Are graduates facing an AI ‘jobpocalypse’? The evidence is unclear

Commencement speeches by VIPs to newly graduating students are usually met with respectful applause. But in recent weeks a series of speakers at American universities have been roundly booed – the latest being former Google CEO Eric Schmidt. Their crime? Talking about the benefits of artifcial intelligence to roomfuls of graduates terrified about their future.

It’s not just students who are worried. On both sides of the Atlantic, the public is turning against AI. About 57% of people here think it will lead to widespread unemployment, while just 17% think it will create more jobs. And 69% are worried about the impact of AI on their own economic future.

AI companies have only themselves to blame, having decided to promote their products by warning of the immense social harm they will cause. Anthropic CEO Dario Amodei has predicted half of all entry-level graduate jobs will be wiped out within four years. His rival, Sam Altman, who set up OpenAI, thinks 40% of all jobs globally will be automated over the same timescale. As PR strategies go, it’s questionable.

The only people who seem to disagree with this analysis are those doing the hiring. Four-fifths of employers surveyed by the Institute of Student Employers say they are anticipating minimal job losses as a result of AI over the next three years. They expect significant changes to entry-level jobs, as more mundane tasks are automated, but not much of a reduction in the overall numbers of employees.

So who’s right? To date the evidence is mixed. Some studies have shown a drop-off in hiring for those sectors most exposed to AI, such as software development, but others have found no discernible pattern, or even an increase in jobs. The lack of consensus suggests that any shifts have been fairly minor so far. Some companies, such as Amazon and Standard Chartered have announced significant job cuts associated with more automation but others have ended up rehiring staff after struggling to implement AI-based alternatives. There’s no doubt there are fewer graduate jobs around, but that might just be down to broader economic troubles. There are also fewer non-graduate jobs in roles that aren’t exposed to AI.

One key question disputed by economists is how elastic demand is for some of these most exposed sectors. Productivity growth leads to job losses only when demand is satiated. So improvements in making cars and TVs have led to a big fall in manufacturing jobs over the past few decades. There’s a fixed limit to how many of these products we need. But the same doesn’t necessarily apply for, say, software development if companies are willing to pay for more activity as the price falls. Healthcare is an obvious example of a sector with highly elastic demand. If AI diagnostic and admin tools make doctors more productive, that won’t lead to job cuts but a reduction in waiting lists.

There’s also huge uncertainty around the speed and nature of improvements in AI. Clearly, the leading models have seen significant advances in the past few years, and rapid progress will continue. But what exactly their capabilities will be within the next year, let alone the next five, is hotly disputed. Some, including Amodei and Altman, think we are five to 10 years away from AGI (artificial general intelligence), which would mean models can do any intellectual task at least as well as a human. Other researchers argue this is decades away.

All we can say for sure is that some change in the labour market is inevitable, as certain tasks become easier to automate. What that means for the total number of jobs is far less certain, though those employer surveys suggest we’re not on the cusp of a “jobpocalypse” in the immediate future.

Beyond these technical questions, though, is a broader one that takes us back to those students booing Schmidt: how will negative perceptions of AI affect its take-up and role in society?

Most people already believe that it will do more harm than good, and that’s before there is much evidence of job losses. For those working in knowledge industries, this has led to a powerful stigma around using AI. When the Nobel prize-winning novelist Olga Tokarczuk told an interviewer she used LLMs (large language models) to help develop ideas and check facts, there was widespread condemnation from the literary world, and she had to put out a clarifying statement.

In the same week, director Seth Rogen was applauded by Hollywood for slamming any screenwriter using an LLM, and pointing out the limitations of AI in film production: “Every time I see a video on Instagram that’s like, ‘Hollywood is cooked’, what follows is the most stupid dog shit I’ve ever seen in my life.” In recent days, there’s been a noisy controversy about whether several winners of the Commonwealth short story prize used LLMs to write their entries.

All this suggests that, as AI use grows, human input will be seen as increasingly high value and status, and a form of rebellion against the dead hand of corporate technology. And this is at current levels of usage.

Now imagine that Amodei and Altman are correct, and their products wipe out whole professions within a few years. The backlash would be extraordinary. They’d be roughly as popular as leprosy. It’s also obviously unsustainable for businesses, including those selling AI, for there to be a vast wave of unemployment that plunges economies worldwide into recession, while all the gains end up with a handful of people. Who would buy their products?

Warnings emanating from Silicon Valley seem bizarrely disconnected to the dystopian scale of the consequences. Dan Schulman, the CEO of the telecoms conglomerate Verizon, recently predicted 20%-30% unemployment within two to five years as a result of the products he is helping to roll out – in both graduate jobs and ones that rely on manual labour because of advances in robotics. His solution was more retraining support. Retraining as what, exactly?

That level of unemployment would be unprecedented for developed countries in the modern era (the UK peak was 12% in 1984). There would be widespread civil unrest and immense pressure on governments to act. States can’t permanently hold off the impact of technological change, but they can slow it via regulation. Firms, too, are subject to social pressure and do not automatically maximise profit to the exclusion of all other considerations (after the financial crash, many kept on more employees than economic models predicted or was strictly necessary given the downturn in demand). Frontier AI labs, such as Anthropic and OpenAI, can also help manage the speed of change via the way they release products. The prospect of being lynched by a mob of angry unemployed graduates is a powerful incentive for caution.

No doubt the stream of dystopian predictions will continue. In reality, changes to the labour market will most likely happen in a manageable way. But, if it looks like we’re on the road to mass unemployment and immiseration, governments and labs, in the face of public pressure, can choose to slow things down. Ultimately, we get to decide how we want to organise society. Just because something is possible, it doesn’t mean we have to do it.

Photograph by Justin Sullivan/Getty Images