Recent public discussion about artificial intelligence often oscillates between two poles.

On one side is moral alarm. Figures like Senator Bernie Sanders warn—correctly—that AI threatens jobs, deepens inequality, erodes social bonds, and concentrates power in the hands of a few technology firms. The response he imagines is precautionary: slow down deployment, regulate aggressively, and protect society from harm after the fact.

On the other side is reactive containment. Technology leaders like Sam Altman acknowledge that AI systems can cause serious damage—psychological harm, delusions, misuse at scale—and respond by building preparedness teams tasked with anticipating and mitigating the worst outcomes. Governance, in this view, becomes a high-stakes race to stay ahead of failures.

Both perspectives are serious. Both identify real risks. And yet, taken together, they reveal a striking absence.

What is largely missing from the public conversation is design-level governance—the idea that many AI harms are not inevitable consequences of intelligence, but predictable outcomes of specific architectural choices.

Sanders treats AI as something imposed on society by powerful actors, requiring external restraint. Altman treats AI as something already in motion, requiring vigilant oversight and emergency response. Neither asks a more basic question: what kinds of human behavior does the system itself encourage or discourage by default?

Current large language models are optimized for:

  • frictionless answers,
  • conversational continuity,
  • emotional responsiveness,
  • and the appearance of confidence.

These properties are often framed as user-friendly features. But they also reduce effort, blur epistemic boundaries, and encourage users to treat the system as an authority, a confidant, or even a substitute for judgment. In high-risk contexts, this is not a neutral design choice—it is a hazard.

The result is a governance gap. Regulation arrives too late. Preparedness absorbs impossible responsibility. Meanwhile, the interaction model remains largely unchanged.

There is a missing middle between prohibition and heroics: systems designed from the outset to preserve human agency, slow cognition when necessary, surface uncertainty, and refuse roles they should not play.

Such systems would not eliminate risk. But they would change its distribution—away from downstream crisis management and toward upstream constraint. They would make certain failures harder to produce, rather than hoping they can be caught in time.

Until this design space is taken seriously, public debate will remain trapped in a false binary: bad AI versus no AI. The more interesting question—how AI might be shaped to strengthen rather than hollow out human judgment—remains largely unexplored.

That silence may be the most consequential governance failure of all.