AI researchers say ChatGPT and rival chatbots are significantly less capable in languages other than English, exacerbating the bias against non-English speakers
AI chatbots are less fluent in languages other than English, threatening to amplify existing bias in global commerce and innovation. LinkedIn: Pascale Fung , Hans Uszkoreit , Yasmina Jraissati , and Sandhya Surendran . Mastodon: @kentborg@social.tchncs.de Tweets: @blader , @dxmartinjr , @aleyda , @senalexpadilla , @alexpadilla4ca , and @dorrego LinkedIn: Pascale Fung : I wanted to promote multilingual LLM research. Hope the message came through in my interview for this article. — https://www.wired.com/... Hans Uszkoreit : It is quite amazing how well the top LLMs handle other languages and also how many languages they respond to. … Yasmina Jraissati : Sharing a fascinating, thought provoking article here. — Speaking of non English languages being left out of the generative AI revolution: we have a real problem in Arabic. … Sandhya Surendran : Well, that's some food for thought. I'm buzzing with observations on this, but one point sticks out - I'm assuming AI has its origins in … Mastodon: Kent Borg / @kentborg@social.tchncs.de : @Techmeme Sounds like a blessing, other languages being less attacked by LLMs. Tweets: Siqi Chen / @blader : the state of technology journalism today: chatgpt is bad actually because it isn't equally competent in all languages yet. why does anybody pay attention to these people any more? [image] Donald Martin, Jr / @dxmartinjr : This can't be solved without a priori cultural context knowledge that comes from real people embedded in the culture. ""a multilingual model might associate the word dove in all languages with peace even though the Basque word for...https://www.linkedin.com/ ... https://www.wired.com/... Aleyda Solis / @aleyda : ChatGPT Is Cutting Non-English Languages Out of the AI Revolution - Happy to see this topic covered in @WIRED (and my quote in there!) 👇 https://www.wired.com/... Thanks @peard33 🙌 [image] Senator Alex Padilla / @senalexpadilla : Nearly 44% of Californians speak a language other than English—and they're being left behind in the AI revolution. That's why I'm pushing for these technologies to be developed more equitably. Read now: https://www.wired.com/... Alex Padilla / @alexpadilla4ca : The information and education benefits of AI cannot be exclusively reserved for English speakers. https://www.wired.com/... David Orrego-Carmona / @dorrego : “Any good Spanish results are a bonus.” The title of the article could easily be that language models are cutting languages out. The position of English in LLMs is problematic because it ends up enhancing language barriers. https://www.wired.com/...
Context & Ripple Effects
The English-language bias of large language models is not a new complaint — after GPT-3, [[a:969997|startups in China, South Korea, Israel, and Germany began building general-purpose language tools for non-English markets]], betting that local capability would be a differentiator. What the researchers cited here — Pascale Fung among them — add is evidence that the frontier chatbots themselves remain significantly weaker outside English, meaning the gap has persisted even as the models got better.
The story also lands alongside a darker feedback loop documented later in the corpus: inaccurate AI translations of Wikipedia pages are flowing back into training data, threatening a doom spiral that marginalizes vulnerable languages further. Fung's push for multilingual LLM research is positioned against exactly that trajectory.
First-order effects
- Non-English speakers — with Arabic, Basque, and Spanish named in the coverage — get measurably lower-quality chatbot output for the same tasks English users take for granted, putting them at a disadvantage in commerce and innovation that increasingly assumes AI fluency.
- Multilingual NLP researchers like Pascale Fung gain a concrete policy argument: the case for funding non-English model research is now an equity issue, not just a technical one.
Second-order effects
- The regional-buildout thesis behind the 2021 wave of non-English language startups strengthens — if frontier labs under-serve these languages, local players have both the opening and the political cover to fill it.
- Training-data quality becomes a competitive battleground: models trained partly on AI-generated translations inherit and compound each other's errors, so labs that invest in genuine native-language corpora pull ahead in markets others write off.
Third-order effects
- If the pattern holds, global AI access splits into a two-track system — high-capability English models and thinner local alternatives — entrenching which populations capture productivity gains and which are left with degraded tooling, a divide regulators and funders will eventually be pressed to address.
The trend: Frontier AI is consolidating around English-first capability while regional models and multilingual research programs race to prevent a durable two-tier system.