How fast it arrived
ChatGPT reached one hundred million monthly users in two months. Facebook took four and a half years, Instagram two and a half, TikTok nine months.
Stanford's 2025 AI Index recorded global private investment in AI at $252 billion in 2024, a jump of more than 40% on the year before. McKinsey's 2024 State of AI survey, run across thousands of organisations, found that 72% now use AI in at least one business function, up from 55% the year before.
The International Monetary Fund estimated in January 2024 that around 40% of jobs globally are exposed to AI-driven change, rising to roughly 60% in advanced economies. "Exposed" does not mean "replaced"; it means the work inside those jobs will look meaningfully different.
Routine work: AI as leveller
Erik Brynjolfsson and colleagues at Stanford and MIT studied more than five thousand customer-support agents using a generative AI assistant alongside their normal work (Quarterly Journal of Economics, 2025). Across the group, productivity rose by about 14%. The gains were not evenly distributed. The newest, least-experienced agents saw their productivity rise by roughly a third. The most experienced agents saw essentially no change.
In routine, scripted work, AI raises the floor. A weaker writer, a less-confident maths student, a junior employee, a worker in a second language: all of them now have access to explanation and feedback at a level unimaginable a decade ago.
Judgement work: AI as amplifier
Where judgement is required, the picture inverts. A 2023 field experiment by Fabrizio Dell'Acqua, Ethan Mollick and colleagues at Harvard Business School worked with consultants at the Boston Consulting Group. Half were given access to GPT-4, half were not. On tasks the AI was suited to, the AI-using group produced work that was about 40% higher quality and 25% faster. On tasks just outside the AI's capabilities, the AI-using group did worse than the group without it. They had trusted the machine where they should have pushed back.
The fundamentals AI amplifies are the familiar ones. The pupil who reads carefully can interrogate a long answer and notice a contradiction; the pupil who skims is skimmed back at. The pupil with real numerical sense can spot a wrong calculation; the pupil without it accepts the number on the screen. The pupil with knowledge of a field asks better questions and recognises better answers; the pupil without it cannot tell brilliance from confident nonsense, and AI produces both. The cognitive parts of most white-collar work are being handed off to AI, and the human role moves up the stack: less typing the report, more deciding what report is needed, briefing the AI, judging whether the draft is correct, and taking responsibility for it.
Learning: where the thinking happens
Every solid lesson has roughly the same shape. Something is taught: the teacher explains a concept, sets it in context, links it to what the pupil already knows. Then the pupil thinks: recalls the idea, applies it to a new problem, gets something wrong, tries again. Finally the work is checked and corrected. The middle stage is where learning is built.
AI is fluent at the first stage and useful at the third. It can also do the middle stage, and that is where learning evaporates. Two 2025 studies have begun to measure this directly.
The MIT Media Lab put fifty-four adults into three groups: write an essay with ChatGPT, with Google Search, or with no tools at all. Neural activity was measured with EEG. The ChatGPT group showed the lowest engagement of the three. When the same group was later asked to write again without AI, they performed worse than the others, and most could not accurately quote even a single sentence from their own earlier essays.
A separate 2025 study by Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about their use of generative AI. The more confidence workers had in the AI's output, the less critical thinking they reported applying. The more confidence they had in their own ability, the more they pushed back on what the AI produced. The workers least equipped to judge what the machine returned were also the most likely to accept it.
Both findings point at the same mechanism. When a tool produces a fluent, plausible answer, the brain stops doing the work that turns information into understanding. Producing the work and learning from producing it are not the same thing.
Three years of evidence point one way. AI raises the floor in routine work and the ceiling in judgement work. Subject knowledge, deep reading, numerical reasoning, and the habit of pushing through hard problems are the fundamentals being multiplied.