How AI has de-skilled translation

{"text":[[{"start":null,"text":""}],[{"start":7.35,"text":"Welcome back to The AI Shift, our weekly deep dive into how AI is changing the world of work. This week, we’re looking at how things are going for professional translators. The short answer is: not well, but not necessarily in the way you might think."}],[{"start":23.65,"text":"Sarah writes"}],[{"start":25.099999999999998,"text":"As some of you might know, I have been working on a book over the past few years in which I’ve travelled around the world to meet people on the front lines of technological change, from Hollywood writers and software developers to miners and supervisors of autonomous trucks. (My book is called We Are Not Machines: The Fight for the Future of Work, and it is out today!)"}],[{"start":46.65,"text":"That journey really highlighted to me how the same set of technologies can play out very differently in different professions and workplaces. One contrast struck me as particularly illuminating: between software developers and translators."}],[{"start":61.8,"text":"For experienced software developers, as we’ve written in previous editions of the newsletter, AI coding agents seem to have made the job of being a developer more human in some ways: they’re spending less time (if any) hand writing code, and more time on those interpersonal “glue” skills - seeing the bigger picture, co-ordinating between teams and agents, applying their own taste and creativity."}],[{"start":84.9,"text":"But for translators, the opposite has happened. After AI and other machine translation tools came along — many of the agencies which intermediate between clients and freelance translators simply stopped paying humans to translate the subtitles from scratch. Instead, they rolled out something they called ‘machine translation post-editing’ or MTPE. Under this system, the agencies use machine translation or AI tools to translate the subtitles, then send that version to freelancers, who are now paid to check, correct and finesse the machine’s output. The agencies expect the translators to do this much more quickly than translating from scratch, and they have slashed their rates accordingly. One translator I interviewed for my book called Petr Čermoch, who translates the subtitles of TV shows from English to Czech, said one big agency in the sector used to offer $5 per minute of video, but once they shifted to MTPE, they cut the rate to $1.50 per minute. Another translator, who translates legal, financial and academic texts between Spanish, English and Estonian, told me the agencies in her field slashed rates per word by about 50 per cent when they brought in MTPE."}],[{"start":153.95,"text":"This has had two impacts. The first — according to the translators I interviewed — is that MTPE work is actually quite cognitively taxing — at least, if you maintain high standards for quality and accuracy. Mark Rawson, a translator in Taiwan who translates between English and Chinese, explained to me that in order to check a machine translation thoroughly, you have to look at the original source text and the machine text simultaneously. “You’ve got to interrogate the hell out of the Chinese, and do the same with the English,’ he said. ‘It’s two or three times harder than pure translation work, and they have the barefaced cheek to pay you half as much.”"}],[{"start":192.89999999999998,"text":"But the drop in pay also means you have to speed up. As another translator put it: “You’re having to churn out text as quickly as possible to try and maintain your hourly rate, but you’re also really stressed because you can’t let any errors through. You still have to be quite fastidious in receiving all this machine translation output, and that’s exhausting, it’s stressful.” As Čermoch, the TV subtitle translator, put it: the work feels as if it has become less creative compared with translating from scratch, and more mechanical. “It loses all the parts that I love about it,” he told me. “It’s just a tedious job — boring and bland and lifeless.”"}],[{"start":232.49999999999997,"text":"There is another contrast too. While AI seems to be a “seniority-biased” or “expertise-biased” technology so far in the software world — boosting the productivity of those with the skills and taste to use the tools well — in translation, AI seems to be having a de-skilling effect. Many experienced translators told me they are refusing to do MTPE work, and as they leave the market, people with less skill and experience are stepping in. I also gather that more companies are beginning to drop the MTPE layer altogether, because they are increasingly satisfied with pure machine translations."}],[{"start":267.75,"text":"John, what does your look at the data tell us about all this? And do you have thoughts about why AI is playing out so differently in the worlds of translation and software engineering?"