Sometimes I think ChatGPT does the broader field of artificial intelligence a disservice. While hallucinations, slop, exam cheats and robot speechwriters dominate the discourse, other uses of the technology are revolutionising science.

Last year, the Nobel Prize for chemistry was awarded to Google DeepMind's Sir Demis Hassabis and John Jumper for their pioneering work on AI-powered protein folding. AlphaFold promises to hugely expedite drug discovery and development and is already being used to combat cancers and other diseases.

OpenAI has since joined the fray, developing a small, specialised version of its large language models tailored for longevity science, with promising early results. Developments like these, coupled with the promise of much greater advances around the corner, have prompted some to make bold claims about radical life extension thanks to AI.

But while exploring the latest study of global mortality trends from the Institute for Health Metrics and Evaluation, I was struck by another thought: are we underrating the risk that AI could exert not only upward but also downward pressure on human life expectancy?

One of the headline findings is that although mortality rates continue to edge downwards for older people, deaths among younger, and middle-aged adults have been on the rise in many countries, most notably the US, Canada and the UK. This trend is turning what would otherwise be solid increases in headline life expectancy into stagnation.

This concerning wave of middle-aged mortality is often attributed to "deaths of despair" -- premature loss of life through suicide, drugs or alcohol, afflicting people in severe distress. But more recent research, which followed thousands of adults in the US over several decades, found that the "despair" framing is not quite right. What marks out those who succumb to self-destructive behaviour is not psychological distress or financial hardship -- specifically, it is long-term joblessness and social isolation.

My analysis of middle-aged mortality on both sides of the Atlantic confirms that an equally simplistic narrative of "it's just the availability of drugs" is only part of the story. Deaths from substance abuse have not risen uniformly across the population; they affect specific cohorts of people who were hit by large negative employment shocks in early life. This is most pronounced in Scotland, where the bulk of deaths from drugs and suicide have afflicted the generation that grew up during spells of high joblessness during the rapid deindustrialisation of the 1970s and 1980s.

But drugs and economic dislocation aren't enough on their own, either. Many countries in western Europe with fentanyl in circulation have endured bouts of unemployment almost as severe without the subsequent waves of mortality.

Separate data highlights the final piece of the puzzle: both young men and people out of work report being significantly more socially isolated in English-speaking countries, markedly increasing the risk of self-destructive behaviour. Several factors in non-anglophone countries may be playing a protective role here, from levels of religious belief and social solidarity to more concrete factors such as close-knit multigenerational family units. These may allow other populations to weather storms that damage anglophone societies.

This dangerous combination of economic dislocation and social isolation brings us back to AI. If future iterations of today's large language models do, one day, bring about joblessness at scale, it will not be a passing spike. Rather it will mean the lasting destruction of professions and careers, permanently scarring the generation affected. Because enduring involuntary unemployment (much more than financial distress) is what causes the harm, even a substantial universal basic income might not be sufficient to replace the loss of purpose, camaraderie and social contact that work can provide.

Compounding this, spending more time conversing with large language models and less with humans risks greater isolation. There is evidence that digital platforms may be helping to accelerate rates of relationship breakdown.

The result could be a life expectancy tug of war: AI-enhanced health for those who make it into their silver years, and AI-elevated vulnerability for the young and middle-aged. Policymakers and individuals alike would be wise to defend against the latter.