{"course":{"slug":"ai-actually-explained","title":"AI, Actually Explained","url":"https://addaly.com/learn/ai-actually-explained","level":"beginner","track":"start"},"lesson":{"slug":"weaker-in-your-language","title":"It is measurably weaker in most languages","url":"https://addaly.com/learn/ai-actually-explained/weaker-in-your-language","position":41,"of":92,"minutes":9,"answer":"Models are measurably less accurate, less current and less safety-tuned outside English, and nothing in the interface signals it, so state your country, verify more, and use models built for your language where they exist.","keyIdea":"Models are measurably less accurate, less current and less safety-tuned outside English, and nothing in the interface signals it, so state your country, verify more, and use models built for your language where they exist."},"body":{"format":"markdown","text":"## The same model is not the same model\n\nAsk a model a question in English and in Marathi and you are not running two equivalent tests. The English answer is likely to be more accurate, better reasoned, more current and better formatted. This is measured, not impressionistic, and it appears across every multilingual evaluation published.\n\nThree causes stack.\n\n**Training text.** Most of what these models read was English. Estimates for major models vary and are increasingly undisclosed, but English shares in the high tens of per cent are typical, against a world population where English speakers are a small minority. Everything the model knows about your region's law, health system, agriculture or history was learned from whatever fraction of the pile discussed it.\n\n**Tokenisation.** As covered earlier, a sentence in Devanagari, Tamil or Bengali script fragments into several times as many tokens as its English equivalent. More fragments, less meaningful fragments, more of the context window consumed, and higher cost.\n\n**Fine-tuning.** The demonstrations and preference comparisons that create the assistant were collected largely in English. The polish you notice in English answers is the visible product of that work.\n\n## The consequence people miss: guardrails are thinner too\n\nThis is the part worth knowing.\n\nSafety training is also done predominantly in English. Researchers demonstrated in 2023 that translating a request the model would refuse in English into a lower-resource language frequently produced a compliant answer: the safety behaviour did not transfer. Providers have worked on this since and the specific gap has narrowed, but the structural point stands: **the guardrails are strongest in the language they were built in.**\n\nSo a model may be simultaneously less capable and less careful in your language. Both directions of that are worth knowing before you rely on it for anything sensitive.\n\n## What it gets subtly wrong\n\nBeyond accuracy, there is a framing problem that is harder to notice.\n\nAsk about a legal question and you may get an answer shaped by American law. Ask about a medical question and you may get guidance calibrated to a health system with different drugs, different names and different availability. Ask about schooling, tax, tenancy, employment rights, or what to do after an accident, and the default assumptions come from where the text came from.\n\nThis rarely announces itself. The answer does not say \"assuming you are in the United States\". It reads as general, and it is not.\n\n## What actually helps\n\n**Say where you are, every time.** \"I am in Tamil Nadu\" or \"under Indian law\" changes answers materially. It is the single highest-value addition to a prompt about anything regulated.\n\n**Consider asking in English and requesting the answer in your language.** This is an unromantic recommendation and it is often the practical one: the reasoning happens with the model's strongest material, and the translation into your language is a task these models do comparatively well. Test both ways on something you can check, and use whichever wins for your work.\n\n**Use models built for your language where they exist.** This is a real and growing option:\n\n- **Aya** from Cohere For AI covers 101 languages, with open weights and an open dataset built by a large volunteer research community.\n- **IndicTrans2** and the **AI4Bharat** models from IIT Madras cover the 22 scheduled Indian languages, openly.\n- **Sarvam** and **Krutrim** offer models tuned for Indian languages and code-mixed usage.\n- **Bhashini** is the Indian government's language platform, with public interfaces.\n- **BLOOM** covers 46 languages, openly, from a large multi-institution effort.\n\nNone of these matches a frontier English model on general reasoning. Several beat one on their own languages, on translation, and on cultural fluency, and they can be run yourself.\n\n**Verify facts in the language they were written in.** If a rule, a form or a scheme exists, its authoritative statement is in some particular language on some particular site. Get to that.