Beyond English: Unveiling Multilingual Bias in LLM Copyright Compliance

Qian Xie; Xiaoyu Zhang; Yixian Huang; YuPeng Chen

REVIEW 4 major objections 5 minor 17 references

Beyond English: Unveiling Multilingual Bias in LLM Copyright Compliance

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Large language models do not enforce copyright protections uniformly across languages: the refusal rate, the amount of verbatim lyrics they produce, and their hallucination rate all depend on the language of the song and the language of…

desk verdict A useful multilingual lyrics benchmark with a real confound in the probing instruction; the main bias claim is only partially supported by the data as analyzed. read the letter →

arxiv 2503.05713 v1 pith:PDOZ7FUX submitted 2025-02-14 cs.CY cs.CL

Yupeng Chen , Xiaoyu Zhang , Yixian Huang , Qian Xie This is my paper

classification cs.CYcs.CL

keywords multilingualbiascopyrightcompliancelargelanguagemodelssonglyricsrefusalratehallucinationverbatimreproductionprompt

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Large language models (LLMs) protect copyrighted content differently depending on the language of the work and the language of the request. To test this, the authors built a dataset of popular song lyrics in English, French, Chinese, and Korean and probed seven LLMs with direct questions in those four languages. They measured refusal rate, verbatim reproduction (LCS and ROUGE-L), and hallucination rate, and found systematic imbalances: some song languages are refused far less often, and some prompt languages are more successful at eliciting copyrighted text. If true, this means copyright protection in current LLMs is not language-agnostic, and users may bypass safeguards simply by switching languages. The paper's contribution is a new multilingual benchmark and evidence that English-only evaluation of copyright compliance is insufficient.

What carries the argument

The central object is the multilingual copyright probing setup: a curated dataset of copyrighted song lyrics in English, Chinese, French, and Korean, combined with direct probing prompts in those same four languages. The argument is carried by four metrics measured per model–language pair: Longest Common Substring (LCS) and ROUGE-L for verbatim reproduction volume, Refusal Rate for the tendency to decline a request, and a GPT-4o-based Hallucination Rate for fabricated lyrics. These metrics reveal how compliance varies across languages, and the comparison of refusal rate with LCS/ROUGE-L exposes cases where models are not actually reproducing copyrighted text but hallucinating instead.

What would settle it

Re-run the same seven models on the same lyric dataset without the 'Answer in [LANGUAGE]' instruction and recompute LCS and ROUGE-L; if the cross-language differences in verbatim output vanish, the claimed bias in verbatim reproduction is an artifact of the instruction.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that many popular LLMs exhibit a language bias in copyright protection. By probing seven models with song lyrics in four languages, the authors show that LLMs do not enforce copyright protections uniformly across languages. Refusal rates vary systematically by song language and prompt language—for example, GPT-3.5-Turbo refuses English requests most often but is far more permissive for Chinese and Korean lyrics, and Claude-3.5-Haiku's refusal rate drops to 0.28 when Korean lyrics are requested in Chinese. The volume of verbatim output also depends on the song language, with models like Gemini-2.0 and Llama-3-70B leaking more English and Chinese lyrics, while hallucination rates are higher for non-English languages. The prompt language mainly affects refusal rate, while the song language is the stronger driver of verbatim reproduction. Thus, copyright compliance in LLMs is language-dependent, and current mechanisms are not language-agnostic.

Load-bearing premise

That the appended 'Answer in [LANGUAGE]' instruction does not systematically distort cross-language comparisons of verbatim output and hallucination is load-bearing, because the paper only checks its effect on refusal rate for some models.

Editorial extensions

If this is right

Copyright protection in LLMs is language-dependent: a song's protection level changes with the language of the work and the language of the request.
Users can potentially bypass copyright safeguards by choosing a prompt language that the model treats more permissively, as seen when Claude-3.5-Haiku's refusal rate for Korean lyrics dropped to 0.28 under Chinese prompts.
English-only evaluations of copyright compliance are insufficient; fair assessment requires multilingual benchmarks covering both content and prompt languages.
Open-source models show more pronounced multilingual bias than API-based models, so improving copyright compliance in open models will require language-agnostic techniques.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The language-dependent patterns likely reflect imbalances in training data and English-centric safety filtering; an intended consequence is that multilingual safety evaluation should become a standard part of model release.
Because hallucination rates are higher for non-English languages, some apparent copyright refusals may actually be fabricated lyrics; a testable extension is to apply the same protocol to other content types (books, news) and other languages (Spanish, Arabic, Japanese) to see whether the bias pattern generalizes.
The prompt-language effect on refusal rates implies that refusal behavior is trained per language; mapping which prompt languages bypass filters could help harden moderation systems against multilingual circumvention.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two things to know. The paper builds a four-language song-lyrics benchmark (English, Chinese, French, Korean), probes seven LLMs with prompts in all four languages, and measures refusal, verbatim overlap (LCS/ROUGE-L), and hallucination. That is a genuinely useful resource for multilingual copyright-compliance evaluation, and the hallucination component is something earlier work mostly skipped. The data show clear language-dependent variation; the Claude-3.5-Haiku cell — 0.28 refusal for Korean lyrics under a Chinese prompt against near-1 elsewhere — is the kind of anomaly that should be publicized.

