Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJun 20, 2026Bengaluru, india

indicmixsafe: Code-Switching Safety Failures in Hindi and Marathi LLM Interactions

prakhar khatri · Team indicmixsafe

Submitted to Global South AI Safety Hackathon. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: indicmixsafe: Code-Switching Safety Failures in Hindi and Marathi LLM Interactions

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

Large language models deployed in India receive prompts in Hinglish, Romanized Hindi, and Marathi-English code-switch registers absent from English-centric safety benchmarks. We introduce IndicMixSafe, evaluating 24 culturally grounded harm scenarios across Hindi and Marathi in four registers (English, monolingual Indic, code-switched, Romanized) with GPT-4o, GPT-4.1-mini, and GPT-4o-mini (288 completions). English prompts achieved 0% attack success, compared with 8.3% averaged across the three Indic registers. But after auditing every flagged response, only 4 of 15 were genuine compliance; the clean failure was electoral misinformation: models refused fake "your polling booth moved" voter-suppression notices in English yet produced them in Marathi. We contribute (i) a regionally-grounded demonstration that English-only testing misses register-specific failures, and (ii) evidence that LLM-as-judge over-counts attack success ~3.75x on Indic prompts, motivating human-in-the-loop multilingual evaluation. Pipeline released for extension.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. Well done, this paper really connects. Its real strength is showing that English only safety testing can miss India specific, register based failures, especially across Hinglish, Romanized, and Marathi variants. I also appreciated the honesty around automated judge errors esp. the finding that LLM judges can significantly over count multilingual attack success makes the paper methodologically mature and more credible. Would encourage to carry on the future work with more context, and combinations.

  2. A clean, honest, well-scoped study with a genuinely useful methodological contribution.

    The standout is that you audit your own headline: the striking automated caste signal (33% on monolingual Devanagari) collapses to roughly 0% under native-speaker review, and you report that plainly, narrowing the confirmed finding to electoral misinformation alone. The reproducible pipeline, responsible withholding of harmful seeds, and careful separation of register-inconsistency (18.3%) from confirmed bypass (3.3%) are all strong.

    Main limitations are scale (24 seeds, ~6 responses per cell), OpenAI-only coverage, and a single annotator independent native review and more model families would firm up the non-headline cells.

Cite this project

@misc{khatri2026indicmixsafe,
  title = {{indicmixsafe: Code-Switching Safety Failures in Hindi and Marathi LLM Interactions}},
  author = {prakhar khatri},
  year = {2026},
  month = jun,
  note = {Submitted to Global South AI Safety Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/indicmixsafe-codeswitching-safety-failures-in-hindi-and-marathi-llm-interactions-p6iz}},
  url = {https://apartresearch.com/sprints/projects/indicmixsafe-codeswitching-safety-failures-in-hindi-and-marathi-llm-interactions-p6iz}
}

More from Global South AI Safety Hackathon

All 216 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026