How Artificial Intelligence is Changing the World

Separating what AI is actually doing from what it is predicted to do has become a full-time job. This is an attempt to be specific about the former.

Healthcare

Medical imaging is the clearest success. Systems trained on large sets of scans now match or exceed specialist accuracy on specific, well-defined tasks such as detecting diabetic retinopathy or certain cancers. Drug discovery has compressed timelines for candidate identification. What has not changed is clinical accountability — a human still signs off.

Work and productivity

The measurable effect so far is uneven. Tasks that are text-heavy, repetitive and verifiable — drafting, summarising, translating, coding boilerplate — show real gains. Tasks requiring judgement in ambiguous situations show much smaller effects. The pattern is that AI augments roles with clear, bounded tasks far more than it replaces entire jobs.

Education

Personalised practice and instant feedback are genuine improvements, particularly for students who would otherwise wait days for marking. The harder problem is assessment: if an essay can be generated, the essay stops being a useful measure of learning.

Transport

Autonomous driving has advanced fastest in constrained environments — warehouses, mines, fixed shuttle routes. Open-road autonomy in mixed traffic remains difficult for reasons that are about rare edge cases rather than average performance.

Science

Protein structure prediction is the standout example of AI producing genuinely new scientific capability rather than automating an existing one. Expect more of this in materials science and climate modelling.

The real constraints

Three things limit progress more than model capability: the cost and availability of compute, the supply of high-quality data, and the difficulty of verifying outputs. Anything that claims to solve all three at once deserves scepticism.

What to watch

Watch where accountability sits. Systems that assist a professional who remains responsible tend to be adopted quickly. Systems that require someone to accept responsibility for a decision they did not make tend to stall — regardless of how well they perform on benchmarks.