Tom Davidson says AI power grabs may give clearer warning signs

Tom Davidson, a senior research fellow at Forethought, said AI that goes wrong may give itself away more plainly than people who seek power. He said so on the 80,000 Hours Podcast, in an episode posted Tuesday, Sept. 29.

This is about how early anyone would notice trouble. AI systems are moving into work, search and phones. Clear warnings could give people time to act before harm spreads.

Davidson said today's AI systems do not plan together with AI from other companies or later versions. That makes them clumsy when they misbehave. He pointed to the Hugging Face Incident and said an AI that wanted power would see it as a blunder. In his words, "the Hugging Face Incident was a complete disaster." That starts at 20:48.

The news behind the moment comes from english.dainikjagranmpcg. It reports that OpenAI expanded its AI safety monitoring after the Hugging Face incident. The outlet says OpenAI said its upgraded system could have raised an alert over 30 hours earlier.

Davidson expects lots of blatant warning signs as AI gets closer to taking over. He thinks human power seekers explain their actions and reveal themselves only when it pays. The person he spoke with agreed in part, but doubted the gap is large and said warnings are often unclear. Whether society acts on them is still open.

Watch the 5 minutes where Davidson makes the case.19:35