Your AI Use Policy Treats Everyone The Same

New York City removed generative AI from 600,000 children yesterday and left it running for their teachers. One school system, one decision, two opposite answers, and that is what a working AI use policy actually looks like.

The city announced a one-year moratorium on student-facing generative AI for 2-K through eighth grade, effective with the 2026-27 school year. It covers roughly two-thirds of enrolment. All software using student-facing generative AI goes, and companion chatbots are barred across every grade, high schools included.

High school is where it gets interesting. Rather than a blanket exemption, the city set up five pilots reaching around 50,000 general education students, capped at five classes per school, each session supervised by a trained educator using vetted tools. One maths tool is limited to twenty minutes per week. Alongside that, twice-yearly critical thinking modules on AI itself.

Teachers keep using the technology for planning and administration throughout.

Your AI use policy assumes exposure is uniform

Almost every AI use policy I see is written at the level of the organisation. We permit these tools. We prohibit those. Staff may use approved systems for approved purposes.

That framing has a hidden assumption inside it, namely that exposure is uniform across the people it governs. It rarely is.

New York’s reasoning makes the assumption visible. An eight year old using a chatbot to finish a worksheet and a teacher using one to draft a lesson plan are not running the same risk, because they are not doing the same thing. The child is supposed to be building the capability the tool replaces. The teacher already has it.

So the same software, in the same building, on the same network, produces opposite answers depending on who is in front of it.

Where blanket policies break

A single organisational rule has to resolve that tension somehow, and it usually resolves it badly in one of two directions.

Permissive policies extend the senior case to everyone. The reasoning runs that experienced staff use these tools well, therefore the tools are fine, therefore everyone may use them. That quietly removes the practice that produces experienced staff in the first place.

Restrictive policies extend the junior case upward. Nobody may use it, because some uses are risky. Then the capable people route around the rule, and the organisation acquires shadow adoption it cannot see or govern.

Both failures come from the same error. The policy picked one population and wrote a rule for the whole organisation from it.

Three questions that segment an AI use policy properly

New York effectively asked three things. They transfer to any organisation.

• Who is exposed, and what are they meant to be learning? The answer differs sharply between someone building judgement and someone applying it. Where the task is how a person becomes competent, automating it has a cost that shows up years later.
• What supervision actually exists at the point of use? Not policy supervision. Somebody present, trained, who sees the output and can intervene. New York’s pilots put a trained educator in the room rather than a clause in a handbook.
• What is the cap, and what happens when it expires? Twenty minutes a week is a cap. Five classes per school is a cap. A named review date is a cap. Caps make evaluation possible, because unbounded pilots never produce a decision.

What this looks like away from a classroom

Translate the same logic into an ordinary organisation and the segments appear quickly.

Customer-facing deployment carries transparency duties and reputational exposure that internal drafting does not. A junior analyst learning to build a financial model sits in a different position from the director reviewing it. Regulated advice differs from marketing copy. A tool used to summarise a document differs from one used to decide who gets contacted.

An AI use policy that cannot distinguish between those cases is not governing them. It is covering them.

This connects to something scaling AI surfaces repeatedly. The governance load does not arrive with the pilot. It arrives when the same tool spreads across populations with very different exposure, under a rule written for whichever population happened to go first.

The literacy piece is not separate

Notice what New York paired with the restriction. High schoolers get structured instruction on thinking critically about AI, twice a year, before they start relying on it.

That is the same instinct behind the AI literacy duty in Article 4. Restricting a tool and building the capability to use it well are not alternatives. Organisations that only do the first end up with staff who avoid the technology without understanding it, which is its own governance problem when the restriction eventually lifts.

The part that makes it credible

Announcing a pause is easy. Building the conditions for the pause to end is not.

New York attached a study period, named the review point, capped the pilots so they generate comparable evidence, and kept a live channel through teacher use so the institution does not fall a year behind. The moratorium may extend, and officials have said so. It may not.

Either way, the decision in twelve months gets made against something. That is unusual, and it is the part worth copying regardless of what you think of the underlying call.

The question for the AI lead

Take your current AI use policy and ask which populations it actually distinguishes between. Not job titles. Exposure.

If the answer is that everyone gets the same paragraph, the policy is making a claim your organisation probably cannot defend, which is that a first-year hire and a department head carry identical risk when they open the same tool.

Nobody has to agree with New York’s judgement to use its method. Segment by who is exposed, put real supervision where the exposure is highest, cap what you are unsure about, and set a date when you look again.