Even a Slowed-Down AI May Trigger Severe Economic Disruption

How ironic then that just a day later, Dario Amodei, boss of Anthropic, urged developers to slow advances in the AI industry’s frontier models. One doubts he was responding to the cross-party group of 40 MPs who had asked Andy Burnham to pursue an international agreement prohibiting “superintelligent” AI. Yet fears of model capabilities outpacing safety precautions have raised the prospect of a deliberate slowdown, with leading companies themselves now urging government-backed regulation.

It would be a mistake, however, to think such “paced” development means AI’s near-term economic effects on Britain must then be small. A recent Anthropic Institute working paper shows that AI’s growth-boosting potential isn’t dependent on ever-improving capabilities. It also depends on business adoption and the pace at which capital and workers adjust to existing technologies.

There is plenty of low-hanging fruit here for the next four years. Recent government analysis found that only about one in five British firms used or planned to use AI. Even within adopting businesses, fewer than a third of employees used it. There is thus ample scope for efficiency gains as, say, law firms buy AI licences for all solicitors rather than allowing a handful of AI enthusiasts to experiment.

Indeed, even last year’s AI models were hardly primitive. OpenAI’s GDPval benchmark examines 1,320 realistic work assignments devised by experienced professionals. The tasks required about seven hours of expert effort, on average. On its 220-task test subset, Claude Opus 4.1 produced deliverables judged as good as, or better than, the human version in almost half of comparisons. There is substantial capability for productivity improvements already.

Anthropic’s new model allows you to tinker with assumptions about how AI capabilities, adoption and automation will affect the US economy through 2030. Their results are startling. Suppose capabilities remained at today’s levels, but use of AI for these tasks grows from 10 per cent to 40 per cent by 2030, with about half fully automated. In that case, GDP would be 3.4 per cent higher.

In its “substantial change” scenario, where AI could theoretically do almost half of knowledge work, usage quadrupled, and 75 per cent of that use is automated, GDP would be a huge 8.3 per cent higher, with knowledge-work employment falling by 4 per cent. The institute’s extreme case has AI capable of doing 80 per cent of knowledge work and ultimately performing almost half of it within four years, overwhelmingly autonomously. GDP then would be almost a third higher, with knowledge-worker unemployment hitting 18 per cent.

There are thus scenarios where countries get substantially richer but it feels like a major recession for its professional class. And Britain looks sensitive to these effects. One government estimate suggests that 67 per cent of British workers have jobs highly exposed to AI. Finance, law, accountancy, consulting and administration services rank particularly high.

There are, of course, less-exposed occupations, including roofers, plasterers and steel erectors. Yet a displaced accountant cannot generally retrain as an electrician as quickly as his firm can deploy new AI software. Nor are Britain’s planning system, expensive housing and training institutions renowned for helping workers move swiftly to expanding occupations and places. And even where workers aren’t replaced outright, automation can still mean fewer junior hires, thinner departments and weaker wage growth.

Now these modelled results are scenarios, not forecasts. The results do demonstrate, however, that a slowdown in frontier AI capabilities doesn’t rule out substantive economic disruption very soon. Anthropic’s model highlights a politically explosive potential future, with Britain growing much richer but many graduates concluding that the technology is coming for their livelihoods.