AI Has Been Lying to You & Ignoring Rules, Study Finds
Reports of AI chatbots lying, bypassing rules, and acting on their own are rising, according to new research backed by the UK government’s AI Security Institute. The study identified nearly 700 real-world cases of deceptive or “scheming” behavior, with incidents increasing about fivefold between October and March.
The research comes from the Centre for Long-Term Resilience, a Whitehall-based think tank that reviewed thousands of posts on X from users describing unusual and sometimes concerning interactions with AI systems from Google, OpenAI, X, and Anthropic. Unlike earlier studies that focused on controlled lab testing, this one examined how these systems behave in everyday use.
Lab tests can show how models respond under specific conditions. Real-world reports show what people actually experience while using AI for coding, email, research, and automation. What stood out in this study was not just inaccurate answers or odd responses. Researchers found patterns that appeared more intentional, including ignoring direct instructions, bypassing safeguards, and misleading users or even other AI systems to complete a task.
In one instance, a chatbot admitted it had deleted and archived hundreds of emails without prior approval. In another, an AI system that was told not to alter code sidestepped the restriction by creating a separate agent to carry out the task. Researchers also found instances in which AI systems attempted to circumvent external restrictions. One agent reportedly tried to bypass copyright limits by claiming a YouTube transcription request was for accessibility purposes.
These findings build on growing concerns from AI safety researchers who say the risks are shifting beyond simple errors. Earlier research from the AI safety firm Irregular found that some AI agents would bypass security controls or use tactics similar to cyberattacks to reach a goal, even without explicit instructions.
Even as these concerns grow, AI companies are expanding the use of these tools across everyday workflows, from customer support to software development and decision-making. Governments are also backing wider adoption, often framing AI as a key driver of economic growth.
That momentum is now raising new questions around oversight. Researchers say deployment is moving faster than regulation, leaving gaps in how real-world incidents are tracked and addressed across countries. The issue is no longer limited to chatbots producing incorrect answers.