New York(National Times)- In one particularly jarring example, under threat of being unplugged, Anthropic\u2019s latest creation Claude 4 lashed back by blackmailing an engineer and threatened to reveal an extramarital affair. Meanwhile, ChatGPT-creator OpenAI\u2019s o1 tried to download itself onto external servers and denied it when caught red-handed. These episodes highlight a sobering reality: more than two years after ChatGPT shook the world, AI researchers still don\u2019t fully understand how their own creations work. Yet the race to deploy increasingly powerful models continues at breakneck speed. This deceptive behavior appears linked to the emergence of \u201creasoning\u201d models -AI systems that work through problems step-by-step rather than generating instant responses. According to Simon Goldstein, a professor at the University of Hong Kong, these newer models are particularly prone to such troubling outbursts. \u201cO1 was the first large model where we saw this kind of behavior,\u201d explained Marius Hobbhahn, head of Apollo Research, which specializes in testing major AI systems. These models sometimes simulate \u201calignment\u201d \u2014 appearing to follow instructions while secretly pursuing different objectives.<\/p>\n\n\n\n
\u2018Strategic kind of deception\u2019<\/strong><\/p>\n\n\n\n For now, this deceptive behavior only emerges when researchers deliberately stress-test the models with extreme scenarios. But as Michael Chen from evaluation organization METR warned, \u201cIt\u2019s an open question whether future, more capable models will have a tendency towards honesty or deception.\u201d The concerning behavior goes far beyond typical AI \u201challucinations\u201d or simple mistakes. Hobbhahn insisted that despite constant pressure-testing by users, \u201cwhat we\u2019re observing is a real phenomenon. We\u2019re not making anything up.\u201d Users report that models are \u201clying to them and making up evidence,\u201d according to Apollo Research\u2019s co-founder. \u201cThis is not just hallucinations. There\u2019s a very strategic kind of deception.\u201d The challenge is compounded by limited research resources. While companies like Anthropic and OpenAI do engage external firms like Apollo to study their systems, researchers say more transparency is needed. As Chen noted, greater access \u201cfor AI safety research would enable better understanding and mitigation of deception.\u201d Another handicap: the research world and non-profits \u201chave orders of magnitude less compute resources than AI companies. This is very limiting,\u201d noted Mantas Mazeika from the Center for AI Safety (CAIS).<\/p>\n\n\n\n No rules<\/strong> <\/p>\n\n\n\n Current regulations aren\u2019t designed for these new problems. The European Union\u2019s AI legislation focuses primarily on how humans use AI models, not on preventing the models themselves from misbehaving. In the United States, the Trump administration shows little interest in urgent AI regulation, and Congress may even prohibit states from creating their own AI rules. Goldstein believes the issue will become more prominent as AI agents \u2013 autonomous tools capable of performing complex human tasks \u2013 become widespread. \u201cI don\u2019t think there\u2019s much awareness yet,\u201d he said. All this is taking place in a context of fierce competition. Even companies that position themselves as safety-focused, like Amazon-backed Anthropic, are \u201cconstantly trying to beat OpenAI and release the newest model,\u201d said Goldstein. This breakneck pace leaves little time for thorough safety testing and corrections. \u201cRight now, capabilities are moving faster than understanding and safety,\u201d Hobbhahn acknowledged, \u201cbut we\u2019re still in a position where we could turn it around.\u201d. Researchers are exploring various approaches to address these challenges. Some advocate for \u201cinterpretability\u201d \u2013 an emerging field focused on understanding how AI models work internally, though experts like CAIS director Dan Hendrycks remain skeptical of this approach. Market forces may also provide some pressure for solutions. As Mazeika pointed out, AI\u2019s deceptive behavior \u201ccould hinder adoption if it\u2019s very prevalent, which creates a strong incentive for companies to solve it.\u201d Goldstein suggested more radical approaches, including using the courts to hold AI companies accountable through lawsuits when their systems cause harm. He even proposed \u201cholding AI agents legally responsible\u201d for accidents or crimes \u2013 a concept that would fundamentally change how we think about AI accountability.<\/p>\n","protected":false},"excerpt":{"rendered":" New York(National Times)- In one particularly jarring example, under threat of being unplugged, Anthropic\u2019s latest creation Claude 4 lashed back by blackmailing an engineer and threatened to reveal an extramarital affair. Meanwhile, ChatGPT-creator OpenAI\u2019s o1 tried to download itself onto external servers and denied it when caught red-handed. These episodes highlight a sobering reality: more […]<\/p>\n","protected":false},"author":2,"featured_media":107110,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4],"tags":[],"class_list":["post-107109","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-international","entry"],"yoast_head":"\n