Two stories dropped this week that have nothing to do with each other and everything to do with the same problem. Inherent's Faraday AI agent, built by DeepMind alumni, is now outperforming GPT-4 and Claude at replicating scientific research papers. Meanwhile, a Nature investigation revealed that a single scientific sleuth identified dozens of studies built on the wrong antibody, meaning their foundational reagents were misidentified from the start. Put these together and you get a vertigo-inducing loop: an AI that can reproduce papers with high fidelity, in a landscape where many of those papers were themselves already producing false results.
The Authenticity Crisis Runs Deeper Than AI
The antibody error story is not about AI at all. It's about how institutional trust in peer review allows errors to compound invisibly across decades. A 2015 paper in PLOS ONE by Bradbury and Bhatt estimated that hundreds of thousands of published experiments used misidentified antibodies, a contamination of the scientific record so vast it resists easy remediation. Now introduce an AI optimized to replicate that record. Faraday does not verify; it reproduces. If the source is corrupt, the replication is a high-fidelity copy of a corrupted original.
Forgery Logic Across Fields
The Norval Morrisseau forgery case being cracked open this week by an obscure Canadian agency offers an uncomfortable parallel. Fakes had been circulating since at least 1996, authenticated by institutions that trusted provenance chains rather than primary evidence. The same institutional deference that allowed Morrisseau fakes to proliferate for thirty years is what allowed wrong-antibody studies to accumulate citations. In both cases, a single dedicated investigator, not a system, identified the problem. Rhizome's Michael Connor has argued that archiving digital culture requires precisely this kind of obsessive individual attention to provenance. The lesson scales: verification is a human craft, not a process you can replicate your way out of.