What If AI Is Really Good but Not That Disruptive?

If you use AI tools daily, the capability is obvious. You can draft, summarise, translate, debug, and rubber-duck a problem at two in the morning. The gap between a blank page and a working first draft has genuinely shrunk. That is real, and it is not nothing.

And yet walk into most workplaces, schools, or living rooms and very little looks different. The reason is the gap between a tool being powerful and a tool being adopted. Surveys keep landing on the same uncomfortable pattern: a lot of people have tried AI once or twice, a much smaller group use it every day, and a large slice of the population is not using it at all. Day-to-day usage has barely moved in the past year, even as the models got better and the headlines got louder.

That is not a failure of the technology. It is what diffusion actually looks like.

We have seen this film before. In 1987 the economist Robert Solow quipped that you could see the computer age everywhere except in the productivity statistics. Personal computers were visibly transforming offices, yet measured productivity growth stayed flat for years. Economists named the puzzle the productivity paradox: a powerful general-purpose technology can take a decade or more to show up in the numbers, because the technology is the easy part and rewiring how organisations work is the hard part.

AI is a textbook general-purpose technology, which means it is a strong candidate to repeat the pattern. The model is cheap. Redesigning a workflow, retraining a team, rebuilding the surrounding processes, and earning the trust to let software touch real decisions is slow, expensive, and deeply unglamorous. The capability arrives in months. The impact arrives over years.

Here is the awkward counterweight to everything above. On the narrow question of what the models can actually do, the hype has, if anything, been too timid. Benjamin Todd, a writer on AI who sits on an expert forecasting panel called LEAP, went back through the forecasts made since early 2025 and found that almost every group expected capabilities to improve more slowly than they did (his review of the forecasts).

The gaps are not small. Even the more bullish panel members expected AI to score roughly a third on FrontierMath, a benchmark of research-grade maths problems, by the end of 2026; the best systems reached 40% by the end of 2025 and 94% by September 2026. Superforecasters, the group with the best record on geopolitical questions, put OpenAI and Anthropic's combined annualised revenue at US$90bn by the end of 2026, and that mark was passed by July. In a December 2024 survey of more than 1,000 published AI researchers, the median guess for AI solving a genuinely unsolved maths problem was around 2054. Superforecasters were the most consistently too slow; AI company leaders were among the few who erred on the optimistic side, and usually only on timing.

So how does that square with an economy that still looks mostly normal? The two claims measure different things. Capability is what a model can do on a test. Impact is what happens once people, firms and regulators reorganise around it, and that is the slow part the productivity paradox describes. The same review puts the share of work tasks done with AI help at about 5.7% at the end of 2025, only a little above forecasts, and notes that most forecasters expected no large effect on GDP or employment by now. None has shown up.

Keeping the two apart is the useful habit. “It's all hype” has been a poor bet on capability: forecasters who assumed things would stay roughly where they were have been wrong again and again. “Everything changes next year” is a poor bet on impact, because diffusion has its own speed limit. The honest position holds both at once: the tools are improving faster than almost anyone predicted, and the world around them is changing at the usual, slower pace.

Boring is what durable looks like. The spreadsheet did not trend; it just quietly became the substrate of every finance department on earth. Email, search, and the smartphone followed the same arc. None of them announced themselves as a revolution on a Tuesday. They seeped in, became infrastructure, and then became invisible precisely because they were everywhere.

A technology that gets absorbed into the plumbing is far more valuable than one that burns bright and flames out. If AI is heading toward being the boring default in a thousand small tools, that is a bigger deal than any single dramatic launch, not a smaller one. The least exciting version of the AI story is also the most economically significant.

Watch what the biggest players do, not just what they say. In early 2026 OpenAI began testing advertising inside ChatGPT for free and lower-cost users, alongside a new low-cost subscription tier, while keeping its top paid plans ad-free (as NBC News reported). That is not the move of a company expecting imminent superintelligence to make economics irrelevant. It is the move of a company building a durable consumer product that needs to pay for itself the same way Google and Meta do: with ads and subscriptions.

An ad-supported chatbot is a very normal business. The pivot is a tell that even the frontier labs are quietly planning for a world where AI is a profitable utility, not a rupture in the fabric of work.

None of this means nothing changes. The most concrete shifts are happening in the plumbing, where they are easy to miss. Payment networks are starting to give AI agents their own verifiable identities so they can transact on your behalf within limits you set. Software companies are reorganising around agents that move information between systems that never used to talk to each other.

That is genuine change, but notice the shape of it. It is infrastructural, gradual, and mostly invisible to the person at the end of it. You will not wake up to a transformed world. You will, over a few years, notice that a handful of annoying tasks quietly stopped needing you. That is how real disruption usually arrives: not as an event, but as a slow accumulation of things that used to be your job and are not anymore.

The healthiest stance is neither breathless nor dismissive. Use the tools enough to form a real opinion rather than absorbing one by proximity, and keep asking what evidence would change your mind. If your honest answer is nothing, you have stopped holding a view and started defending an identity.

Practically, that means treating AI like any other capable tool. Learn where it genuinely saves you time, stay sceptical of anything it asserts with confidence, and ignore both the people promising utopia next quarter and the people insisting it is all a bubble. The truth is duller and more useful than either: a powerful tool, unevenly adopted, slowly reshaping the boring middle of how work gets done.

The most likely future for AI is not a thunderclap and not a flop. It is a slow, uneven, deeply useful absorption into the tools you already use, the same path the spreadsheet, email, and search all walked. Capability racing ahead of forecasts while impact trails behind them is not a contradiction; the two simply run on different clocks. Calling the result boring is fair. Calling it unimportant is a mistake. Boring infrastructure is the most powerful kind there is, and it is usually the part nobody notices until they cannot work without it.