A polished answer can conceal a missing ability. That is a problem for a school assessing learning, and for a business assessing whether a new employee can handle an unfamiliar situation. AI makes the distinction between producing an answer and understanding it harder to ignore.

What happened

On 2 September 2026, New York City announced a one-year moratorium on student-facing generative AI for pupils from 2-K through eighth grade during the 2026-27 school year. The announcement covered nearly 600,000 children. High schools were assigned a different approach involving limited pilots and critical-thinking work. This was a defined education policy, not a permanent prohibition on every form of AI. The city’s announcement sets out its scope.

That decision does not, by itself, prove that AI damages learning. It does make a practical tension visible: institutions must decide when assistance supports development and when it removes the practice they are trying to provide.

Why it matters

Employers face a related design problem when they introduce AI into junior roles. If a system drafts every explanation, compares every option and identifies every exception, where does a newcomer learn to do those things?

The answer need not be a blanket workplace ban. I would separate learning activities from delivery activities. A person might first explain a customer problem unaided, then compare their reasoning with an AI suggestion, and finally prepare a checked response using approved tools. Each stage serves a different purpose.

The manager should assess more than the final response. Can the employee explain the decision, identify an assumption and recognise when the available information is insufficient? Those questions reveal something that a fluent paragraph cannot.

This requires time from experienced colleagues. Buying an assistant while removing mentoring capacity may undermine the very capability the organisation expects people to use when the assistant is wrong. Training budgets should therefore include guided practice and feedback, not simply access and an introductory demonstration.

The bigger shift

My interpretation is that AI adoption makes the purpose of a task more important. The same activity can be either work to complete or practice through which someone learns.

Drafting a routine note for an experienced manager may be a sensible use of assistance. Asking a trainee to draft that note independently may be useful because it exposes whether they understand the customer, the evidence and the consequence of an unclear promise. Neither decision follows from the task label alone.

Organisations can make this explicit. For each developing role, identify a few abilities people must retain without assistance, the tasks where assistance is encouraged and the situations that require an experienced reviewer. Review the arrangement as competence improves.

This is a governance decision because it affects accountability and resilience, as discussed in the CEO’s guide to AI governance. A company needs people capable of questioning a system, not merely operating its interface.

My take

I would resist turning an education policy into a universal rule for business. Children learning foundational skills and experienced professionals completing routine administration have different needs.

The useful challenge is more specific: make sure the organisation can explain what people are supposed to learn while AI helps them work.

Choose one junior workflow this month. Observe the reasoning it normally develops, decide which part still needs independent practice and give a mentor responsibility for reviewing that practice. A successful AI programme should make useful work easier while keeping the path to expertise visible.