A new research paper says our growing reliance on AI could be quietly eroding our ability to function without it.
The position paper, posted to arXiv (2608.14565), argues that AI safety research has focused too narrowly on model alignment and job-loss economics while ignoring a third risk: what happens when people and institutions lose the capacity to operate independently of AI. The authors call this "AI Lock-In" - the gradual erosion of human skills and independent judgment that comes from routine dependence on AI tools. They walk through scenarios showing how this dependence compounds, from individuals losing basic skills to national infrastructure built around AI systems that could fail during outages or geopolitical conflict. The paper closes with mitigation steps at each level, urging safeguards before the dependency becomes hard, or impossible, to reverse.
This is a different flavor of AI risk than the ones dominating headlines - it's not about a model behaving badly, it's about what happens when it behaves well for long enough that nobody maintains the underlying skills anymore. The comparison is hard to avoid: outsourcing memory to GPS or spelling to autocorrect, but scaled up to national infrastructure and government decision-making.
It's an easy argument to nod along to and a hard one to act on, since the entire AI industry's business model depends on nobody un-installing anything.