Two of the most influential people in artificial intelligence spent a couple of years warning that the technology would wipe out a large share of office jobs. Then, fairly quietly, they changed their tune. The reversal is worth understanding, because the reason behind it is not a public relations exercise. It is an old and well documented pattern in economics, and it has direct consequences for how any organisation should plan its next few years.
The prediction that did not hold
Around 2023 and 2024, the warnings were stark. Sam Altman of OpenAI suggested that AI could take over a great many jobs. Dario Amodei of Anthropic went further and put a number on it: he said the technology might remove as much as half of entry-level white-collar work within a few years. The claim travelled fast. It was repeated in boardrooms, in newspapers and in policy debates, and it left a lot of people genuinely worried.
By the middle of 2025, both men had softened their position. Amodei accepted that his earlier framing had been too blunt. AI would change the shape of many jobs, he said, but the long-running pattern of new technology creating fresh categories of work tends to reassert itself. Altman shifted his emphasis too, talking more about AI as something that supports people rather than replaces them.
This was not simply a matter of tidying up the message. The job market had not collapsed. In several sectors, demand for skilled workers actually rose as AI use spread. There were real pockets of displacement, but the wider picture looked nothing like the predicted wave of redundancies. To see why, it helps to go back to the nineteenth century.
Jevons and the coal that would not run out
In 1865, the British economist William Stanley Jevons noticed something odd about coal. Steam engines were becoming far more efficient, which meant they needed less coal to do the same amount of work. The sensible expectation was that total coal use would fall. Instead it climbed.
The reason was straightforward once he spelled it out. Cheaper, more efficient steam made it worthwhile to run engines in far more places than before. Factories that could not previously justify the cost now could. Efficiency did not shrink demand. It widened it. The saving per unit was real, but it was swamped by the sheer growth in how much the technology got used.
This is now known as Jevons Paradox, and it maps neatly onto what is happening with AI. When a task becomes cheaper to perform, organisations rarely bank the saving and stop there. They do more of the task, and they start doing things that were never economical before.
More capacity, more to do
Writing software is a good example. As AI tools made it quicker to produce code, the obvious worry was that companies would need fewer developers. What many found instead was that cheaper code led them to build more software. Backlogs that had sat untouched for years suddenly became affordable to clear. Internal tools that nobody could justify before were now worth the effort. The work did not disappear. It expanded to fill the new capacity, and much of it still needed people to specify, review and maintain it.
The same logic shows up across other functions. When analysis gets cheaper, organisations analyse more. When drafting gets cheaper, they produce more drafts and spend the saved time on judgement and decisions. The cost of one unit falls, the appetite for the underlying activity grows, and the net effect on headcount is far less obvious than the early predictions assumed.
None of this means the change is painless. Jevons Paradox describes the aggregate, not the individual. The total amount of work can grow while specific roles still shrink or shift. Someone whose job was a narrow, repeatable task can be displaced even as overall demand rises. The shape of work changes faster than the volume of it falls.
What this means for planning
For anyone running a team or a department, the practical lesson is to be wary of both extremes. The story that AI will quietly remove half your staff has not held up. Neither has the idea that nothing will change. The realistic middle is that the same number of people, or more, end up doing different work, with the routine parts handed off and the harder parts left firmly with humans.
That points to a few sensible moves. Treat AI as a way to raise what your existing people can take on, rather than as a headcount-reduction exercise. Expect demand for work to rise as it gets cheaper to deliver, and plan capacity with that in mind. And pay close attention to the parts of a job that do not automate well, because that is where the value, and most of the employment, will sit.
The people who got the prediction wrong were not careless. They reasoned from the efficiency of the tool and stopped there. The missing piece was demand. Once a capability gets cheap enough, people find far more uses for it than anyone expected, and the work that surrounds it grows rather than shrinks. That has been true since the age of steam, and so far it is holding for AI.