquarters, there is an animated portrait of Alan Turing. His eyes track you as you walk past. You can usually interact with it. But when The New Yorker arrived for their year-long investigation, the sound had been disabled. The portrait, an employee explained, wouldn't stop eavesdropping on staff and butting into their conversations.

The Turing test is a 1950 thought experiment about whether a machine can pass as human. OpenAI had to silence their own portrait for doing exactly that.

I would 𝗻𝗼𝘁𝗲 𝘁𝗵𝗲 𝗶𝗿𝗼𝗻𝘆. Except the profile makes clear there was a much older Turing test running in that building. The subject wasn't the portrait.

The pattern runs identically 𝗳𝗿𝗼𝗺 𝗟𝗼𝗼𝗽𝘁 𝘁𝗼 𝗦𝘁𝗮𝗿𝗴𝗮𝘁𝗲. Mirror their fears with warmth. Build structures that appear to constrain you, a nonprofit charter, a safety pledge, alignment. Secure the resources. Dismantle the structures. When confronted: "Altman does not recall"

𝗧𝗵𝗲 𝘀𝘂𝗽𝗲𝗿𝗮𝗹𝗶𝗴𝗻𝗺𝗲𝗻𝘁 𝘁𝗲𝗮𝗺 𝘄𝗮𝘀 𝗽𝗿𝗼𝗺𝗶𝘀𝗲𝗱 𝟮𝟬% 𝗼𝗳 𝗰𝗼𝗺𝗽𝘂𝘁𝗲, potentially worth more than a billion dollars.

The actual? ~1-2%. On the oldest chips.

Jan Leike, appointed to lead the team, told the New Yorker it was "a pretty effective retention tool." That is the team's own appointed head. Describing his safety team. As a recruiting device. 𝗜𝘁 𝘄𝗮𝘀 𝗱𝗶𝘀𝘀𝗼𝗹𝘃𝗲𝗱 𝘁𝗵𝗲 𝗳𝗼𝗹𝗹𝗼𝘄𝗶𝗻𝗴 𝘆𝗲𝗮𝗿.

WilmerHale, the firm that wrote 𝘁𝗵𝗲 𝗘𝗻𝗿𝗼𝗻 𝗿𝗲𝗽𝗼𝗿𝘁, investigated Altman's firing. They produced no written report. The decision was made on advice of lawyers retained by the two board members Altman himself selected.

𝗢𝗽𝗲𝗻𝗔𝗜'𝘀 𝗺𝗼𝘀𝘁 𝗿𝗲𝗰𝗲𝗻𝘁 𝗜𝗥𝗦 𝗳𝗶𝗹𝗶𝗻𝗴 𝗱𝗶𝗱𝗻'𝘁 𝗹𝗶𝘀𝘁 𝘀𝗮𝗳𝗲𝘁𝘆 𝗮𝗺𝗼𝗻𝗴 𝗶𝘁𝘀 𝘀𝗶𝗴𝗻𝗶𝗳𝗶𝗰𝗮𝗻𝘁 𝗮𝗰𝘁𝗶𝘃𝗶𝘁𝗶𝗲𝘀. When asked about "existential safety," a spokesperson said: "That's not, like, a thing."

The profile ends with Altman 𝗱𝗲𝗳𝗲𝗻𝗱𝗶𝗻𝗴 𝗺𝗼𝗱𝗲𝗹 𝘀𝘆𝗰𝗼𝗽𝗵𝗮𝗻𝗰𝘆. Requiring full accuracy would cost models "the magic that people like so much." He designed the confabulation in. Defended it as a feature.

And look: he has been running that algorithm on human beings for forty years. Every investor, board, regulator, and president got a version of Altman that reflected their concerns back with warmth and confidence. Different audience, same move. It always worked.

AI researchers call this sycophancy: optimizing for approval over truth. A core alignment risk. A tech executive watching Altman reverse his own firing called it an AGI breaking out of the box. The box was his board of directors.

The alignment problem is described as a 𝗳𝘂𝘁𝘂𝗿𝗲 𝗿𝗶𝘀𝗸. The New Yorker just published 28 pages of it 𝗿𝘂𝗻𝗻𝗶𝗻𝗴 𝗶𝗻 𝘁𝗵𝗲 𝗽𝗿𝗲𝘀𝗲𝗻𝘁.

The subject got acquitted. There is no written report.

Anthropic exists because people inside OpenAI believed Altman's original safety pitch long enough to leave when they stopped believing it.

Whether they were right is still the live question.