That’s alarming since adoption has to soar for AI firms to justify their eye-watering valuations. It’s not just shareholders who need this. Big tech firms, all of whose fortunes depend on AI, accounted in 2025 for 39% of the S&P 500’s capitalization, and for 49% of US GDP growth. A tech crash would wipe out trillions of dollars in value and very likely precipitate recessions in real economies.
The Census Bureau’s findings reinforce a McKinsey report that only about a third of enterprises have moved beyond AI pilots to meaningfully incorporate them in business functions, and less than 5% are making money from them. Research shows that public trust and optimism about AI have also fallen.
We have to ask, then, just how transformative is this technology? Three years after ChatGPT 3 exploded onto the world, the threatened jobs apocalypse hasn’t happened – unemployment is near historic lows in most rich countries. Productivity growth is flat. The chatbot on your phone seems no more likely to take over the world than your cat. Has it all been hype?
Conversation around AI has so far been dominated by “Boomers” and “Doomers”. The former argue that AI will solve everything from climate change to cancer, and the latter that AI will bring about our redundancy and eventual extinction. Both views treat AI as unprecedentled in human history. Findings like the Census Bureau’s puncture this exceptionalism. AI is behaving less like a deity and more like a utility; a run-of-the-mill technology constrained by organizational inertia and the long grind of adoption.
Glacial pace
Narayanan and Kapoor of Princeton diagnose this as a category error: we’ve confused the exponential speed of AI innovation with the glacial pace of diffusion. They draw a parallel to electrification which took 40 years to show up in productivity stats because it required factories to be physically torn up and redesigned.
There’s a case, though, that surveys like the Census Bureau’s are asking the wrong question. When pressed on adoption, tech firms point to the volume of tokens – the units of text AI models read and generate – that the world is consuming. In early 2023, reports The Economist, global usage was about 5 trillion tokens a month. Today it exceeds 100trn, a 20-fold increase. Token consumption captures the volume of work AI is doing, not just how many people touch it.
Focus on the pockets of the economy where AI adoption has been fast and it’s harder to maintain that AI is business as usual. In software development, Google reports that over 25% of its new code is machine-generated, and industry data suggests the figure is as high as 41% across some teams. In customer service, The Economist reports that fintech firm Klarna’s AI now handles two-thirds of customer chats – doing the work of 700 full-time agents – while slashing resolution times from 11 minutes to two.
AI’s most important advances may be happening largely out of the office. DeepMind’s AlphaFold2 has predicted the structure of nearly every known protein, vastly accelerating drug discovery. AI has revolutionised materials science, designing new alloys, batteries and catalysts decades faster than human researchers could. In automated labs, AIs analyse data, propose experiments, then instruct robotic systems to pipette, mix chemicals, grow cells, or run assays without human intervention. The results are fed back into the AI, which updates its predictions and designs the next round of tests. None of this shows up in the Census, but it may matter far more than any chatbot ever will.
Joe Smith is Founder of the AI consultancy 2Sigma Consultants. He studied AI at Imperial College Business School and is researching AI’s effects on cognition at Chulalongkorn University. He is author of The Optimized Marketer, a book on how to use AI to promote your business and yourself. Contact joe@2Sigmaconsultants.com.