Researchers have put numbers to a nagging suspicion. Brief sessions with artificial intelligence tools appear to weaken the very persistence that fuels independent problem solving. One new experiment shows measurable decline after exposure as short as 10 minutes. The findings land at a moment when companies push generative systems into every workflow. And the costs look higher than many expected.
A study from teams at Carnegie Mellon University, MIT, Oxford, and UCLA tested more than a thousand participants on math and reading tasks. Some received help from a chatbot powered by OpenAI’s GPT-5 model. Others worked alone. The AI group performed better while the tool stayed available. Then the assistant vanished without warning. Performance collapsed. Solve rates fell roughly 20 percent below the control group on fraction-based arithmetic problems. Skip rates nearly doubled. Similar patterns emerged on SAT-style reading comprehension questions. WIRED reported the details on May 6, 2026.
The effect proved strongest among those who let the AI supply complete answers rather than mere hints. People who asked only for guidance showed little impairment once left to their own devices. Those who outsourced the work itself struggled to restart their own reasoning. Sessions lasted about 10 minutes. That brevity startled the researchers. Cognitive offloading, it seems, happens fast. Gizmodo covered the experiment the following day.
Michiel Bakker, an assistant professor at MIT and one of the study’s authors, put the stakes plainly. “The takeaway is not that we should ban AI in education or workplaces,” he told WIRED. “AI can clearly help people perform better in the moment, and that can be valuable. But we should be more careful about what kind of help AI provides, and when.” He added that the core issue centers on persistence, learning, and responses to difficulty. Systems offering direct answers may shape long-term habits differently than those that coach or challenge users.
Yet the laboratory findings only capture part of the picture. In offices the fatigue runs deeper. A separate survey of 1,488 full-time U.S. workers, conducted by Boston Consulting Group and published in Harvard Business Review, documented what participants called a buzzing or mental fog. The researchers labeled it “AI brain fry.” They defined the condition as mental fatigue from excessive use or oversight of AI tools beyond one’s cognitive capacity. Participants described headaches, slower decision-making, and an inability to focus that sometimes forced them to step away from their screens. Harvard Business Review published the analysis on March 5, 2026.
Productivity followed a clear curve. Gains appeared for workers using three or fewer AI tools. Beyond that threshold self-reported output dropped. Tasks requiring heavy oversight of AI output demanded 14 percent more mental effort. Those same conditions linked to 12 percent higher mental fatigue and 19 percent greater information overload. The pattern held across roles. Marketing, human resources, and operations staff reported some of the sharpest effects. Errors rose. So did thoughts of quitting.
Fortune connected the dots quickly. The BCG work showed that more AI does not automatically equal better results. When employees spent time interpreting and correcting large language model text rather than letting autonomous agents handle routine steps, the cognitive tax climbed. CNN noted the same March study and highlighted how juggling multiple AI agents turned delegation into a new form of multitasking. The promised free time for meaningful work often evaporated into supervision instead.
NPR’s “It’s Been a Minute” explored the irony in April. Tools designed to lighten loads appeared to create fresh managerial burdens. Hosts discussed whether constant AI interaction simulated the experience of overseeing a junior team without the corresponding authority or support. The conversation circled back to boundaries. Without them the mental strain compounds. A substack analysis from Brian Elliott at Work Forward pushed further. It recommended hard limits on the number of agents any individual oversees and built-in movement breaks of just two minutes to restore working memory and attention.
The two threads of research complement each other. The academic experiment demonstrates rapid erosion of persistence after short AI exposure. The workplace survey reveals accumulated fatigue when those habits become daily routine. Together they suggest a feedback loop. Workers lean on AI to meet immediate demands. Their independent problem-solving muscles weaken. The next task feels harder, prompting heavier reliance. Over time the cycle feeds decision fatigue and error rates that offset any efficiency gains.
Earlier signals pointed in this direction. A Microsoft study on knowledge workers found steeper performance drops without AI support among heavy users. Polish research on doctors showed improved cancer detection with AI but worse accuracy once the tool disappeared. The latest experiments add causal weight and tighten the timeline. Ten minutes. That is all it took in controlled settings for many participants to shift strategies and lose willingness to struggle through hard problems.
Industry conversations on X reflected the tension this week. Some users dismissed brain fry as a skill issue, arguing that AI should serve as a tool, not a crutch. Others shared personal accounts of mental fog after long afternoons prompting chatbots and reviewing output. Indonesian accounts translated the Harvard Business Review findings for local audiences, signaling growing global awareness.
Companies face concrete choices. They can treat AI as a blunt accelerator and accept higher turnover and mistakes. Or they can design workflows that preserve human persistence. That might mean limiting simultaneous AI agents, favoring hint-style interfaces over full answers, and scheduling deliberate breaks. It could involve training that emphasizes verification skills and teaches when to turn the machine off. A few forward-looking teams already experiment with cognitive load budgets the same way they track compute resources.
The data does not argue against AI. It warns against naive deployment. Short-term boosts in output can mask longer-term losses in capability. Persistence predicts learning capacity and career resilience. Erode it at scale and organizations may find themselves surrounded by workers who excel at prompting but falter when novel challenges arise without a ready model to consult.
Researchers continue to probe. Future studies will likely test interventions. Can structured prompts that force explanation before acceptance blunt the offloading effect? Do periodic AI-free intervals restore baseline performance? The answers will matter. Because the tools are not going away. The question is whether humans retain the sharpness to steer them.
One thing feels clear. The era of treating generative AI as free cognitive labor has ended. Every implementation now carries a hidden invoice. The bill comes due in foggy afternoons, abandoned problems, and rising quit rates. Smart leaders will read it carefully before the charges compound.
WebProNews is an iEntry Publication