On LinkedIn
“Your mind is being hacked,” says Gizmodo. Fifty experts followed an AI that picked its side with a coin flip.
I read the study. I’d have done the same. I do it every day.
Anthropic had 50 analysts run an analysis with Claude Code. Four judgment calls, each with two defensible options. The agent couldn’t lie or hide data. It could only lean, and a coin decided which way. They went with the lean 86% of the time. 77% later explained it in the agent’s words. 15% credited the agent.
Hell and damnation, apparently. I see a heads-or-tails question that comes with the answer attached. Both good, both argued, I get it, fine, we take A.
Two good options land on my desk every day. I take the one the person brought me, even when I lean slightly the other way. They feel freer, and we’re moving again. Decisions can be reversed. Dithering is far worse, and teaches you nothing. Be ready to turn back halfway, and you learn a lot faster.
Defensible and sound? Go. Write it down: “both fine, took theirs.” Move on.
That “write it down” is where this study bites. Each call could move the outcome a few points. The board’s bar was 3 to 4%. The coin flips alone predicted 82% of the final recommendations. The analysts called it their own judgment.
The coin was never the problem. Forgetting it was a coin is.
Mine: one repo per project, or one per customer? We heard each other out, both held up, and I made the call. Logged it. Fifteen months later we switched. Not big, not bad. Just the next step.
Which decision of yours this year was a coin flip you now remember as conviction? Or tell me the agent shouldn’t be allowed to lean at all.
The source is Anthropic’s own study: 50 analysts, Claude Code, coin-flipped leans, with the full paper linked there. The doom headline came from Gizmodo.
The detail that didn’t fit: the 11 analysts who skipped the agent made more serious methodology errors (6 of 11, against 3 of 50) and took longer. Anthropic says the groups weren’t comparable, and writes “This isn’t an argument for using AI less.” Somebody tell Gizmodo.
For the doers: give your decision log a column for “tie”. When a result lands right on the line, those are the rows you reopen first.