{"text":[[{"start":7.6,"text":"Visions of an AI-infused world can be a little scary. Perhaps our brains will dull as we outsource intellectual struggle to our digital assistants. Perhaps — brace yourself — your jaunty economic analysis will come from a confident large language model, rather than a harried human. I prefer to daydream about a sunnier scenario, in which our new digital tools deliver huge productivity gains. So perusing the latest data and evidence, where are the glimmers of light?"}],[{"start":43.34,"text":"One lies in the excitingly strong headline labour productivity growth in both the UK and US. Admittedly, my threshold for excitement here is quite low, though more importantly there are easier explanations for the uptick than an AI-fuelled boom. In the US, tariff uncertainty could have made companies hesitant about hiring, while in the UK a higher minimum wage may have been clearing out low-paid jobs. Both could raise measured productivity, but not really in the way we want."}],[{"start":77.27000000000001,"text":"More promising signs came from digging into the more granular data. I don’t mean the assortment of corporate anecdotes that have the informational value of parental boasting at the playground. One compilation by Goldman Sachs put the average productivity boost from AI at 32 per cent. And based on the playground chatter, the average child is a chess prodigy, a lover of Swiss chard and a semi-professional trombonist."}],[{"start":104.80000000000001,"text":"Instead, I mean correlations at the industry level. If AI were helping companies squeeze more out of their employees, you would expect the industries adopting AI most enthusiastically to be enjoying the strongest labour productivity growth. In the US that correlation has started to show up in recent data. Though of course, correlation isn’t causation and it could be that more innovative industries were most likely to adopt AI in the first place. "}],[{"start":134.58,"text":"In a post published by the Federal Reserve Bank of St Louis, some economists try to improve on this analysis in two ways. First, rather than blunt AI adoption metrics, they ask people to estimate how much time AI tools saved them at work. Second, they look at recent labour productivity growth between the introduction of ChatGPT and the second quarter of 2025, relative to its trend between 2015 and 2019. This was supposed to strip out any pre-existing trends that could mess up the results."}],[{"start":171.52,"text":"Combining these two metrics, they found that the industries where workers were saving the most time using AI were also the ones seeing unusually fast labour productivity growth. These included information services as well as professional, scientific and technical services. And updating the data to the third quarter of 2025, it looks like the correlation strengthened slightly."}],[{"start":null,"text":"
"}],[{"start":200.20000000000002,"text":"I wouldn’t take the self-reported time savings too literally, not least because not everyone is as diligent as me, reallocating the time I save using ChatGPT (to find data) towards making my output even more jolly. If some people use the extra time to perfect passive-aggressive emails to their colleagues, again, that’s not really the kind of change we want to see."}],[{"start":223.53000000000003,"text":"It’s also reasonable to be sceptical of these correlations because LLMs have only recently graduated from “precocious 11-year-old” to “cocky graduate intern”, and towards the end of 2025 self-reported AI adoption by US businesses was still below 20 per cent. So my final bright spot comes in the form of a study taking a longer view of the data and a broader view of the technology."}],[{"start":248.58000000000004,"text":"Jonathan Haskel, one of the study’s authors, explained that 2017 was the real technological turning point, when a famous “deep learning” paper introduced the transformer architecture in machine learning (the “T” in ChatGPT), boosting generative AI. Which is why they compare the period between 2017 and 2024 with the one between 2012 and 2017."}],[{"start":276.32000000000005,"text":"More specifically, the authors study US investment in software and estimate how much it has contributed to growth. This involves various assumptions, as they try to include both the productivity gains associated with companies becoming better at producing the software, as well as the effects of other industries using it. They estimate that together, these contributed as much as half of the increase in productivity growth between the two time periods."}],[{"start":null,"text":""}],[{"start":305.6,"text":"All of this is suggestive — the sun clearly hasn’t come out fully yet. We don’t have the data to repeat that last analysis in Europe. And when senior McKinsey adviser Tera Allas examined the British data, she couldn’t find any evidence that AI-adopting industries were experiencing unusually high productivity growth. Still, I’m trying to stay positive. Otherwise maybe you will decide that your columnists are better in artificial form."}],[{"start":343.96999999999997,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1770717966_3379.mp3"}