About ten years ago, Andrew McAfee thought radiologists were in trouble. McAfee is a principal research scientist at MIT and co-directs its Initiative on the Digital Economy. He co-wrote The Second Machine Age with Erik Brynjolfsson in 2014. At that time, he said, many white-collar workers looked vulnerable because the technology "was better than they were at the thing they were getting paid to do".
In a four-guest debate on The Diary of a CEO, published on 17 September 2026, he took that back. "I want to own this," he said. "I was dead flat wrong about that."
McAfee pointed to unemployment "at historic lows" across rich countries. He said the bigger problem was that "we can't find qualified people to do the work that needs to get done, not that there's not enough work to go around." Asked whether unemployment would be higher in ten years, he said he expected employers to still be "struggling to find enough people to do the work". He said he did not expect "a massive trend break". He accepted that ten years is a long time in AI, but added that the last four years had been a long time too, and that the effect on jobs had been "essentially crickets".
His reversal shows that one well-known forecast failed over one decade. Whether the next decade follows the same pattern is what the rest of the debate, and the newest hiring data, pushed on.
A scare that came up in the episode
The host introduced the topic with a report released "the other day" by Anthropic, the company that makes the Claude AI models. The report models what AI could do to unemployment. The host compared it with current US unemployment of 4.1%. He said the report showed 11.9% overall and 17.9% for knowledge workers by 2030 in its "more extreme scenario". He suggested the pitchforks would be out without some safeguard if "one in five adults" were out of work.
Anthropic's working paper models the US economy through 2030 under three scenarios, called modest, substantial and extreme:
- Unemployment across all workers: 3.9%, 4.6% and 11.9%.
- Unemployment among cognitive workers: 2.9%, 4.5% and 17.9%. These are people whose jobs are mainly thinking and information work.
So 17.9% applies only to cognitive workers in the most extreme case. It is not a figure for all adults. The paper's numbers are conditional model outputs that depend on the assumptions put in, not a single prediction.
The model also shows that unemployment is only one way the pressure can show up. The paper treats wages as able to adjust. When wages adjust faster, more of the loss appears as lower pay. When they adjust more slowly, more of it appears as lost jobs. In the extreme scenario, the economy is 32.4% larger than it would be without AI. But cognitive workers' wages are 11.5% lower, other workers' wages are 33.6% higher, and total labour income is only 0.5% higher. The authors conclude that a richer economy does not automatically compensate the workers who are displaced.
What the study McAfee cites actually measures
McAfee called one paper "the best work about the faint signals about AI and job loss". It is Canaries in the Coal Mine, co-written by Brynjolfsson, whom McAfee described as his co-author on four books and co-founder of a company. McAfee summarised the finding this way: the effect shows up in the most AI-exposed professions, such as software engineering, and among new entrants to the workforce, "where you've got to teach them before they can become really productive." "That's exactly what we'd expect," he said.
He drew a distinction. Employers are not hiring fewer of these young people than before, he said. They are hiring fewer than they would in a world without AI. In his words, the rate of growth in employment "has slowed down", while growth in those professions overall is "still really, really healthy".
The August 2026 revision of the paper measures something related to that description, but different from it. Brynjolfsson and his co-authors Chandar and Chen used payroll records from ADP, a large payroll processing company. They followed the same set of client firms from January 2021 to June 2026, covering 3.5 to 5 million full-time workers each month. That is a sample, not all of ADP's clients. The findings:
- Employment of 22- to 25-year-olds in highly AI-exposed jobs was 19% below the path suggested by similar young workers in less-exposed jobs.
- In absolute terms, employment of 22- to 25-year-olds in the two most exposed groups of jobs fell about 11% between November 2022 and June 2026. In the three least-exposed groups, employment of the same age group grew about 10%.
- Experienced workers showed no comparable gap.
- The gap had grown since the original 2025 version of the paper.
- It came mainly from less hiring, not from more people losing their jobs. The authors say this slowdown in entry-level hiring in AI-exposed jobs has driven overall stagnation in employment growth for young workers.
- It was concentrated where AI does tasks instead of workers. Where AI assists workers, employment was flat or rising.
The authors add their own limits. They found no widespread displacement across the economy. Employment changed more clearly than base pay. The patterns weakened once education was taken into account, some trends began before generative AI, and national survey data showed weaker evidence. They describe the results as descriptive indicators, not estimates of what AI caused.
That fits McAfee's point that there has been no broad jump in unemployment. It does not match his description of young workers in exposed jobs as a case of slower growth only: in the most exposed jobs, their employment actually fell. The paper shows a gap for one group of early-career workers that has widened since its first version, and that shows up in who gets hired rather than in who gets laid off.
Capable is not the same as used
Roman Yampolskiy, a computer scientist known for work on AI safety, spoke after the host noted he had been "writing a lot of notes". He split the question into two cases. As long as AI works as a tool, he said, people become more productive and unemployment stays low. Someone can already start a company with an "artificial accountant, web designer, logo designer". Ten years out, he saw two possibilities. In one, someone builds superintelligence (AI far more capable than humans), which he said would leave the human population at zero: "Apply unemployment numbers to that." In the other, people make the "smart decision" not to build it and are left with "really cool tools" and low unemployment.
He also separated what AI can do from what people actually adopt. Video phones were invented in the 1970s, he said, but were not deployed until the iPhone. "Just because I can automate something doesn't mean I want to automate it." He declined to predict how much customers will want to be served by humans. But once a job can be automated, he said, unless he has a strong preference for a human doing it, "I'll go with the cheaper option."
Anthropic's scenario explorer builds a similar distinction into its model. It treats each job as a changing bundle of tasks. Its nursing example separates physical care from AI-assisted discharge instructions, automated charting, and new work reviewing AI recommendations. Whether AI can do a task and whether organisations use it are separate inputs. In the substantial scenario, AI can do half of knowledge work by 2030, but most knowledge-work tasks are still done without it. The model leaves out government policy responses, business cycles, financial shocks, catastrophic risks, and highly capable robots.
The horse and the car
Not everyone on the panel shared McAfee's view. Another panelist predicted that unemployment will rise, but not because of large language models (the AI systems behind chatbots). That panelist allowed that there is "probably some effect on jobs" because companies have been "shoving it everywhere".
One of the guests challenged the idea that a quiet labour market so far means a quiet future. The guest asked listeners to imagine horses watching the car improve and deciding it had only a few narrow uses and mostly complemented horses. That would have been true while the car was being developed. "And then there was a time when the car was just better than the horse." The comparison was then extended to ChatGPT: researchers had watched earlier GPT models slowly improve until one "crossed a point where it was sort of like good enough to do a bunch of people's homework", and "then suddenly it's everywhere."
Another panelist dismissed this as "another threshold argument", in which the threshold is human capability. The reply was that "you can't actually just make things not happen by assigning a name to the argument." A nuclear comparison followed. In one device, 100 neutrons produce 99, then 98, then 97, and you have "a hot rock". In another, 100 produce 101, then 102, then 103, and you have an explosive "that can level a city". Steady improvement can cross a line, the speaker said, such as the point where AI is better than humans at a job, and that could happen "in some fields, but not others."
This argument comes with no unemployment figure. "I don't know what it's going to do to employment," the speaker said. If things move fast, the speaker warned, many people could lose their jobs and be unable to relocate into new work: "It's going to be chaos." Turning to what unemployment might look like in ten years, the speaker raised a more basic worry: if "we don't stop with this AI stuff," we would be "very lucky to have 10 years."