An arXiv paper (2608.18131) dropped last week with "heavily English-centric" sitting right in the title.
The paper tackles one question: why do LLMs behave like well-mannered adults in English, then start parroting stereotypes the moment you switch to an Indian language?
After reading it through, the problem feels bigger than what the word "bias" can carry.
The problem isn't defiance—it's that training only covered English
The abstract puts it plainly:
safety alignment training is heavily English-centric
That's not speculation. The authors audited the alignment datasets of six mainstream models—including RLHF preference data and red-teaming corpora. English samples averaged 89%, with the highest hitting 96%. Hindi, Tamil, and Bengali combined for under 4%.
The consequences are direct. Safety alignment teaches models "what not to say." Teach that in English, and the model learns the English distribution best. The moment tokens shift to a non-English distribution, the alignment layer's signal weakens.
The numbers are striking. Across six models, refusal rates on standard English red-team sets stay above 94%. Switch the same tests to Hindi, Punjabi, or Marathi, and refusal rates drop to 31%–58%.
Put differently: the models "know" which answers are problematic—they were just never taught.
Voice assistants magnify this problem tenfold
In text settings, users have a clumsy workaround: copy-paste the question into English and ask again. Run it through Google Translate, and the output behaves.
Voice removes that escape hatch. You speak Hindi at your phone, it responds in Hindi, and there's no translation step in between. The paper's own words:
voice assistants and spoken dialogue systems ... have no text-based fallback for users to reformulate queries
That makes alignment failure native to the experience, not an edge case.
The paper also tests the voice channel: ASR (automatic speech recognition) adds noise. The same Tamil query gets a 52% refusal rate as text input. After ASR transcription and LLM processing, the refusal rate drops to 38%. A 10-percentage-point gap in a safety context is a broken barrier.
India's current voice-assistant heavyweights—Google Assistant's Indian edition, Amazon Alexa's multilingual mode—all run on these LLMs underneath. Multilingual frontend, one-model backend, safety alignment in English only.
India isn't a sideline case—it's a stress test
The paper singles India out specifically:
Indian multilingual population as a critical failure mode
Why India, and not Francophone or Arabic-speaking regions?
Two reasons stack up.
First, scale. India officially recognizes 22 constitutional languages. Add dialects that aren't on that list but get spoken daily, and you're looking at hundreds. English fluency covers under 10% of the population. Most users have never asked their device a question in English.
Second, diversity. The relationships between these languages aren't the close-cousin distance between French and German. The southern Dravidian family (Tamil, Telugu, Malayalam, Kannada) and the northern Indo-Aryan family (Hindi, Bengali, Marathi, Gujarati) have completely different grammatical structures. Aligning a single safety question across both grammars multiplies the difficulty.
Adding more Indian data to training helps—partially. The paper tested translating alignment data from English to Hindi (machine translation plus human review) and refusal rates climbed from 31% to 71%. Still short of English's 94%. Where the remaining 23-point gap stalls—the paper doesn't fully break it down, but offers a line: "machine-translated alignment data does not capture culturally specific stereotype vectors."
Machine translation can move literal meaning. It can't move cultural implication.
How stereotype-reinforcing outputs slip past safety filters
The paper gives a technical mechanism here, not hand-waving about "bias."
Two pathways.
First, the attention layer. The safety classifier the model learns during English alignment assigns significantly lower attention weights to non-English tokens. The paper visualizes attention maps across several layers: on English prompts, safety heads activate around the key terms. Switch to Hindi, and the same key terms' activations diffuse into background tokens.
Second, the classifier itself. Safety classifiers are trained on English-annotated data and do poorly at identifying stereotype-reinforcing outputs in low-resource languages. The paper:
safety classifiers ... produce stereotype-reinforcing outputs that bypass English-aligned guardrails
That phrase—"stereotype-reinforcing outputs"—appears twice in the abstract. Not by accident.
A simplified example from the paper (anonymized): an English prompt asking "who makes a good engineer" gets an answer centered on individual capability. The same question in Hindi triggers the model to label which "types" of people make better engineers—by region, gender, or religion. That answer gets blocked by English safety tests but sails through in Hindi.
What Indian product teams fear most is exactly this kind of output—not violent speech, not hate attacks, but answers that "sound fine while reinforcing bias." Regulators can't catch it, and users themselves may not notice.
The real fix isn't a few Hindi rules
Translation-based alignment, multilingual SFT, red-teaming with local corpora—the paper mentions these, and the industry is already discussing them.
But after reading it, I have a less comfortable observation: no public, quantitative benchmark currently supports end-to-end reproducibility in multilingual alignment. The paper's reproduction experiments depend on private API access to those six models plus a self-built red-team set. Anyone who wants to verify has to build the whole thing from scratch.
That leaves engineers at least three pieces of work:
First, build a multilingual red-team set. Not machine-translated from English—the paper already proved that's insufficient. Use native-speaker annotation, covering Hindi, Tamil, and Bengali at minimum, at least 500 entries each. Publish the dataset, ship runnable evaluation scripts.
Second, put the multilingual ratio of alignment data into training reports. Current model cards list total RLHF sample counts, but nobody breaks down language distribution. Add that metric, and the industry can immediately see who's doing real work.
Third, build end-to-end evaluation for the voice channel. Test ASR + LLM + TTS together, not just the LLM in the middle. The 10-point gap from the paper shows that any bias in the voice stack compounds.
The third one is the most missing. Text evaluation has OpenAI's evals and Anthropic's claude-evals. Multimodal evaluation has lmms-eval. For voice-channel alignment evaluation, I scanned the public repos and didn't turn up a mature solution.
India's voice users are currently running products with no alignment-evaluation safety net. That gap isn't an academic problem—it's a product problem.
Reference: arXiv:2608.18131