In a university library in the American Midwest, a student stares at a blank document, then types a prompt into ChatGPT: “Write a 2,000-word essay on the causes of the First World War.” Within seconds, a structured argument appears. The student copies the text, adjusts a few phrases, and submits. The glow of the screen fades, and with it, the struggle that once forged understanding. This scene, repeated in dorm rooms and coffee shops from Manchester to Melbourne, is not an outlier. It is the new normal of intellectual life, and it has become the subject of a quiet but urgent scientific inquiry.

A British-American study of 1,222 people, still under peer review, found that using AI tools for arithmetic and reading comprehension boosted immediate performance but eroded long-term results and, crucially, the willingness to persist without the tool. “These findings are particularly concerning because persistence is foundational to skill acquisition,” the authors wrote. Grace Liu, a doctoral student at Carnegie Mellon University and the study’s lead author, notes that AI’s generality sets it apart from a calculator: it can be applied to “pretty much any intellectual, reasoning, cognitive activity,” removing learning opportunities at every turn. A 2025 MIT study that went viral suggested students using generative AI to write essays displayed less critical thinking. In France, Johann Chevalère, a researcher at the CNRS, describes a phenomenon of “cognitive offloading”—even “cognitive surrender.” “Human beings have a strong tendency to save energy,” he says. “If there are activities you never do, the brain won’t go to the trouble of maintaining connections that aren’t being used.”

While cognitive scientists measure the slow unravelling of human persistence, a parallel crisis is unfolding in boardrooms and government ministries: the erosion of trust. In Australia, a survey of 48,000 people across 47 countries found that four in five Australians are “very concerned” about the societal risks of AI, and 70 per cent believe current regulation is insufficient. Dr Nicole Gillespie of the University of Melbourne points to AI literacy as the key driver of trust. “We know it’s a lot harder to trust AI systems when you don’t understand them,” she says. Andrew Hinchliff, group chief risk officer at Commonwealth Bank, frames the issue in stark terms: “No trust, no scalability.” Without trust, he warns, Australia risks failing to capture the economic value AI can deliver. In the UAE, where the government has announced that half of all public services will transition to autonomous AI within two years, the conversation has shifted from adoption to resilience. The hard part of AI, analysts in the Gulf note, is no longer access to models but whether they can operate inside real business conditions—with trusted data, accountable reasoning, and the flexibility to adapt under pressure.

In Indonesia, the trust deficit takes a more visceral form. Deepfake technology, powered by the same generative AI that writes student essays, is fuelling what experts call an “infocalypse.” Nina Schick, a geopolitical analyst, describes a world where video, audio, and images can be manipulated so perfectly that people begin to reject genuine evidence. The “liar’s dividend,” a concept from legal scholars Bobby Chesney and Danielle Citron, allows public figures caught in real scandals to dismiss authentic recordings as AI fabrications. The result, as one Indonesian analysis puts it, is not just that people believe falsehoods, but that they stop believing in any reality at all. This corrosion of shared truth runs parallel to the cognitive offloading documented in Western universities: both are symptoms of a society that is outsourcing not just tasks, but judgment itself.

Faced with these twin challenges, technology companies are building what they call “Socratic” safeguards. OpenAI’s ChatGPT now offers a “study mode” that provides hints and questions rather than answers. Google’s Gemini has “guided learning.” Microsoft’s Copilot warns users about the risk of errors and reminds them to verify information. Yet the researchers interviewed across continents agree that large-scale, long-term studies are still missing. Until they arrive, the burden falls on the user. “It’s up to us to use AI in a smart way,” Chevalère says. In that American library, the student closes the laptop. The essay is done, but the question lingers: what was not learned?