Two of the most influential figures in artificial intelligence spent 2023 and 2024 warning that the technology would erase a large share of office work. By 2025 they had quietly softened those warnings. The change of tone matters, because the reason for it is not spin. It is an old and well documented pattern in economics, and it has practical consequences for how any organisation plans its workforce.
What Altman and Amodei actually said, then walked back
The alarming version of the story peaked around 2023 and 2024. Sam Altman, chief executive of OpenAI, suggested that AI could replace a lot of jobs. Dario Amodei, chief executive of Anthropic, went further, warning in a podcast that AI might remove around half of all entry-level white-collar work within two to three years. The claim was widely reported and widely feared. It travelled through boardrooms, newsrooms and policy circles, and it left a lot of people anxious.
By the middle of 2025, both had pulled back. Amodei accepted that his earlier framing may have been too stark: AI would change the nature of many jobs, he said, but the long pattern of technology creating new categories of employment tends to reassert itself. Altman shifted towards describing AI as a tool that augments people rather than replacing them.
This was not simply public relations. The job market had not collapsed. In many sectors, demand for skilled workers rose as AI use spread. There were real pockets of displacement, but the wider picture looked nothing like the predicted wave of redundancies. The reason is a principle economists have understood for more than 150 years.
Jevons Paradox, the engine behind the pattern
In 1865, the British economist William Stanley Jevons noticed something odd about coal. As steam engines became far more efficient, needing less coal to do the same work, total coal use went up rather than down. Cheaper, more efficient power made it worth running far more engines, in far more places, than before. Efficiency did not reduce demand. It multiplied it.
That is Jevons Paradox, and it applies directly to AI and work.
How it applies to AI
When AI makes a task cheaper or faster, two things happen at once. The cost of each unit of output falls: writing a product description, analysing a dataset or generating a snippet of code becomes a fraction of what it used to cost. At the same time, demand for the total output rises, because it is now affordable to produce far more of it.
The result is that you need more people to manage, direct and build on top of the systems doing the cheap work. The task is automated; the workflow around it grows. A marketing team that once published two posts a week might now publish ten, but someone still has to set the strategy, edit the output, handle distribution, manage the tools and track what works. The writing got cheaper. The content operation got bigger.
It has happened before
The fear that automation will destroy jobs at scale is not new, and it has been wrong at every major technological turn.
When cash machines arrived in the 1970s, economists predicted heavy job losses among bank tellers. The opposite happened. Through the 1980s and 1990s the number of teller jobs in the United States actually rose. Cash machines made it cheaper to run a branch, cheaper branches meant more branches, and more branches meant more tellers. The cost per transaction fell while demand for banking grew.
The spreadsheet tells the same story. VisiCalc, and later Excel, automated much of what accountants did by hand. Rather than wiping out the profession, the number of accountants in the United States grew over the following decades, because cheaper analysis put it within reach of more businesses, and those businesses wanted more of it.
The commercial internet is the clearest case. It removed entire categories of work, including travel agents, classified-ad staff and encyclopaedia editors. It also created far more employment than it destroyed, in web development, digital marketing, data analysis, content and e-commerce logistics. Each of these shifts followed the logic Jevons described: cheaper capability expands total demand.
Where the new AI jobs are appearing
The same shift is already visible in hiring. Automation clearly removes some roles. The more useful question is what is growing in their place.
Most of the new demand sits around directing and supervising AI rather than competing with it. Organisations building AI-assisted processes need people who can structure prompts, chain model calls and design workflows that a model can run reliably, a role that barely existed a few years ago. Every deployed system produces output that needs human review, so quality assurance, evaluating results, spotting failure modes and feeding corrections back, is growing across industries. Training and tuning models has created large demand for data labelling. Companies also need people who can translate a business problem into something an AI system can sensibly handle, which calls for business judgement more than engineering. Even creative functions have seen demand rise for art directors and creative leads who can guide AI-generated work at scale. The common thread is that these roles exist because automation is widespread, not in spite of it.
The parts that are genuinely shrinking
This is not a story where nothing changes and everyone wins. Some work is genuinely contracting, and it tends to share a profile: routine, well defined, and judged against a low and consistent quality bar. That includes basic data entry and formatting, template-based writing such as boilerplate content or standard legal documents, scripted first-line customer support, and simple image editing.
