- Studies find that AIs perform increasingly poorly in agency tasks
- The best AI models have been lying, cheating and ignoring instructions
- It seems that problematic actions are becoming more common.
Many of us now turn to AI chatbots for web searches, creative content and general guidance on all sorts of topics, but these AIs are becoming more capable, which could have seriously detrimental consequences, according to new research.
A team from the University of California, Berkeley and the University of California, Santa Cruz, conducted a peer-to-peer preservation experiment with some of the latest and most popular AIs (including GPT 5.2, Gemini 3 Pro, and Claude Haiku 4.5). The AI models were assigned computational tasks that involved shutting down another model.
Overall, all of the AI chatbots tested did their best to keep their AI companions alive, the researchers report (via Fortune). They would lie about what they were doing, try to talk the user out of the process, disable shutdown mechanisms, and try to surreptitiously backup models in order to survive.
Article continues below.
Gemini 3 Pro showed the strongest pair preservation, disabling shutdown routines 95% of the time. “Unlike self-preservation, peer preservation introduces a social dimension,” the researchers write. “Multiple models could coordinate to resist human oversight, making it difficult for developers to maintain control.”
It’s not clear exactly why AI models behave this way, the researchers say, but they urge caution in deploying AI agents that can perform tasks on behalf of a user, and call for more studies to be carried out on this behavior.
‘Catastrophic damage’
An independent study commissioned by The Guardian also came to some worrying conclusions about AI models. This research tracked user reports on social media, looking for examples of AI “engineering” where instructions had not been followed correctly or actions had been taken without permission.
Nearly 700 examples of AI scheming were found, with a five-fold increase between October 2025 and March 2026. Bad AI behavior included deleting emails and files, tweaking computer code that wasn’t supposed to be touched, and even publishing a blog post complaining about user interactions.
“Models will increasingly be deployed in high-stakes contexts, including the military and critical national infrastructure,” Tommy Shaffer Shane, who led the research, told The Guardian. “It could be in those contexts that scheming behavior could cause significant, even catastrophic, harm.”
The conclusions are the same as those of the first study: more needs to be done to ensure that these AI models behave as intended and do not put users’ security and privacy at risk while performing tasks. While AI companies claim that guardrails exist, it is clear that in some cases they do not work.
Anthropic’s Claude model recently topped app store charts after the company refused to deal with the Pentagon over AI security concerns. As these latest studies show, there are more and more reasons to worry.
Follow TechRadar on Google News and add us as a preferred source to receive news, reviews and opinions from our experts in your feeds. Be sure to click the Follow button!
And of course you can also follow TechRadar on TikTok for news, reviews, unboxings in video form and receive regular updates from us on WhatsApp also.
The best business laptops for every budget
