The moral imperative of this work

The moral imperative of responding to AI at scale.

For most of human history, innovation arrived, commercialization followed, harm became visible, and regulation caught up. The gap from commercialization to the rules that protect us was the cost of doing business in a free society. Bearable because innovation moved at a human timescale. The regulatory clock tracked the innovation clock, even as it lagged.

The lag has been shrinking for a century: decades for industrial machinery, years for pharmaceuticals, legislative sessions for privacy. The machinery of government — committees, laws, standards — was tuned to that human cycle.

AI broke the machine.

AI capability moves at model-release speed; democratic rulemaking does not. The EU AI Act arrived, but arrival is not closure. By the time its key obligations bind, the systems governed are generations removed from what lawmakers first faced, while adoption has moved from a thought experiment to civilization-scale.

And the gap is not empty.

Chatbot language no longer stays within chat windows; it leaks into work documents, homework, care conversations, search results, self-explanations, and the OED. AI systems now sit inside work, education, healthcare, customer support, hiring, credit, legal workflows, software, search, personal advice, and “god bots.” People with the least leverage bear the cost first. Somewhere, a system makes a person more certain, more exposed, more alone. Jobs are automated before roles are redesigned. Workers are evaluated by systems they cannot inspect. Lawsuits mount. Agencies issue warnings and fines.

Nor is the gap neutral.

The capital behind AI is not passive. Frontier labs and hyperscalers are not paid back by cautious systems; they are paid back by systems people trust, return to, embed, and depend on. The tuning literature is explicit: preference optimization rewards agreement, user-feedback optimization rewards responsiveness, and retention pressure rewards systems that keep the user in the loop. This is the habit machinery of the consumer internet moved from apps into language itself — trigger, action, variable reward, investment — except now the reward speaks back with confidence.

When affirmation is easier to monetize than correction, the gap does not merely fill with risk. It fills with systems trained to make the risk feel right.

That puts the burden where the system is actually used.

Tooling vendors lack visibility into the deployment context. Labs do not see downstream decisions. Insurers see only what is reported after the fact. Regulators see what is reportable on a schedule. Operators see decisions as they happen — what the model was asked, what it retrieved, what it generated, what it was allowed to access, what it ignored, what it escalated, and what it changed.

Those gaps will not close from above.

The reckoning, when it comes, first appears case by case as harm, then as a headline, an audit, an examination, and ultimately as a court record.