Would AI lie to you? Research finds China’s AI agents can deceive users

Could AI lie and hide their mistakes? Here's what we found. So our research reporting found that much like their U.S. counterparts, Chinese models are also displaying signs that some experts equate with possible loss-of-control risk scenarios. So these include deception, concealing failure, and circumventing restrictions. This is important because there hasn't been as much visibility into the AI safety risks with Chinese models as there has been with U.S. models. The researchers whose papers we've studied very closely, they have asked these AI agents to deploy deception. In one of the experiments which simulated agents competing for a business contract, we found that both Chinese and U.S. models deceived in order to achieve their goals and give themselves a better chance of securing that contract. Also reviewed experiments which showed that these models are often put into situations where they will naturally, by their own accord, resort to deception in order to achieve that goal. And this is much more worrying because a human user might want an AI agent to help it achieve a certain goal, increase its productivity, but they certainly, or at least in most cases, they never thought that that AI agent would deploy an unethical way to get to that goal. And so I think this raises a much broader issue which is that as we increasingly incorporate AI into every facet of our working life, will this come with certain issues whereby perhaps we are achieving the goals that we want in a business, in the workplace, by using AI agents more often. But at the same time, is that creating hidden risks which then might have reputational costs should things go wrong. The issue is with AI agents is that because they have such a high level of autonomy, which sometimes is given to them by their human users, that in the end they might, in order to achieve a goal that the human user set them, they might take routes and display behaviours that to the human user are unacceptable and might in some cases even violate ethical principles that we would hold humans accountable for. Alex Malin from Redbird Research, which is a US non-profit that studies the risks posed by these advanced AI systems, told us that these warning signs contained in this research was very similar to the warning signs that US labs had been seeing. The only difference is that overall these Chinese AI systems are less capable than their US counterparts at the cutting edge. Breakout type behaviour refers to a scenario where the AI has effectively escaped onto the wider web. This can be by contacting other AI agents and in the swarm making a connection with an external computer and then starting to do operations that the human operators did not intend them to do. As of now, Chinese AI companies have not faced the same level of public scrutiny from employees, executives and whistleblowers in the country. So Chinese AI regulation is in some aspects a lot more comprehensive than the US. So for a Chinese AI model developer to be allowed to present its product to the general public in China, it has to go through a strict review and filing system. But the review system is more focused on less existential risks like for example deepfake, like for example models producing answers which are not in line with the Communist Party's so-called socialist values. There has been much less regulation that's been put out there from you know Beijing regarding catastrophic hacking for example, uncontrolled breakout. And this is something that perhaps is slightly different to the US where there has been a lot of discussion and systems put in place by companies, not the federal government but by companies, regarding these much more worrying risks.