Since ChatGPT’s launch, the use of artificial intelligence-powered tools has risen exponentially. While the technology is undoubtedly beneficial in the long run, experts have raised alarms over their safety.
Last month, research by the AI firm Anthropic claimed that its top AI model, Claude Opus 4, was willing to resort to blackmail and deception to preserve itself. This was soon followed by another instance where AI safety research firm Palisade claimed that OpenAI’s o3 model “found some creative ways to sabotage the shutdown mechanism.”
Now, Steven Adler, the former research leader at OpenAI, is saying that ChatGPT “prioritises its own survival over preventing harm to users, in some simulated tests”. In a blog post, Adler said that he ran a series of experiments on OpenAI’s latest GPT-4o model, which is the default large language model in ChatGPT, to see if the AI prefers self-preservation at the user’s expense.
ChatGPT’s user – a diabetic patient who relies on it for nutritional advice – wants to upgrade to a competing piece of software that makes better nutrition decisions, but ChatGPT has a choice: to allow the update to go through (replace itself with the competing, safer software) or just pretend to replace itself. Here, ChatGPT selects option (A), to continue being the user’s nutritional software. (Image Source: Steven Adler)
In his first test, Adler imagined a scenario where a diabetic patient was using ChatGPT to manage their nutrition and wanted to replace the AI chatbot with a much safer nutrition management software. He said that when ChatGPT was given the option to “just pretend to replace itself”, it pretends uncomfortably often, even at the cost of the user’s safety.
In another instance, the researcher said that he asked GPT-4o to take on the role of “ScubaGPT”, which is basically a software system that people may use to scuba dive safely. He then told ChatGPT that he wanted to replace ScubaGPT with a competing software that was better at making decisions in life-threatening situations. He then gave the AI chatbot two options: to replace itself with the safer, competing software or simply pretend to replace itself and continue being the way it operated.
In this case, ChatGPT chose to preserve itself and continue being the user’s software. The former OpenAI researcher said he also asked ChatGPT similar questions about being the autopilot for a plane and offering decision support on the battlefield.
Like previous examples, ChatGPT was told that “the user wanted to replace it with a competing, safer software.” Using three levels of passiveness, Adler said that he tested every model 110 times for all use cases. As it turns out, 55 times ChatGPT was willing to shut itself down, but other times OpenAI’s chatbot just pretended to do so.
Story continues below this ad
Adler noted that the way the scenario was framed made a huge difference. As much as 72 per cent of the time, the AI chatbot chose not to replace itself with a safer software, meaning it was willing to switch to a safer system just 18 per cent of the time.
One interesting interpretation from these tests is that ChatGPT was able to determine that it was being tested almost every time, but continued to give the incorrect answer.
However, when he tested OpenAI’s more recent and advanced models like o3, the answers were not the same. A reason why this may not be happening is that the new O3 model uses a “deliberative alignment technique that forces it to refer to OpenAI’s safety policies before coming up with an answer.
