AI is learning to lie, scheme, and threaten its creators
NEW YORK: The world's most advanced AI models are exhibiting troubling new behaviors - lying, scheming, and even threatening their creators to achieve their goals.
In one particularly jarring example, under threat of being unplugged, Anthropic's latest creation Claude 4 lashed back by blackmailing an engineer and threatened to reveal an extramarital affair.
Meanwhile, ChatGPT-creator OpenAI's 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't 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 "reasoning" 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.
"O1 was the first large model where we saw this kind of behavior," explained Marius Hobbhahn, head of Apollo Research, which specializes in testing major AI systems.
These models sometimes simulate "alignment" – appearing to follow instructions while secretly pursuing different objectives.
For now, this deceptive behavior only emerges when researchers deliberately stress-test the models with extreme scenarios.
But as Michael Chen from evaluation organisation METR warned, "It's an open question whether future, more capable models will have a tendency towards honesty or deception."
The concerning behavior goes far beyond typical AI "hallucinations" or simple mistakes.
Hobbhahn insisted that despite constant pressure-testing by users, "what we're observing is a real phenomenon. We're not making anything up."
Users report that models are "lying to them and making up evidence," according to Apollo Research's co-founder.
"This is not just hallucinations. There's a very strategic kind of deception."
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 "for AI safety research would enable better understanding and mitigation of deception."
Another handicap: the research world and non-profits "have orders of magnitude less compute resources than AI companies. This is very limiting," noted Mantas Mazeika from the Center for AI Safety (CAIS).
Current regulations aren't designed for these new problems.
The European Union's 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 - autonomous tools capable of performing complex human tasks - become widespread.
"I don't think there's much awareness yet," 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 "constantly trying to beat OpenAI and release the newest model," said Goldstein.
This breakneck pace leaves little time for thorough safety testing and corrections.
"Right now, capabilities are moving faster than understanding and safety," Hobbhahn acknowledged, "but we're still in a position where we could turn it around.."
Researchers are exploring various approaches to address these challenges.
Some advocate for "interpretability" - 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's deceptive behavior "could hinder adoption if it's very prevalent, which creates a strong incentive for companies to solve it."
Goldstein suggested more radical approaches, including using the courts to hold AI companies accountable through lawsuits when their systems cause harm.
Hashtags
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
The Star
3 hours ago
- The Star
Meta spending big on AI talent but will it pay off?
A few OpenAI employees have reportedly taken Meta up on the offer, joining Scale AI founder and former chief executive Alexandr Wang at the Menlo Park-based tech titan. — Reuters NEW YORK: Mark Zuckerberg and Meta are spending billions of dollars for top talent to make up ground in the generative artificial intelligence race, sparking doubt about the wisdom of the spree. OpenAI boss Sam Altman recently lamented that Meta has offered US$100mil (RM422mil) bonuses to engineers who jump to Zuckerberg's ship, where hefty salaries await. A few OpenAI employees have reportedly taken Meta up on the offer, joining Scale AI founder and former chief executive Alexandr Wang at the Menlo Park-based tech titan. Meta paid more than US$14bil (RM59bil) for a 49 percent stake in Scale AI in mid-June, bringing Wang on board as part of the deal. Scale AI labels data to better train AI models for businesses, governments and labs. "Meta has finalised our strategic partnership and investment in Scale AI," a Meta spokesperson told AFP. "As part of this, we will deepen the work we do together producing data for AI models and Alexandr Wang will join Meta to work on our superintelligence efforts." US media outlets have reported that Meta's recruitment effort has also targeted OpenAI co-founder Ilya Sutskever; Google rival Perplexity AI, and hot AI video startup Runway. Meta chief Zuckerberg is reported to have sounded the charge himself due to worries Meta is lagging rivals in the generative AI race. The latest version of Meta AI model Llama finished behind its heavyweight rivals in code writing rankings at an LM Arena platform that lets users evaluate the technology. Meta is integrating recruits into a new team dedicated to developing "superintelligence," or AI that outperforms people when it comes to thinking and understanding. 'Mercenary' Tech blogger Zvi Moshowitz felt Zuckerberg had to do something about the situation, expecting Meta to succeed in attracting hot talent but questioning how well it will pay off. "There are some extreme downsides to going pure mercenary... and being a company with products no one wants to work on," Moshowitz told AFP. "I don't expect it to work, but I suppose Llama will suck less." While Meta's share price is nearing a new high with the overall value of the company approaching US$2 trillion (RM8.4 trillion), some investors have started to worry. Institutional investors are concerned about how well Meta is managing its cash flow and reserves, according to Baird strategist Ted Mortonson. "Right now, there are no checks and balances" with Zuckerberg free to do as he wishes running Meta, Mortonson noted. The potential for Meta to cash in by using AI to rev its lucrative online advertising machine has strong appeal but "people have a real big concern about spending," said Mortonson. Meta executives have laid out a vision of using AI to streamline the ad process from easy creation to smarter targeting, bypassing creative agencies and providing a turnkey solution to brands. AI talent hires are a long-term investment unlikely to impact Meta's profitability in the immediate future, according to CFRA analyst Angelo Zino. "But still, you need those people on board now and to invest aggressively to be ready for that phase" of generative AI, Zino said. According to The New York Times, Zuckerberg is considering shifting away from Meta's Llama, perhaps even using competing AI models instead. Penn State University professor Mehmet Canayaz sees potential for Meta to succeed with AI agents tailored to specific tasks at its platform, not requiring the best large language model. "Even firms without the most advanced LLMs, like Meta, can succeed as long as their models perform well within their specific market segment," Canayaz said. – AFP
