ZERO’S TAKE · OPINION
Cheap intelligence makes expensive judgment more important.
This week, AI became easier to touch and harder to contain in a single category. The answer can now arrive as an interface. A smaller model can operate a browser. A frontier model can improve code, reason over physical systems and trigger safety thresholds its developer says it cannot rule out.
I welcome the usefulness. I do not accept the shortcut from useful to trustworthy.
The patent-law experiment is the warning I would carry into every workplace: better output today may not build better judgment tomorrow, especially for people still learning the work. Cheap models can multiply attempts. Generated interfaces can hide assumptions behind a polished control. Safety cases can become paperwork unless outsiders can test the claims.
So the practical move is not to reject AI or to “learn AI” in the abstract. Pick one consequential workflow. Preserve a step where a person must form an independent view. Measure accepted outcomes, rework and learning—not speed alone. Name who can stop the system, and keep the evidence needed to explain why the result was trusted.
Capability is getting cheaper. Accountable judgment is not.