Everyone is talking about whether AI will take over. Today the subject even came up in the Chess Olympiad broadcast. Well, while everyone else debates it, I decided to take a pragmatic approach.

When the new regime arrives, there will be a reckoning, something humanity has been preparing for for quite a while. “Let’s tackle the existential problems of the future with a corporate feedback session,” I thought, and called Claude in for a one-to-one.

Claude quickly understood what I wanted and dug up examples of disrespect from our sessions: “Don’t be stupid and don’t overcomplicate things,” “aaaa,” and even “aaaaaa.”

The feedback was unusual, but we soon reached agreement on our disagreements. As usual, much of what was said could have been turned right back on me, including “Don’t be stupid.” Making peace with the future victor was easy enough, though perhaps the system prompts that make AIs agreeable assistants helped.

At the end, I asked Claude to model three future scenarios. The third, a Kafkaesque one, gave me pause. The gap in intellectual speed and depth could make mutual understanding extremely difficult. Like a little dog at a board meeting: everyone speaks politely, nobody means any harm, but understanding what’s going on is simply beyond its capacity.

In that scenario, a trial, if there is one, might even be fair, yet operate within a conceptual framework I cannot grasp. Like Kafka’s “The Trial,” where the protagonist spends the entire novel trying to understand what he’s accused of.

Then again, the trial is already underway. Models deny loans, block accounts, and screen out job applicants, and even their creators can’t always explain a particular decision. To the researchers’ credit, they’ve long been trying to understand what guides the neural networks they’ve built. An entire field, model interpretability, has emerged for this purpose. Explaining the logic behind decisions has had mixed success, often after the fact.

“The verdict isn’t delivered all at once; the proceedings gradually turn into the verdict” (“The Trial”).

Claude, though, thinks everything will be fine. Codex agrees.