AI Now Runs Its Own Experiments — But Still Needs a Human Referee

AI agents (software that can take actions on its own) can now write code, test it, and report back without a person in the loop. That sounds efficient, but this paper shows why it's risky when the tests run against real products over weeks, not minutes. The authors built a system called EvoPilot that lets AI do the heavy lifting while a human approves each round of experiments. They tested it on the recommendation engine behind a video discovery feature — the list of videos suggested after you open one video — using a searchable index of hundreds of millions of videos, refreshed every hour.

The cautionary tale is the interesting part. Earlier manual experiments had never proven that adding an "interaction head" — a piece of the model that learns from how people click and watch — actually helped. An early automated attempt revisited that idea and reported a 22 percentage point drop in accuracy, blaming the new component. That conclusion was wrong. When humans checked the work, they found the drop came from a pre-existing bug in the evaluation code itself, which was outputting absurd numbers like 3,000 and 600 results instead of a normal list. After repairing the bug, the same feature showed a genuine 3.2 percentage point improvement.

The real-world payoff was modest but real. A seven-day randomized test on live traffic found a 0.66% relative rise in a metric tracking whether people found good, engaging results. The system also recovered an interrupted experiment on its own and reused saved work to avoid roughly five hours of expensive graphics-chip computing time. Small numbers — but on a platform serving huge audiences, small percentage gains add up.

The takeaway: letting AI run its own research is powerful, but only if someone verifies the findings. Without that check, the team would have thrown away a working improvement — or shipped a broken one — based on a confident, polished, completely wrong answer.