The Economist has deployed AI-translated content on its budget-friendly “snack-sized” app Espresso after deciding the technology had reached the “good enough” mark.
Ludwig Siegele, senior editor for AI initiatives at The Economist, told Press Gazette that AI translation will never be a “solved problem”, especially in journalism because it is difficult to translate well due to its cultural specificities.
However he said it has reached the point where it is good enough to have introduced AI-powered, in-app translations in French, German, Mandarin and Spanish on The Economist’s “bite-sized”, cut-price app Espresso (which has just over 20,000 subscribers).
Espresso has also just been made free to high school and university students aged 16 and older globally as part of a project by The Economist to make its journalism more accessible to audiences around the world.
Related
Siegele said that amid “lots of hype” about AI, the questions to ask are: “What is it good for? Does it work? And does it work with what we’re trying to do?”
He added that the project to make The Economist’s content “more accessible to more people” via Espresso was a “good point to start”.
Content from our partners
“The big challenge of AI is the technology, at least for us, is not good enough,” he continued. “It’s interesting, but to really develop a product, I think in many cases, it’s not good enough yet. But in that case, it worked.
“I wouldn’t say that translation is a solved problem, it is never going to be a solved problem, especially in journalism, because journalism is really difficult to translate. But it’s good enough for that type of content.”
The Economist is using AI translation tool DeepL alongside its own tech on the backend.
“It’s quite complicated,” Siegele said. “The translation is the least of it at this point. The translation isn’t perfect. If you look at it closely it has its quirks, but it’s pretty good. And we’re working on a kind of second workflow which makes it even better.”
The AI-translated text is not edited by humans because, Siegele said, the “workflow is so tight” on Espresso which updates around 20 times a day.
“There is no natural thing where we can say ‘okay, now everything is done. Let’s translate, and let’s look at the translations and make sure they’re perfect’. That doesn’t work… The only thing we can do is, if it’s really embarrassing, we’ll take it down and the next version in 20 minutes will be better.”
One embarrassing example, Siegele admitted, is that the tool turned German Chancellor Olaf Scholz into a woman.
But Siegele said a French reader has already got in touch to say: “I don’t read English. This is great. Finally, I can read The Economist without having to put it into Google Translate and get bad translations.”
The Economist’s AI-translated social videos
The Economist simultaneously launched AI-translated videos on its social platforms in the same four languages.
The videos are all a maximum of 90 seconds meaning it is not too much work to check them – crucial as, unlike the Espresso article translations, they are edited by humans (native language speakers working for The Economist) taking about 15 minutes per video.
For the videos The Economist is using AI video tool Hey Gen. Siegele said: “The way that works is you give them the original video and they do a provisional translation and then you can proofread the translation. So whereas the translations for the app are basically automatic – I mean, we can take them down and we will be able to change them, but at this point, they’re completely automatic – videos are proofread, and so in this way we can make sure that the translations are really good.”
In addition they are using “voice clones” which means journalists who speak in a video have some snippets of themselves given to Hey Gen to build and that is used to create the finished product.
The voice clones are not essential, Siegele explained, as translations can be done automatically regardless. Journalists can opt out of having their voices used in this way, and any data stored will be deleted if the employee leaves The Economist. But the clones do mean the quality is “much better”.
They have a labelling system for the app articles and videos that can show they are “AI translated” or “AI transformed”. But, Siegele said, they are “not going to have a long list of AI things we may have used to build this article for brainstorming or fact checking or whatever, because in the end it’s like a tool, it’s like Google search. We are still responsible, and there’s almost always a human except for edge cases like the Espresso translations or with podcast transcripts…”
Economist ‘will be strategic’ when choosing how to roll out AI
Asked whether the text translation could be rolled out to more Economist products, Siegele said: “That’s of course a goal but it remains to be seen.”
He said that although translation for Espresso is automated, it would not be the goal to do the same throughout The Economist.
He also said they still have to find out if people are “actually interested” and if they can “develop a translation engine that is good enough”.
“But I don’t think we will become a multi-linguistic, multi-language publication anytime soon. We will be much more strategic with what we what we translate… But I think there is globally a lot of demand for good journalism, and if the technology makes it possible, why not expand the access to our content?
“If it’s not too expensive – and it was too expensive before. It’s no longer.”
Other ways The Economist is experimenting with AI, although they have not yet been implemented, include a style bot and fact-checking.
Expect to see “some kind of summarisation” of articles, Siegele continued, “which probably will go beyond the five bullet points or three bullet points you increasingly see, because that’s kind of table stakes. People expect that. But there are other ways of doing it”.
He also suggested some kind of chatbot but “not an Economist GPT – that’s difficult and people are not that interested in that. Perhaps more narrow chatbots”. And said versioning, or repurposing articles for different audiences or different languages, could also follow.
“The usual stuff,” Siegele said. “There’s only so many good ideas out there. We’re working on all of them.” But he said he wants colleagues to come up with solutions to their problems rather than him as “the AI guy” imposing things.
Email pged@pressgazette.co.uk to point out mistakes, provide story tips or send in a letter for publication on our "Letters Page" blog