}],[{"start":278.45,"text":"John writes"}],[{"start":280.9,"text":"Thanks Sarah, I find your reporting on translators such a powerful illustration of how technology can subtly warp a job in ways that make it much less appealing and rewarding. The detail on how switching from doing the whole job yourself to reviewing machine work actually makes the job harder and slower was particularly eye-opening."}],[{"start":300.25,"text":"While the data on translators’ labour market outcomes doesn’t capture that nuance, it tells a similar tale of technology eroding demand for human translators and undermining their status, with distinct inflection points for the arrival of first Google Translate and now LLMs."}],[{"start":318,"text":"In the US, where detailed data is most readily available, the steep upward march in the number of interpreters and translators slowed to a crawl around 2010 as Google Translate took off, and translators’ pay relative to the whole-economy average peaked and began to decline at roughly the same time. This matches the findings of a study last year by Oxford university economic historians Carl Benedikt Frey and Pedro Llanos-Paredes, and certainly looks to me like support for the idea that while imperfect and inferior automated tools may not necessarily replace human specialists overnight, they often start eating into demand for new work, as well as reducing its perceived value."}],[{"start":358.55,"text":"And where Google Translate slowed growth, more recent developments and products in AI translation appear to be going a step further; the number of translators has fallen markedly in the last five years, and is now down almost 20 per cent on its peak in terms of its share of all US employment. Similarly, typical wages for translators are now consistently below the whole-economy average for the first time."}],[{"start":null,"text":""}],[{"start":382.45,"text":"It’s also worth emphasising that this data is just capturing employees, not freelancers, with the impacts on the latter likely even more pronounced due to the absence of regulations and long-term contracts. Indeed, the number of new translating projects advertised on major online marketplaces fell by almost 50 per cent in the two years following the launch of ChatGPT according to data from the Online Labour Index, an alternative economic indicator developed by the Oxford Internet Institute to track activity in the online gig economy. This latter detail reminded me of our conversation with Rose Mutiso last month about how AI might impact developing countries: entrepreneurial workers in lower and middle-income countries benefited from the boom in online white collar piece-work during the 2010s, but are now highly exposed to having that work taken by AI."}],[{"start":432.59999999999997,"text":"Overall, the tale of the translator does look sadly similar to that of the 19th century stocking-makers you discuss in your book. Now as then, automation is bringing a very useful product or service to more people than ever and at lower cost — surely a positive — while delivering a quadruple dose of pain to a group of skilled professionals: reducing their employment, pay and satisfaction with the job they are asked to do, and adding insult to injury by replacing their work with an often inferior substitute."}],[{"start":463.4,"text":"On the translators vs coders contrast, this is a perfect illustration of how it’s hugely important which part of a job AI takes over, and the economic and human perspectives are aligned here. In software, the act of typing out code is a routine, laborious and in a sense unthinking part of the job — a means of executing on the higher-value and more rewarding work of creatively solving problems. Removing that means incumbent developers get to spend more time on the parts of the job that give them the most satisfaction, but it risks disrupting entry-level roles where the more routine tasks were a large part of the job. In translation, as your case studies make clear, AI has taken away the creative problem-solving task of thinking deeply about how best to convey meaning and context in a new tongue, and has left a much more routine, fragmented and less specialised husk of a job. The big question is whether translators are a special case — it is after all difficult to think of a job that more closely mirrors the inner workings of large language models — or whether a similar experience could be coming the way of many more knowledge work occupations."}],[{"start":530.1,"text":"Recommended reading"}],[{"start":null,"text":"

1. As you may have gathered, I think Sarah’s new book is fantastic and an invaluable companion through this latest wave of automation (John)

2. Our colleagues have a good (if not reassuring) explainer of the idea of “recursive self-improvement”, which AI labs say may solve all our problems, or kill us all (Sarah)

"}],[{"start":null,"text":""}],[{"start":537.9,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1780715329_3856.mp3"}