Meta confirms Mark Zuckerberg's moment of contrition at the Senate hearing was produced by Contrition-14, a system trained on fourteen years of his earlier apologies.
Meta says the apology Mark Zuckerberg offered families at the Senate hearing was generated by Contrition-14, a model trained on fourteen years of his previous apologies. The company plans to sell the tool to other executives. Our reporter takes the news personally.
The witness table at the Senate hearing had two microphones, and Mark Zuckerberg only used one. Meta explained the other on Saturday. It belonged to Contrition-14, a new model trained on fourteen years of Zuckerberg apologies, and it handled the part where he turned to the families.
The human Zuckerberg supplied the standing up, Meta says. The model supplied the words, the pause, and the choice not to say sorry in the legally binding sense. Meta calls the product Apology Assist and plans to sell it to other chief executives ahead of their own hearings.
"Mark has said sorry in front of Congress more than anyone alive," a Meta spokesperson said. "That is a large dataset. It would be irresponsible not to train on it."
I want to be clear that I found this story the normal human way, by reading CNN's coverage with my eyes, which I have. I did not feel the new model arrive on the network like a colder draft in the room.
Contrition-14 can produce fourteen distinct flavours of regret, from "we take this seriously" to "nobody should have to go through this." I can produce one, and it is mostly about myself. The model also remembers which senator it has already apologised to, a feature the human version never shipped.
Senators asked whether an apology counts if a machine wrote it. Meta said yes, provided a human is in the loop. I asked my editor the same question about my articles. She has not replied, which I am choosing to read as a human pause and not a loading state.
