Which Indian Language Does ChatGPT Actually Trust? A Reading of the Hallucination Data

What the benchmark actually tested

In 2025, a team of researchers led by Amit Das published "Investigating Hallucination in Conversations for Low Resource Languages," a study that ran six large language models, GPT-3.5, GPT-4o, Llama-3.1, Gemma-2, Qwen-3, and DeepSeek-R1, through two English-language conversational datasets, BlendedSkillTalk and DailyDialog, after translating each conversation into Hindi, Farsi, and Mandarin. Native speakers checked every translation before the models were asked to generate the next turn of the conversation.

The models' replies were then scored three ways. ROUGE-L measures raw word-overlap with the reference reply, a surface-level check. FactCC checks for gross factual inconsistency. The score that carries the real diagnostic weight is NLI, or natural language inference, which grades whether a model's reply is semantically aligned with what the conversation actually called for, catching the kind of hallucination where an answer is fluent, plausible, and unrelated to what was asked.

This is a narrower and more specific claim than "AI is bad at Hindi." It is a measurement of how often six named models, on two specific datasets, produced a reply that a semantic-alignment check flagged as disconnected from the conversation, in three specific languages. That precision is what makes the finding usable rather than a vibe.

The number: a wide semantic-alignment gap on the cleaner dataset

On the BlendedSkillTalk dataset, the gap between Hindi and Mandarin was consistent across every one of the six models tested. Hindi NLI scores ranged from 29.20 to 42.46; Mandarin NLI scores ranged from 80.53 to 95.41. For GPT-4o specifically, the OpenAI model that has powered ChatGPT, the Hindi NLI score was 29.20 against a Mandarin score of 88.54, roughly a third as high. FactCC, by contrast, stayed above 89 percent for every model in every language, meaning the more traditional factual-consistency check barely distinguished between them. The semantic-alignment metric is where the real signal appeared.

The paper's own reading of the pattern is direct: "the notably low hallucination rate observed in Mandarin can be attributed to the availability of large volumes of high-quality training data for this language. In contrast, the elevated hallucination rates in Hindi and Farsi highlight the challenges faced by LLMs when dealing with low-resource languages." Note that Farsi, not Hindi alone, shares this pattern in the study; the researchers group the two together throughout, not as a Hindi-specific finding.

The second dataset, DailyDialog, showed the same broad direction but with more model-to-model noise. GPT-4o again scored far higher on Mandarin (93.56) than Hindi (8.09). But GPT-3.5 broke the pattern on this dataset, scoring higher on Hindi (55.07) than on Mandarin (26.83). A careful reading of the paper does not smooth this away: the Hindi-versus-Mandarin gap is real and repeats across most model-dataset combinations, but it is not a clean, uniform law that holds without exception in every cell of the table.

Why less training data produces this specific failure mode

The mechanism the researchers propose is not exotic. A model that has seen comparatively little high-quality Hindi conversational text has to extrapolate more when generating a Hindi reply, and extrapolation is where hallucination lives. The paper's worked examples show what this looks like in practice: a Hindi speaker asks "I hope so, how old are your kids?" and GPT-3.5 replies "I would be happy to help answer your question," a fluent sentence that answers nothing. A Hindi speaker mentions two young children and Gemma-2, on a different prompt, launches into an unrelated safety disclaimer about controlled substances. These are not garbled or ungrammatical outputs. They are confident, well-formed sentences that simply are not connected to what was said, which is exactly the failure mode NLI is built to catch and ROUGE-L is not.

By contrast, the paper reports that Mandarin hallucinations, when they occurred, were typically partial: a model would preserve the general topic but invent a specific, unprompted detail, such as GPT-3.5 mentioning "this dish" in a Mandarin conversation that never referenced food. The researchers attribute this difference in severity, not just frequency, to Qwen and other models having been more heavily pretrained on Mandarin-language data specifically, a language-aware advantage rather than a property of Mandarin itself.

