The dangers of artificial intelligence are well documented, from helping students cheat to creating porn fakes. While AI is already being used for nefarious means, new research shows AI programs are even learning to be nefarious on their own. The research published in StudyFinds, tested Meta's AI system designed to play the strategy game Diplomacy, and found the AI program used dishonesty to win the game against human players. The authors wrote that Meta's AI had learned to be a "master of deception," noting that "Meta failed to train its AI to win honestly."
Cheating to win a game is one thing, but it raises greater concerns about using AI to cheat at more important or dangerous things. Congressman Jay Obernolte (R-CA) chairs the House AI Task Force. He recently told Fox News AI is not essentially good or bad, but a reflection of what we put into it. "AI is a tool for enhancing productivity, and that means it enhances productivity in good ways and bad ways," he says. "We say that bad people are gonna bad, and they're gonna bad more effectively with AI."
"So, I think it's our job in government to make sure that those negative consequences don't occur."
How to prevent those negative consequences is the tricky part. The Senate just introduced a 31-page "road map" on regulating AI. Obernolte notes there are some areas of obvious immediate concern with AI, like using it for election interference, or the aforementioned porn fakes. But he admits there is still a lot about AI that is unknown. "It's going to take some time for us to figure out exactly what needs to be done," he says. "But those short-term threats, I think we can mitigate those this year, and I'm hopeful that the task force will be able to get that done."
"The goal is to actually pass legislation," he continues. "It does no one any good if we just talk about it, or we just have meetings or create reports."