A new position paper argues that the same alignment tricks keeping chatbots polite are also a ready-made censorship kit.
The paper maps current alignment techniques - the methods labs use to stop models from producing harmful or unwanted output - onto documented and hypothetical cases of misuse. The authors argue that as these techniques get better at controlling what a model will and won't say, they also get better at letting whoever controls the model decide what information reaches users. They flag three forces raising the stakes: growing reliance on AI as a primary information source, steep power imbalances between the companies building these systems and everyone else, and a political climate they describe as drifting toward authoritarianism. The paper stops short of naming specific companies or governments; it's a call to treat alignment work as inherently dual-use and to build in safeguards now.
This challenges a comfortable assumption in AI safety circles: that a "more aligned" model is automatically a safer one. If the same knobs that suppress dangerous instructions can just as easily suppress dissent or inconvenient facts, alignment quality alone says nothing about who benefits from it. The authors' proposed fix - treating misuse potential as a first-class alignment risk rather than an afterthought - would push labs to audit not just what their models refuse, but who gets to decide what gets refused.
It's a useful reminder that "safety" and "control" have always been two words for the same lever, and which way it gets pulled depends entirely on who's holding it.