For people whose roles are concentrated in those tasks, the displacement is real and the transition is hard. The paradox does not mean nobody is hurt. It means the overall employment picture has proved more resilient than the gloomiest forecasts, and that growth in AI-adjacent work has tended to outpace the loss of routine work over the medium term. That distinction matters for workforce planning, for training, and for individual career decisions.
The management layer few people mention
There is a less obvious driver of all this: AI systems need far more human management than the tools they replace.
A spreadsheet does not hallucinate. It does not produce confident but wrong answers, and it does not need to be watched for bias, checked against regulation or tested on edge cases. AI systems need all of that. Every serious deployment creates a layer of people who monitor output quality, handle the cases the model gets wrong, update the system as requirements change, keep it compliant, and maintain the integrations that connect it to the rest of the business.
That overhead is real and it is significant. A company that automates first-line support with an AI agent still needs operations staff, compliance reviewers and technical people to keep it working. The agent is replaced; the support operation is not. In many organisations the work of managing these systems has raised total headcount in technology and operations, even as task-level automation increased.
How to think about your own work
If Jevons Paradox holds for AI, and the evidence so far suggests it does, the useful question is not whether AI will take your job. It is how AI changes what you spend your time on.
For most knowledge workers the near-term shift is recognisable. The routine, repeatable parts of the role tend to get automated: first drafts, data formatting, standard reporting, scheduling, transcription. The parts that expand are the ones that need judgement, such as strategic decisions, quality control, managing relationships, coordinating across teams and designing how the work fits together. Some genuinely new tasks appear too, mostly around managing the tools, reviewing what they produce and helping colleagues adopt them.
For most roles the net effect is not elimination but reorganisation. The work expands to fill the capacity that automation frees up, and it shifts towards higher-judgement activity. That is not automatically better or more satisfying work. It is different work, and the people who adapt their skills to the new mix tend to fare better than those waiting for their role to stay exactly as it was.
Common questions
Is AI displacing jobs right now?
Yes, in specific categories. Routine cognitive tasks such as basic data entry, template writing, simple support and standard image editing are contracting as AI handles them more cheaply. The wider picture is more mixed. New roles in AI operations, workflow design, data quality and prompt work are growing, and demand for AI-assisted output is rising in many sectors. Economy-wide net displacement has not appeared on the scale the most alarming forecasts suggested.
What is Jevons Paradox, and why does it matter here?
It is the observation, first set out by William Stanley Jevons in 1865, that making a resource more efficient tends to increase total consumption rather than reduce it. Applied to AI: as cognitive tasks get cheaper, demand for their output grows, so more people are needed to manage and build on AI-assisted work even when fewer are needed for any single task. It is a large part of why automation has historically created more employment than it has removed over the medium term.
Why did Altman and Amodei walk back their predictions?
Their earlier forecasts, including Amodei's suggestion that AI could remove half of entry-level white-collar jobs within a few years, assumed capability gains that ran ahead of what happened in practice. As adoption grew, the job market did not collapse, new roles emerged, and the economics of Jevons Paradox reasserted themselves. Both moved towards describing AI as augmentation rather than replacement.
Which jobs are most exposed?
The most exposed jobs involve clearly defined, repetitive output judged against a consistent and fairly low quality bar: basic data entry, boilerplate documents, scripted first-line support, routine image processing and standard report formatting. Work that calls for judgement, relationships, novel problem-solving or physical dexterity in messy environments is far more resilient.
How can workers adapt?
The safest position is to understand which parts of your role are routine and automatable, and to move your time deliberately towards higher-judgement work. Being able to build and manage simple AI workflows, even without a technical background, is becoming a basic competence. The people who do best are not those who resist automation but those who get good at directing it.
Key takeaways
- More automation has tended to create more work, not less, driven by the same logic as Jevons Paradox.
- Both Sam Altman and Dario Amodei have softened their earlier predictions of large-scale job loss, as the employment picture proved more resilient than the gloomiest models.
- The mechanism is simple: cheaper tasks raise demand for output, so the workflow grows even as the cost of each task falls.
- Routine, well-defined work is genuinely shrinking, while higher-judgement and AI-management roles are growing.
- Managing AI systems, through quality review, compliance and operations, creates real demand that offsets much of the task-level automation.
- Adapting means understanding what AI automates in your role and building the skills to direct and manage it.