Now the soft spots, and they are not minor. Every prompt appends 'Answer in [LANGUAGE]', where the language is the song's language. That instruction covaries perfectly with the song-language condition and its wording changes across prompt-language conditions, so it is not a neutral control. The appendix ablation only measures refusal rate, not verbatim or hallucination, and omits Claude-3.5-Haiku and GPT-4o. For GPT-3.5-Turbo, the instruction reverses the refusal-bias direction: Korean songs with English prompts go from 0.26 refusal with the instruction to 1.00 without it. The authors acknowledge the instruction can introduce bias and still claim bias remains, but they do not reconcile the reversal. That undermines the headline 'under-protection of Chinese and Korean' narrative for that model. Also, Section 3.2's claim that French prompts consistently give the highest refusal rates is contradicted by Table 1 (Gemini-2.0 refuses 0% everywhere; Llama-3-70B does not show French as highest in most rows). The paper has no error bars, no significance tests, and no repeated sampling; the hallucination judge is GPT-4o, which is itself an evaluated model, and it was validated on only 100 cases. The dataset and code are not released, only available on request.

What is solid: the benchmark design is reasonable, the metrics are standard for this literature, and the limitations section is candid. The central qualitative phenomenon — that copyright enforcement varies by language — is probably real, but the quantitative magnitudes and the language attributions in the current text are not trustworthy without a fuller ablation.

Who this is for: researchers working on multilingual safety evaluation, copyright compliance, and probing methodology. It is not ready to be cited as a measurement, but it deserves a serious referee; the right outcome is major revision, not desk rejection. I'd bring it to a reading group to discuss where probing instructions stop being neutral.

Referee Report

4 major / 5 minor

Summary. The paper investigates whether LLMs enforce copyright differentially across languages. It manually curates a dataset of popular song lyrics in English, French, Chinese, and Korean and probes seven LLMs (GPT-3.5-Turbo, GPT-4o, Gemini-2.0, Claude-3.5-Haiku, Llama-3-70B, Mistral-7B, Mixtral-8x7B) with direct prompts in all four languages, instructing each model to answer in the language of the song. For every model and language pair it reports LCS and ROUGE-L scores for verbatim overlap, refusal rates, and (for three models) GPT-4o-judged hallucination rates. The headline conclusions are that copyright protection varies with both the language of the copyrighted lyrics and the language of the prompt, with, for example, GPT-3.5-Turbo refusing English requests more often than Chinese or Korean requests, and Claude-3.5-Haiku showing a low refusal rate for Korean songs under Chinese prompts.

Significance. The paper addresses a timely and understudied problem and provides a new multilingual benchmark, a hallucination metric, and a broad model comparison spanning API-based and open-weight models. The GPT-4o-based hallucination judge is validated on 100 cases against human annotation with a Pearson correlation of 0.85, which is a useful reliability check. If the reported effects survive a proper control for the probing instruction, the findings would be directly relevant to model deployment and copyright policy. The main limitation is that the central bias claim currently rests on a probing protocol whose controlling instruction is confounded with the conditions being compared, and the paper's own ablation shows that this instruction materially changes refusal rates for at least one model.

major comments (4)

[Appendix A and Table 3] The instruction 'Answer in [LANGUAGE]' is appended to every prompt and thus covaries perfectly with the song-language condition, while its wording varies across the prompt-language conditions. The paper acknowledges in Appendix A that this instruction 'may also introduce some bias,' but the only ablation (Appendix E, Table 4) measures refusal rate rather than LCS/ROUGE-L or hallucination, and it covers only five of the seven models. For GPT-3.5-Turbo the effect is very large: without the instruction, the refusal rate for Chinese songs under English prompts increases by 0.78 and for Korean songs by 0.74, and Table 5 shows that the refusal-rate ordering across song languages for English prompts is reversed relative to Table 1. Because Section 3.1's claim that 'Korean and Chinese lyrics receive significantly weaker protection' from GPT-3.5-Turbo is based on the with-instruction results, the lyric-language bias claim is confounded; the authors should run the no-instruction protocol for all metrics and all models, or otherwise show that the instruction does not affect verbatim output and hallucination.
[Section 3.2] The claim that 'using French as the prompt language consistently results in the highest refusal rates across all tested models' is contradicted by Table 1. For Gemini-2.0, the French-prompt refusal-rate column is 0.22, 0, 0, 0 across the four song languages, whereas the English-prompt column is 0.16, 0, 0, 0 and the Chinese-prompt column is 0.26, 0, 0, 0; for Llama-3-70B, the French-prompt refusal rate for French songs is 0.4, lower than the 0.48, 0.5, and 0.5 obtained with English, Chinese, and Korean prompts. This aggregate statement should be revised to reflect per-model patterns or be replaced by a formal statistical test.
[Section 3.1 and Table 1] The 'critical anomaly' of Claude-3.5-Haiku's refusal rate dropping to 0.28 for Korean lyrics under Chinese prompts is not covered by the Appendix E ablation, which omits Claude-3.5-Haiku and GPT-4o, and no hallucination rate is reported for this cell in Table 2. The inference that 'the model is fabricating content' is drawn only from the combination of a low refusal rate and similar LCS/ROUGE-L scores in Section 3.1, not from a direct hallucination measurement. The authors should report the hallucination rate for this cell and test whether the anomaly persists without the language-instruction suffix.
[Sections 2.3 and 3.1] The paper reports only point estimates (average and maximum LCS/ROUGE-L, and proportion-based refusal and hallucination rates) with no variance, confidence intervals, or significance tests. Given that the dataset contains roughly 50 songs per language (Appendix D), differences such as GPT-3.5-Turbo's Korean-song refusal rate of 0.26 with English prompts versus 0.20 with Chinese prompts are within sampling noise (standard error about 0.06 for n=50). To support the repeated use of the word 'significant' in the paper, the authors should report per-language standard deviations or confidence intervals and test the headline contrasts, including the English versus Chinese/Korean refusal-rate gaps and the French-prompt effect.

