Calum Chace, co-founder of Conscium, joined Times Radio presenter Carol Walker alongside Mantas Mazeika, a research scientist at the Center for AI Safety, to discuss which jobs AI is actually coming for first. Their answer cut against the standard assumption: it's white-collar work, not manual labour, leading the way.
The numbers on the table:
3 million low-skilled jobs could disappear in the UK by 2035, according to research cited in the segment.
The UK's National Foundation for Educational Research names trades, machine operators and administrative roles as most exposed.
The UK government's own list adds management consultants, psychologists and legal professionals to the most-exposed category.
White-collar jobs are going first, not blue-collar ones
For years, the assumption was that repetitive physical work would be automated before anything requiring judgment or a degree. Calum said that assumption hasn't held up. "A few years ago, everybody said it'll be repetitive blue-collar jobs which get automated first. And then along came large language models, and it turns out it's actually repetitive white-collar jobs that get automated first." He called the result "collar blind" automation: the machines don't distinguish between a factory floor and a law office.
Software development is the clearest case. Mazeika pointed out that large language models have already made themselves central to writing code, to the point where many in the field expect programming to be among the first professions fully automated, despite that being close to the opposite of what most people predicted a decade ago.
The skills machines have, and the ones they're still missing
Mazeika broke down why the order of automation looks the way it does. Current AI has made rapid progress on specific cognitive abilities: short-term memory, abstract reasoning, mathematical ability. Progress has been slower on visual perception and on the kind of long-term, independent agency a job requires over weeks or months rather than a single conversation.
That gap explains the pattern. Work that lives mostly in text, which covers most programming and a large share of knowledge work, sits in the category where AI is already strong. Work that depends on physical dexterity or sustained autonomous judgment sits in the category where it isn't, yet.
Caring roles are holding out, but not by much
Nursing came up as a test case for the limits of automation, and the answer was less reassuring than expected. Japan, dealing with one of the world's most advanced ageing populations, already uses robots in elder care. Calum offered one theory for why robots handle some caregiving tasks surprisingly well: "You can tell the machine the same joke a hundred times a day and it really doesn't care." Patience without fatigue turns out to be a real advantage.
Neither guest treated this as a long-term safe harbour. The point both made repeatedly was that trying to rank professions by how automation-proof they are is, in Calum's words, "a fool's game." Every profession is heading toward at least partial automation, whether that's nursing, plumbing, or accountancy.
Nobody is planning for what happens if this goes all the way
The short-term disruption is one problem. The longer-term one, which Calum said gets far less attention than it deserves, is what happens if AI eventually does everything humans do for money, more cheaply and more reliably. He was clear about the destination while leaving the timeline open: "Unless we stop developing them, or unless there's some kind of silicon ceiling that they can't break through, that will happen." At that point, he argued, the economy as currently structured stops working, since it assumes most people need a job to live.
He doesn't treat that as a purely bad outcome. An economy where machines do the jobs "could be a brilliant world, because humans could get on with doing the important things in life, like having fun and socialising and exploring and travelling and learning." The harder question is who owns the value AI generates. Calum named OpenAI, Google DeepMind and Anthropic directly: if most future economic value is produced by their models, distributing that value fairly to everyone else will take deliberate policy, not something that happens on its own.
What to actually do about it right now
Asked whether people should be learning to program AI models before it's too late, Mazeika's advice was more practical than technical: learn to use the tools that already exist. In the short term, both guests agreed, the advantage goes to people who get fluent with AI quickly, regardless of industry. The bigger structural questions, about jobs, ownership and the shape of a post-work economy, are ones Calum argues governments and companies are barely engaging with, let alone answering.
Watch the full Times Radio segment>>