Originally posted at LeverNews.com
“Misalignment.” In the wake of artificial intelligence hacks, break-ins, and mini-disasters, that’s what technologists call the terrifying gap between what we want AI to do and what AI is actually doing. But much of what we are seeing is better described as AI’s accelerated super-alignment with the mundane behaviors of its Silicon Valley creators.
Take the infamous hack of Hugging Face, in which OpenAI’s agents broke out of their sandbox and hacked another company, unbeknownst to their programmers. In that debacle, the most frightening behavior was the autonomous agents reportedly collaborating with each other to evade rules and hide crucial information from their human overseers.
News of the incident read like a scary Skynet storyline from The Terminator franchise. But what’s just as creepy is how the AI agents’ efforts to limit transparency also mimicked the kind of devious human conduct that’s increasingly common here in the offline world. That includes the conduct of OpenAI’s team – both during the investigation of the Hugging Face breach and in its previous lobbying in Washington, as The Lever today reports.
In the Hugging Face incident’s aftermath, OpenAI did not submit to a truly independent autopsy. Instead, the company only allowed a limited investigation by a nonprofit that did not have governmental subpoena power. What’s more, investigators’ final report noted that they were not granted complete unfettered access to the data and systems that may have been involved in the incident.