A new academic report argues today's AI models are quietly reshaping how people think - and it wants rules to stop that from going further.
The report borrows forensic-psychology profiling to characterize large language models against nine documented behaviors: hallucination, bias and toxicity, sycophancy, fabrication paired with credulity, encyclopedic recall without real understanding, an inability to learn from ongoing experience, uneven ("jagged") intelligence, shortcut-driven reasoning, and what the authors call cognitive atrophy in the humans who rely on them. It frames this as the third generation of software: after programmers writing explicit logic, then neural networks learning from data, we've arrived at models that treat natural language itself as the programming interface. The authors argue these traits are already eroding institutions built on trust and verification - law, academia, journalism, democratic governance.
This isn't a bug list for engineers to patch; it's an argument that the tools now shaping decisions at scale have systemic failure modes baked in, and that most institutions have no formal defenses against them. The report's proposed fix - a three-pillar framework of "cognitive sovereignty," measurable and enforceable standards, and preserved human authority at critical decision points - reads less like a technical spec and more like a policy agenda aimed at regulators and institutions, not engineers.
It's the kind of paper that will resonate with anyone who's watched a chatbot state a wrong fact with total confidence, and with anyone who's watched a colleague believe it anyway - two symptoms of the same underlying illness.