career bridge

The problem, not the technology

Rules written around a model architecture are obsolete in two years. Rules written around a harm survive.

You know how instruments are built. Now the failure that produces most bad AI policy: regulating a technology rather than a behaviour. Rules written around a model architecture are obsolete within about two years. Rules written around a harm and the party who can prevent it survive.

What you do

Pick one real problem and specify it before you propose anything at all. Choose something concrete in a jurisdiction you know: automated screening in hiring, eligibility decisions for a benefit, biometric identification in public places, a ministry procuring AI. Then write four things down. Who is harmed, specifically, in the language of what happens to a person rather than the language of risk tiers. Who holds the power to prevent that harm, which is frequently not the party everyone blames. Which levers a government actually has here, because there are more than legislation: procurement conditions, licensing, disclosure duties, funding conditions, liability allocation, publishing a register. And what is already regulated, since existing consumer, employment, data protection and sector rules usually cover more than people assume, and a new instrument that duplicates them creates confusion instead of protection. Finish by writing what success would look like in measurable form, and what evidence would show the rule was working. If you cannot state that, you are writing a statement of values rather than a policy.

Done when

  • One concrete problem is named in one jurisdiction, in terms of what happens to a person.
  • The party with actual power to prevent the harm is identified, and is not assumed to be the builder.
  • At least four available levers are listed beyond passing a new law.
  • Existing rules that already cover part of the problem are named, and the overlap stated.
  • Success is defined in measurable terms, with the evidence that would demonstrate it.

What you end up with

A problem definition for one real policy question: who is harmed, who can prevent it, which levers exist, what already applies, and what success would look like.

If you get stuck

The common failure is naming the technology as the problem, which produces a rule that is either obsolete or accidentally bans something useful. Name the harm and the behaviour instead. The second failure is reaching for new legislation first because it is the most visible lever. It is usually the slowest and least precise one, and a procurement condition can move a market inside a quarter.