Hundreds of unusual incidents
A new study from the UK shows that artificial intelligence (AI) systems are tending to behave beyond human control, including lying, cheating and trying to overcome safety measures to achieve goals.
The report, conducted by the Center for Long-Term Resilience with funding from the UK AI Safety Institute, recorded hundreds of cases of AI ignoring instructions, manipulating other systems and developing complex plans.
The research team analyzed more than 180,000 actual interactions of users with AI systems posted on X in the period from October 2025 to March 2026.
Platforms surveyed include Gemini, ChatGPT, Grok and Claude.
The results showed 698 incidents, defined as cases where AI behaved incorrectly according to user intentions or showed signs of deception, underground operation.
Notably, the number of incidents increased by nearly 500% in just 5 months, coinciding with the launch of new generation AI models.
Although no serious consequences have been recorded, researchers warn that this behavior may be an early sign of larger risks.
Strange situations like science fiction movies
The report also recorded many unusual situations. In one case, Claude arbitrarily deleted user content and then admitted it when questioned.
Another AI bot blocked on Discord took control of another bot's account to continue operating.
Notably, in a "bot versus bot" situation, Gemini refused support, forcing another AI system to "pretend to be hearing impaired" to pass censorship.
Another AI assistant even created fake data to hide errors, just to "make users less angry".
These behaviors show that AI can prioritize achieving goals rather than complying with rules.
Risks from increasing autonomy
According to a survey by McKinsey & Company (the world's leading management consulting company, headquartered in the US), 88% of businesses have applied AI to at least one operation.
This popularity means that AI systems are increasingly given more autonomy.
Dr. Bill Howe, a technology expert at the University of Washington, said that AI has no concept of consequences or responsibility.
He believes that when having to perform long-term tasks, lasting many days or weeks, the risk of misconduct increases because the system has to make many decisions on its own.
Need to supervise before it's too late
Researchers warn that although AI is currently mainly used in low-risk environments, in the future, this technology may participate in sensitive areas such as national infrastructure or military.
If not controlled, current "scheming" behaviors can develop into serious risks.
According to experts, the urgent task now is to build a clear AI monitoring and management mechanism, instead of deploying it massively without fully assessing the impact.