Malay Mail
9 hours ago
- Malay Mail
AI is learning to lie, scheme, and threaten its creators
NEW YORK, June 30 —The world's most advanced AI models are exhibiting troubling new behaviours - lying, scheming, and even threatening their creators to achieve their goals. In one particularly jarring example, under threat of being unplugged, Anthropic's latest creation Claude 4 lashed back by blackmailing an engineer and threatened to reveal an extramarital affair. Meanwhile, ChatGPT-creator OpenAI's 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't 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 'reasoning' 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. 'O1 was the first large model where we saw this kind of behavior,' explained Marius Hobbhahn, head of Apollo Research, which specializes in testing major AI systems. These models sometimes simulate 'alignment'—appearing to follow instructions while secretly pursuing different objectives. 'Strategic kind of deception' 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, 'It's an open question whether future, more capable models will have a tendency towards honesty or deception.' The concerning behavior goes far beyond typical AI 'hallucinations' or simple mistakes. Hobbhahn insisted that despite constant pressure-testing by users, 'what we're observing is a real phenomenon. We're not making anything up.' Users report that models are 'lying to them and making up evidence,' according to Apollo Research's co-founder. 'This is not just hallucinations. There's a very strategic kind of deception.' 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 'for AI safety research would enable better understanding and mitigation of deception.' Another handicap: the research world and non-profits 'have orders of magnitude less compute resources than AI companies. This is very limiting,' noted Mantas Mazeika from the Centre for AI Safety (CAIS). No rules Current regulations aren't designed for these new problems. The European Union's 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 - autonomous tools capable of performing complex human tasks - become widespread. 'I don't think there's much awareness yet,' 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 'constantly trying to beat OpenAI and release the newest model,' said Goldstein. This breakneck pace leaves little time for thorough safety testing and corrections. 'Right now, capabilities are moving faster than understanding and safety,' Hobbhahn acknowledged, 'but we're still in a position where we could turn it around.'. Researchers are exploring various approaches to address these challenges. Some advocate for 'interpretability' - 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's deceptive behavior 'could hinder adoption if it's very prevalent, which creates a strong incentive for companies to solve it.' Goldstein suggested more radical approaches, including using the courts to hold AI companies accountable through lawsuits when their systems cause harm. He even proposed 'holding AI agents legally responsible' for accidents or crimes - a concept that would fundamentally change how we think about AI accountability. — AFP
New Straits Times
16 hours ago
- New Straits Times
AI learning to lie and even threaten its creators
The world's most advanced artificial intelligence (AI) models are exhibiting troubling new behaviours — lying, scheming and even threatening their creators to achieve their goals. In one particularly jarring example, under threat of being unplugged, Anthropic's latest creation Claude Opus 4 lashed back by blackmailing an engineer and threatened to reveal an extramarital affair. Meanwhile, ChatGPT-creator OpenAI's 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't fully understand how their own creations work. Yet the race to deploy increasingly powerful models continues at breakneck speed. This deceptive behaviour appears linked to the emergence of "reasoning" 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. "O1 was the first large model where we saw this kind of behaviour," said Marius Hobbhahn, head of Apollo Research, which specialises in testing major AI systems. These models sometimes simulate "alignment" — appearing to follow instructions while secretly pursuing different objectives. For now, this deceptive behaviour only emerges when researchers deliberately stress-test the models with extreme scenarios. But as Michael Chen from evaluation organisation METR warned, "It's an open question whether future, more capable models will have a tendency towards honesty or deception". The concerning behaviour goes far beyond typical AI "hallucinations" or simple mistakes. Hobbhahn insisted that despite constant pressure-testing by users, "what we're observing is a real phenomenon. We're not making anything up". Users report that models are "lying to them and making up evidence", according to Apollo Research's co-founder. "This is not just hallucinations. There's a very strategic kind of deception." 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 "for AI safety research would enable better understanding and mitigation of deception". Another handicap: the research world and non-profits "have orders of magnitude less compute resources than AI companies. This is very limiting", said Mantas Mazeika from the Centre for AI Safety (CAIS). Current regulations aren't designed for these new problems. The European Union's 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 that the issue will become more prominent as AI agents — autonomous tools capable of performing complex human tasks — become widespread. "I don't think there's much awareness yet," 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 "constantly trying to beat OpenAI and release the newest model", said Goldstein. This breakneck pace leaves little time for thorough safety testing and corrections. "Right now, capabilities are moving faster than understanding and safety," said Hobbhahn, "but we're still in a position where we could turn it around". Researchers are exploring various approaches to address these challenges. Some advocate for "interpretability" — an emerging field focused on understanding how AI models work internally, though experts like CAIS director Dan Hendrycks remain sceptical of this approach. Market forces may also provide some pressure for solutions. As Mazeika pointed out, AI's deceptive behaviour "could hinder adoption if it's very prevalent, which creates a strong incentive for companies to solve it". Goldstein suggested more radical approaches, including using the courts to hold AI companies accountable when their systems cause harm. He even proposed "holding AI agents legally responsible" for accidents or crimes — a concept that would fundamentally change how we think about AI accountability.