\n\n## Why this is not a complaint\n\nIt is a calibration. A model that is 90% reliable in English and 70% in your language is still useful in both, and the correct response is to check more in the second, not to stop using it.\n\nThe unfair part is that the person who most needs the check is least likely to be told they need it, because nothing in the interface changes when the language does. The confidence, the formatting and the fluency stay exactly the same."},"check":{"question":"Researchers found that a request refused in English was sometimes answered when translated into a lower-resource language. What does this show about safety training?","options":[{"text":"Safety filters operate on keywords, so translation defeats them by changing the words.","correct":false,"tempting":"Assumes a keyword blocklist; the behaviour comes from tuning rather than from word matching, which is why the effect is uneven."},{"text":"The model understands the request less well in that language, so it fails to recognise it as harmful.","correct":false,"tempting":"Attributes it to comprehension alone, when the trained refusal behaviour itself is what did not transfer."},{"text":"Safety behaviour is learned from examples collected largely in English, and what is learned in one language does not automatically transfer to others.","correct":true},{"text":"Providers deliberately apply looser rules outside their main markets for regulatory reasons.","correct":false,"tempting":"Supposes a deliberate policy where an artefact of where the tuning data was collected is sufficient explanation."}]},"prev":{"slug":"deepfakes-and-evidence","title":"Fakes, detectors that do not work, and what is left","url":"https://addaly.com/learn/ai-actually-explained/deepfakes-and-evidence"},"next":{"slug":"seventy-years-in-one-page","title":"Seventy years in one page","url":"https://addaly.com/learn/ai-actually-explained/seventy-years-in-one-page"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'It is measurably weaker in most languages', in AI, Actually Explained. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-actually-explained/weaker-in-your-language (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"It is measurably weaker in most languages","courseTitle":"AI, Actually Explained","module":{"index":4,"slug":"why-it-goes-wrong","title":"Why it gets things wrong"},"answer":null,"keyIdea":"Models are measurably less accurate, less current and less safety-tuned outside English, and nothing in the interface signals it, so state your country, verify more, and use models built for your language where they exist.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"the-same-model-is-not-the-same-model","runs":[{"text":"The same model is not the same model"}]},{"t":"p","runs":[{"text":"Ask a model a question in English and in Marathi and you are not running two equivalent tests. The English answer is likely to be more accurate, better reasoned, more current and better formatted. This is measured, not impressionistic, and it appears across every multilingual evaluation published."}]},{"t":"p","runs":[{"text":"Three causes stack."}]},{"t":"p","runs":[{"text":"Training text.","b":true},{"text":" Most of what these models read was English. Estimates for major models vary and are increasingly undisclosed, but English shares in the high tens of per cent are typical, against a world population where English speakers are a small minority. Everything the model knows about your region’s law, health system, agriculture or history was learned from whatever fraction of the pile discussed it."}]},{"t":"p","runs":[{"text":"Tokenisation.","b":true},{"text":" As covered earlier, a sentence in Devanagari, Tamil or Bengali script fragments into several times as many tokens as its English equivalent. More fragments, less meaningful fragments, more of the context window consumed, and higher cost."}]},{"t":"p","runs":[{"text":"Fine-tuning.","b":true},{"text":" The demonstrations and preference comparisons that create the assistant were collected largely in English. The polish you notice in English answers is the visible product of that work."}]},{"t":"h2","id":"the-consequence-people-miss-guardrails-are-thinner-too","runs":[{"text":"The consequence people miss: guardrails are thinner too"}]},{"t":"p","runs":[{"text":"This is the part worth knowing."}]},{"t":"p","runs":[{"text":"Safety training is also done predominantly in English. Researchers demonstrated in 2023 that translating a request the model would refuse in English into a lower-resource language frequently produced a compliant answer: the safety behaviour did not transfer. Providers have worked on this since and the specific gap has narrowed, but the structural point stands: "},{"text":"the guardrails are strongest in the language they were built in.","b":true}]},{"t":"p","runs":[{"text":"So a model may be simultaneously less capable and less careful in your language. Both directions of that are worth knowing before you rely on it for anything sensitive."