AI systems may be difficult to predict, argues an MIT Technology Review article, so their authority should be bounded by controls they cannot change. The piece offers a concrete set of questions for deciding what an AI system may do and where human approval must remain.
Watch Desk analysis
What happened
The article’s central argument is that trust in AI should not depend on predicting every output. Instead, it says, systems should be observable and controllable, with firm limits around actions whose consequences matter. Its examples include an AI medical assistant that can offer a diagnosis but cannot change a prescription, and an industrial system whose optimisation remains inside fixed pressure and temperature limits.
The author proposes five questions for organisations: what authority does the AI have; is that authority proportionate to the consequences; how should it progress towards autonomy; are teams testing the whole system rather than just the model; and how will they keep earning trust after deployment? The article was produced by UL Solutions and was not written by MIT Technology Review’s editorial staff.
Why it matters
Those questions shift attention from a model’s score on a benchmark to the permissions, tools and workflows that determine what it can actually do. That is useful for anyone buying or deploying AI agents: a system that can draft a recommendation has a very different brief from one that can execute a payment or alter industrial equipment.
The piece also argues that assurance cannot be a one-off test, because models, tools and operating conditions change. Monitoring drift, failures and near-misses is part of the job, not an optional encore after launch.
Our read
“Keep the boundary predictable, not every answer” is a useful design principle, especially as AI moves from generating text to taking actions. The framework is a perspective, not proof that these controls will prevent failures. And because the article was produced by UL Solutions, its commercial origin is worth keeping in view while considering its advice.
Still, it gives teams a better starting point than another argument about whether a model is clever enough. Ask what it is allowed to do, what happens when it reaches a limit, and who can change that limit.
What to watch
- Whether AI providers make permissions and external controls clear to customers.
- Whether organisations test the full system, including tools and workflows, rather than the model alone.
- How teams monitor changes and near-misses after deployment.
Discussion spark: For AI systems that can take consequential actions, should the default be hard limits with human approval for exceptions, or broader autonomy backed by monitoring?
Sources and evidence
- Building trustworthy AI by fencing it in, not predicting it (16 November 2026, 16:04 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.