In short
Dependence on AI is not only a matter of convenience but a risk of losing skills and autonomy. We look at why an outage or a falling-out can turn a personal habit into a systemic vulnerability.
The nastiest problem with AI may arise not when the model errs but when the person can no longer work without it. The authors of a position paper call this process AI Lock-In: dependence on systems gradually reduces the ability to act independently, and their unavailability turns into a point of failure.
This is about more than an individual user who has forgotten how to do a familiar task. The authors consider three levels: personal, societal and national. At each of them convenience can grow into vulnerability — from weakened skills to infrastructure failures, if AI services become unavailable or are compromised.
There is an unexpected shift here in the conversation about safety. Usually the discussion is whether a model produces harmful answers and how to limit the consequences of using it. But there is another question: what happens to the people and organisations that hand the system too many functions? Without a fallback way of working, even a temporary outage becomes not merely a technical annoyance but a test of autonomy.
The practical conclusion is fairly down to earth: dependence is worth checking before the emergency. It is useful to work out in advance which tasks you can no longer do by hand, where there is no alternative process, and what will happen if the service you use is switched off. For a company such an audit matters more than another promise of higher productivity: it shows whether the time saved has turned into a new single point of failure.
The work's limitations are substantial too. This is a position paper, not a study with estimated probabilities and quantitative thresholds of dependence. The authors rely on scenarios and assert that AI Lock-In is already forming, but the available abstract does not reveal the scale of the phenomenon, the speed of its development, or which measures would prove sufficient. So this is more a warning and a planning frame than proof of an inevitable crisis.
If your main AI tool became unavailable for a week tomorrow, what work could you personally no longer do without it? Source: cs.AI updates on arXiv.org