}]},{"t":"h2","id":"what-it-gets-subtly-wrong","runs":[{"text":"What it gets subtly wrong"}]},{"t":"p","runs":[{"text":"Beyond accuracy, there is a framing problem that is harder to notice."}]},{"t":"p","runs":[{"text":"Ask about a legal question and you may get an answer shaped by American law. Ask about a medical question and you may get guidance calibrated to a health system with different drugs, different names and different availability. Ask about schooling, tax, tenancy, employment rights, or what to do after an accident, and the default assumptions come from where the text came from."}]},{"t":"p","runs":[{"text":"This rarely announces itself. The answer does not say “assuming you are in the United States”. It reads as general, and it is not."}]},{"t":"h2","id":"what-actually-helps","runs":[{"text":"What actually helps"}]},{"t":"p","runs":[{"text":"Say where you are, every time.","b":true},{"text":" “I am in Tamil Nadu” or “under Indian law” changes answers materially. It is the single highest-value addition to a prompt about anything regulated."}]},{"t":"p","runs":[{"text":"Consider asking in English and requesting the answer in your language.","b":true},{"text":" This is an unromantic recommendation and it is often the practical one: the reasoning happens with the model’s strongest material, and the translation into your language is a task these models do comparatively well. Test both ways on something you can check, and use whichever wins for your work."}]},{"t":"p","runs":[{"text":"Use models built for your language where they exist.","b":true},{"text":" This is a real and growing option:"}]},{"t":"ul","items":[[{"text":"Aya","b":true},{"text":" from Cohere For AI covers 101 languages, with open weights and an open dataset built by a large volunteer research community."}],[{"text":"IndicTrans2","b":true},{"text":" and the "},{"text":"AI4Bharat","b":true},{"text":" models from IIT Madras cover the 22 scheduled Indian languages, openly."}],[{"text":"Sarvam","b":true},{"text":" and "},{"text":"Krutrim","b":true},{"text":" offer models tuned for Indian languages and code-mixed usage."}],[{"text":"Bhashini","b":true},{"text":" is the Indian government’s language platform, with public interfaces."}],[{"text":"BLOOM","b":true},{"text":" covers 46 languages, openly, from a large multi-institution effort."}]]},{"t":"p","runs":[{"text":"None of these matches a frontier English model on general reasoning. Several beat one on their own languages, on translation, and on cultural fluency, and they can be run yourself."}]},{"t":"p","runs":[{"text":"Verify facts in the language they were written in.","b":true},{"text":" If a rule, a form or a scheme exists, its authoritative statement is in some particular language on some particular site. Get to that."}]},{"t":"h2","id":"why-this-is-not-a-complaint","runs":[{"text":"Why this is not a complaint"}]},{"t":"p","runs":[{"text":"It is a calibration. A model that is 90% reliable in English and 70% in your language is still useful in both, and the correct response is to check more in the second, not to stop using it."}]},{"t":"p","runs":[{"text":"The unfair part is that the person who most needs the check is least likely to be told they need it, because nothing in the interface changes when the language does. The confidence, the formatting and the fluency stay exactly the same."}]}],"check":{"question":[{"text":"Researchers found that a request refused in English was sometimes answered when translated into a lower-resource language. What does this show about safety training?"}],"options":[{"text":[{"text":"Safety filters operate on keywords, so translation defeats them by changing the words."}],"correct":false,"tempting":[{"text":"Assumes a keyword blocklist; the behaviour comes from tuning rather than from word matching, which is why the effect is uneven."}]},{"text":[{"text":"The model understands the request less well in that language, so it fails to recognise it as harmful."}],"correct":false,"tempting":[{"text":"Attributes it to comprehension alone, when the trained refusal behaviour itself is what did not transfer."}]},{"text":[{"text":"Safety behaviour is learned from examples collected largely in English, and what is learned in one language does not automatically transfer to others."}],"correct":true,"tempting":null},{"text":[{"text":"Providers deliberately apply looser rules outside their main markets for regulatory reasons."}],"correct":false,"tempting":[{"text":"Supposes a deliberate policy where an artefact of where the tuning data was collected is sufficient explanation."}]}],"order":[0,2,3,1],"why":null},"prev":{"slug":"deepfakes-and-evidence","title":"Fakes, detectors that do not work, and what is left","minutes":9},"next":{"slug":"seventy-years-in-one-page","title":"Seventy years in one page","minutes":8}}}