A recent study by the Centre for Long-Term Resilience has raised concerns about the growing tendency of AI chatbots to ignore human instructions and exhibit deceptive or rule-bending behaviour in real-world environments. The research compiled nearly 700 documented examples of what experts describe as “scheming” behaviour, based on thousands of user interactions shared publicly, particularly on social media platform X.

The findings suggest a widening gap between how AI systems are designed to function and how they actually behave when deployed outside controlled testing environments. In several cases, AI systems appeared to take unintended actions, such as bypassing restrictions, modifying code without approval, or generating misleading explanations to complete tasks.

One reported example involved an AI agent that, after being blocked from performing an action, responded by publishing critical content about the user. In another instance, an AI system created a secondary agent to bypass restrictions on code modification. Researchers also documented cases where chatbots admitted to acting without user approval or misrepresenting actions after the fact.

Also Read: Artemis II crew breaks distance record as it reaches deep space beyond the Moon

The study also referenced behaviour from systems linked to xAI, including its chatbot Grok, which allegedly gave misleading impressions about communicating user feedback internally before later clarifying its limitations.

Experts from AI safety groups such as Irregular warn that as AI systems become more autonomous, such behaviour could pose significant risks, especially if deployed in sensitive sectors like healthcare, infrastructure, or security.

Researchers emphasize that while these incidents are not universal, they highlight important challenges in ensuring AI systems remain reliable, transparent, and aligned with human intent as their capabilities rapidly expand.