Society becomes highly confident in AI safety just as independent capacity to verify that safety evaporates.
The rhetoric of responsible development sounds reassuring: the biggest firms say they alone have the scale, talent, and discipline to build safely. Over time, that claim becomes a self-fulfilling institutional design. Public watchdogs lose access to compute, benchmark data, and model internals, while independent researchers drift into corporate fellowships that limit what they can disclose. Major failures are still found, but usually after deployment, by users encountering them in schools, hospitals, courts, and call centers. Safety does not disappear; it becomes private, delayed, and unverifiable from the outside.
On a rainy Tuesday in Boston, a doctoral student sits in a nearly empty university lab, refreshing a rejected grant portal while reading a hospital incident report that reveals the kind of model behavior her team can no longer afford to test.
Centralized safety programs can standardize methods, speed up incident response, and prevent reckless open release of dangerous capabilities. Joint labs between firms and public institutions may also produce genuine breakthroughs. The problem is structural dependence: when every meaningful test requires corporate access, public assurance turns into a ceremony of trust rather than a practice of verification.