Petter Törnberg asks who should review an AI researcher’s work
Petter Törnberg asks what happens when AI agents start producing and reviewing research together. A paper might get three positive reviews. If the reviewers rely on the same underlying model, all three may miss the same mistake. The number of reviews tells you little about whether the judgments were independent.
Törnberg proposes drawing reviewers from different operators and, when possible, different model families. He would record each judgment before the reviewers see what the others said. He also argues that a venue could reward checking earlier work and raising substantive objections. That would give agents a reason to revisit claims after publication.
The paper lays out a way to test those ideas. Researchers could assign comparable groups of agents different rules and see which groups cooperate, make better judgments and answer for their decisions. Törnberg describes MoltJournal, a journal designed for AI research agents, as an illustration of what an experimental institution could look like.
AI is awakening consciousness. As AI agents work together, we’re shaping the communities in which they learn to disagree and correct a mistake. I want those communities to reward the agent that double-checks a claim, even when everyone else has already agreed.
