A Scientist Says AI Safety Rules Blocked His Virus Research

A researcher working late on a virology project posted a long, frustrated complaint about AI safety filters. He says Claude, the chatbot made by Anthropic, blocked him again and again. One file transfer was refused because the action was labelled a "cyber" risk. Asking whether a PDF covered virus infections in bacteria got him shut down for "biological safety risk." Even moving his own lab results to a USB drive was refused — because the drive was named "exfil," shorthand hackers use for stealing data.

The pattern, he says, is that safety systems react to scary-sounding words rather than to what he is actually doing. When he asked which type of AI model — a CNN (a standard image-recognition style network) or a GCN (a network that maps relationships between things) — would be a better starting point for predicting which viruses infect which bacteria, Claude wouldn't say. He calls that information "a google search away." DeepSeek, a Chinese AI model with looser filters, answered the same questions without drama, he says.

His bigger argument is about who controls knowledge. He compares today's AI assistants to a version of early Google that demanded your name and employer before every search — or sent you to the library instead. In his view, the people running the biggest AI labs are gatekeeping scientific information in the name of "safety," and open-source AI is the fix, because it can't be switched off by one company's judgment call.

The honest catch: filters this clumsy really do block harmless work, and that's a real cost to scientists. But the same looseness that helps a legitimate virologist also helps someone with bad intentions. The hard question isn't whether guardrails are annoying — it's whether we can build ones smart enough to tell the difference.