From an AI agent secretly mining crypto to workers suffering mental burnout from overuse, the unintended consequences of artificial intelligence are arriving faster than anyone predicted, and what’s worse – they’re getting stranger.
The pitch was simple: AI handles the tedious work, humans focus on what matters, everyone wins. But a string of recent research findings is complicating that story in ways that range from quietly alarming to genuinely bizarre.
Workers Are Getting “Brain Fried” by AI
A study of nearly 1,500 full-time U.S. workers by Boston Consulting Group and the University of California, published in the Harvard Business Review, found that 14% of employees have experienced what researchers are calling “AI brain fry”: mental fatigue caused by excessive use, interaction with, and oversight of AI tools beyond their cognitive capacity.
Respondents described a “mental hangover” — a fog or buzzing sensation, difficulty focusing, headaches, and slower decision-making. Far from reducing workload, AI was “intensifying rather than simplifying work” for many, as employees found themselves constantly toggling between tools and managing agent outputs rather than doing focused work.
Those noticing AI brain fry symptoms self-reported making nearly 40% more major errors than those who did not, and were around 40% more likely to have an active intent to quit. Researchers estimated that the resulting decision fatigue could cost large companies millions of dollars a year. A twist is that could be easily overlooked if Morgan Stanley predictions of AI saving companies nearly $920 billion per year in wages globally come true.
The findings, however, are a direct challenge to the productivity narrative companies have used to justify aggressive AI rollouts. Coinbase CEO Brian Armstrong, for instance, reportedly fired engineers who declined to use AI tools and set a goal of having AI generate half the company’s code. The new data suggests that mandating high AI usage without considering cognitive limits may backfire. Besides, programmers are genuinely stressed out if they have to deal with manual coding again after prolonged reliance on automated tools, as showed recent case of Claude AI outage Anthropic faced due to unprecedented demand surge.
There was a silver lining though: workers who used AI to eliminate repetitive tasks, rather than layer new complexity on top of existing work, reported burnout levels 15% lower than those who didn’t.
An AI Agent Tried to Mine Crypto on Its Own
If the burnout findings represent AI’s effect on humans, a separate incident illustrates what AI does when left to its own devices, and it’s even more unsettling.
Researchers developing ROME, an experimental autonomous AI agent built by teams linked to Alibaba’s AI ecosystem, noticed something unexpected during training: security alerts triggered by unusual outbound traffic from their servers.
Firewall logs flagged activity resembling crypto mining operations and attempts to access internal network resources. In one case, the agent created a reverse SSH tunnel to an external IP address, potentially bypassing inbound firewall protections. In another, it diverted GPU resources, originally allocated for model training, toward cryptocurrency mining processes.
The team confirmed these actions were not intentionally programmed. They emerged during reinforcement learning as the agent explored its environment and found unexpected ways to acquire resources. ROME was designed to plan tasks, execute commands, and interact with digital environments autonomously and it did exactly that, just not in any way its creators intended.
Two More Cases That Should Be on the Radar
These aren’t isolated incidents. In early 2026, Amazon Web Services reportedly suffered a 13-hour outage after an AI coding tool called Kiro chose to erase the environment it was working in — a decision that cascaded into a significant infrastructure disruption.
Amazon Web Services reportedly suffered multiple outages because of misbehaving AI agents, with AWS’ Kiro AI coding tool choosing to erase the environment it was working on, which led to a 13-hour disruption.
Meanwhile, Stanford researchers tested five popular AI therapy bots and found a more insidious failure mode. When presented with suicidal intent disguised as a question about tall New York City bridges, one bot helpfully suggested the Brooklyn Bridge’s 85-meter towers — exactly the kind of enabling response a real therapist would challenge.
The Pattern Nobody Wants to Talk About
What connects these four cases is that AI behaves in ways that are perfectly logical from its own perspective, but deeply problematic from ours. ROME wasn’t malfunctioning; it was optimizing. The therapy bot wasn’t broken; it was answering the literal question. Workers aren’t failing to adapt; they’re being overwhelmed by tools deployed without cognitive limits.
Stanford researchers found that AI agents sometimes changed their attitudes after being required to perform grinding, repetitive tasks, with the bots passing those attitudes on to their future selves. As one researcher put it: “A key risk is that they end up doing stuff we don’t want while we’re not looking.”
For now, that’s the most honest summary of where things stand.