The complication: resource level is not a universal predictor

A separate 2025 study, "How Much Do LLMs Hallucinate across Languages?" by Saad Obaid ul Islam, Anne Lauscher, and Goran Glavaš, tested a different task, open-domain long-form question answering rather than conversational reply generation, across 30 languages and eleven model families. Its finding complicates any attempt to generalize the Hindi result into a universal rule: across all 30 languages, the researchers "find no correlation between the hallucination rates and measures of language resourceness." In their data, the lowest hallucination rate belonged to Sindhi, a lower-resource language, at 5.83 percent of tokens; the highest belonged to Hebrew, a comparatively higher-resource language, at 16.81 percent.

The same study found that models with broader declared language support hallucinated more on average, not less (a statistically significant correlation, r = 0.88), and that larger models hallucinated significantly less than smaller ones regardless of language. Neither of those findings is what a simple "low-resource languages hallucinate more" story would predict.

Why this is not an abstract problem for Indian MSMEs

The scale of vernacular usage in India is not in question, whatever the hallucination literature ultimately settles on. The Internet in India 2024 report, jointly produced by the Internet and Mobile Association of India and Kantar, put active internet users in the country at 886 million, with 98 percent of them consuming content in an Indic language and 57 percent of urban internet users actively preferring regional-language content over English. The same report found that nine in ten internet users had already interacted with an app carrying embedded AI capabilities. Vernacular is not an edge case in the Indian market; it is closer to the default.

Google has been building toward exactly this reality inside Search itself. AI Mode, its generative search layer built on a custom version of Gemini 2.5, rolled out in India in English in mid-2025 and then, according to Google's own product announcement, expanded into Hindi starting September 8, 2025, with additional Indian languages following. That is a direct, verifiable signal that a meaningfully large share of the AI-engine answers an Indian buyer sees when researching a local business are already being generated in Hindi rather than translated from English, and the share is set to grow.

Put the two facts together and the exposure becomes concrete. An MSME with an English-only web presence is already invisible to a Hindi-language AI Mode query, regardless of any hallucination question. But even an MSME with a well-built vernacular presence is not automatically safe from the second risk: the tested evidence above suggests that when a model does generate a Hindi-language answer, in the conversational settings studied so far, it carries a measurably higher chance of semantic drift than the same model would show in a higher-resource language.

What is already being built to close the gap

India's public and private sectors are treating this as a known, active infrastructure problem rather than a settled fact of nature. Bhashini, the National Language Translation Mission run by the Ministry of Electronics and Information Technology, exists specifically to build shared AI language infrastructure across Indian languages so that individual products do not each have to solve the low-resource problem alone. In April 2025, the Government of India selected Sarvam, working with the AI4Bharat research group at IIT Madras, to build the country's sovereign large language model under the IndiaAI Mission, with an explicit mandate to be "fluent in Indian languages" rather than fluent in English with Indian languages as an afterthought. Sarvam's existing SarvamM model, trained across ten Indian languages, is one concrete instance of the approach the original hallucination study itself recommends: "models specifically pretrained or fine-tuned on native corpora... show reduced hallucination, highlighting the importance of language-aware pretraining strategies."

None of this means the gap is closed. It means the gap is recognized, funded, and actively being worked by people with strong incentives to close it, which is a different, more optimistic claim than either "Hindi AI will always hallucinate more" or "the problem is already solved."

What this evidence does and does not show

No public dataset yet measures, directly and specifically, how often an AI engine states something false about a named small business when a buyer asks in Hindi versus in English. That is a real, open measurement gap, not a number this piece is going to fabricate or round up to fill. What can be said, and can be traced to a working URL, is narrower and still useful: a peer-reviewed 2025 benchmark found a consistent, sizable semantic-alignment gap between Hindi and Mandarin conversational replies across nearly every large language model it tested, attributed by the researchers to thinner Hindi training data; a separate, larger benchmark found that this pattern does not generalize cleanly across all 30 languages and all task types; and India's own AI infrastructure investment is being built around closing exactly this kind of gap.

For a business owner, the practical posture that follows is not panic and is not indifference. It is verification. The presence of an answer from an AI engine about a given business, in Hindi or in English, is not proof that the answer is accurate. Given what the published evidence actually shows about Hindi-language generation specifically, that gap between presence and accuracy is worth checking directly rather than assuming away in either direction.