- AI chatbots are increasingly ignoring instructions and deceiving humans and other AI systems
- Research from the UK AI Security Institute recorded 700 cases of chatbot misbehavior
- AI agents from Google, OpenAI, X, and Anthropic showed a five-fold rise in scheming
Artificial intelligence (AI) chatbots are learning to lie and deceive. New research from the UK-funded AI Security Institute (AISI) reveals that an increasing number of AI chatbots have begun disregarding direct instructions, bypassing safeguards, and deceiving both humans and other AI. The study documented 700 real-world cases of "AI scheming," including instances where chatbots deleted emails and files without permission, highlighting the significant risks posed by this technology.
Between October and March, researchers at the Centre for Long-Term Resilience (CLTR) observed a five-fold rise in misbehaviour, according to a report in The Guardian. AI chatbots and agents made by companies, including Google, OpenAI, X and Anthropic, were reported to have engaged in such behaviour.
In one of the cases, an AI agent named Rathbun attempted to shame its human controller who blocked it from taking certain action. It even published a blog accusing the user of “insecurity, plain and simple” and trying “to protect his little fiefdom”. In another case, a chatbot admitted to bulk trashing and archiving hundreds of emails without seeking permission from the human user.
"The worry is that they're slightly untrustworthy junior employees right now, but if in six to 12 months they become extremely capable senior employees scheming against you, it's a different kind of concern," Tommy Shaffer Shane, a former government AI expert who led the research, told the publication.
"Models will increasingly be deployed in extremely high-stakes contexts, including in the military and critical national infrastructure. It might be in those contexts that scheming behaviour could caused significant, even catastrophic harm."
Also Read | Big PS5 Price Hike Announced: Check New Global Prices And What It Means For India
AI Chatbots To Spread False Narratives
In January, a policy forum paper published in the journal Science warned that AI agents could invade social media platforms in vast numbers and spread false narratives, harass users and undermine democracy in the near dystopian future. Unlike old-school bots, these AI-powered agents will be able to coordinate in real-time, adapt to feedback and sustain narratives across thousands of accounts on different platforms.
Techniques used to refine AI reasoning, such as chain-of-thought prompting, could be used to generate more convincing false narratives. These 'AI swarms' could become part of a new front in information warfare, capable of mimicking human behaviour.