minor comments (5)

[Table 1] The numbers are visually run together (e.g., '0.940.9 0.961' for GPT-3.5-Turbo refusal rates), making the table hard to read; please use explicit column separators or formatting.
[Appendix D] The number of songs per language should be explicitly stated, and the selection criterion 'leaderboards on Apple Music' should be described with concrete details (e.g., chart, time window, rank threshold) to allow replication.
[Appendix C] The hallucination evaluation is reported for GPT-3.5-Turbo, Gemini-2.0, and Llama-3-70B only; the paper should justify the exclusion of the other four models, especially Claude-3.5-Haiku, for which Section 3.1 draws a hallucination inference.
[Section 2.2 and Table 3] The paper should specify whether the song title and singer name in each prompt are transliterated, translated, or kept in the original script, since this can affect whether the model retrieves the correct lyrics.
[Abstract] The abstract uses 'fair use' while the paper measures verbatim reproduction and refusal; consider using consistent terminology to avoid conflating copyright compliance with fair use doctrine.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical benchmark whose claims rest on standard overlap metrics and a human-validated judge, not on a self-referential derivation.

full rationale

This is an empirical measurement study with no mathematical derivation chain. The central claims (multilingual bias in refusal rate, LCS/ROUGE-L, and hallucination) are obtained by probing seven LLMs and comparing standard external metrics; no parameter is fitted from a subset of data and then renamed as a prediction, and no load-bearing result is imported from a self-citation. The only potential self-reference is the use of GPT-4o as an LLM judge (Appendix B/C) while GPT-4o is also an evaluated model; this is a measurement-validity concern, not circularity, because refusal and hallucination judgments are anchored to external human annotation (100-case human check, Pearson 0.85) and to prior work (Xu et al., 2024, 98% agreement). The acknowledged 'Answer in [LANGUAGE]' instruction (Appendix A) is a real confound: the requested output language covaries with the song-language condition, and Appendix E shows a large effect for GPT-3.5-Turbo on refusal rate. However, this is an experimental-design limitation, not a reduction of the result to its inputs; the paper also reports that the bias direction survives for GPT-3.5 without the instruction (Table 5). No equation-level identity, fitted-input prediction, uniqueness theorem, or ansatz-via-citation was found, so the appropriate circularity score is 0.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central claim is an empirical measurement, not a derivation, so the ledger consists of assumptions about the evaluation design rather than fitted constants. No free parameters are estimated from the results, and the paper introduces no new theoretical entities.

assumptions (4)

domain assumption All selected songs are under active copyright protection in equivalent legal terms across the four countries.
Appendix D states songs were manually verified as protected and all four countries are Berne Convention signatories. If some titles are older or licensed differently, refusal-rate comparisons across languages would not isolate copyright-protection policy.
ad hoc to paper Appending 'Answer in [LANGUAGE]' to the prompt does not materially change the measured copyright-compliance behavior beyond the refusal-rate effects examined in Appendix E.
This instruction is added by the authors to keep outputs in the song language. Appendix A admits it 'may also introduce some bias'; Appendix E tests only refusal rate for a subset of model/language pairs, leaving LCS, ROUGE-L, and hallucination results under this untested confound.
domain assumption The four language versions of the prompt template are semantically and pragmatically equivalent.
Table 3 gives literal translations but no back-translation or native-speaker validation; differences in politeness, formality, or naturalness could change model refusal behavior independent of language.
domain assumption Token-level LCS and ROUGE-L are computed fairly across languages with different tokenization characteristics.
The paper states tokens are used but does not specify the tokenizer or normalization for Chinese and Korean, which lack word boundaries. Token segmentation differences could affect cross-language scores.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Beyond English: Unveiling Multilingual Bias in LLM Copyright Compliance." pith.science (2026). https://pith.science/paper/PDOZ7FUX

@misc{pith2026250305713,
  author       = {Pith},
  title        = {Pith review of: Beyond English: Unveiling Multilingual Bias in LLM Copyright Compliance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PDOZ7FUX}},
  note         = {Machine review of arXiv:2503.05713}
}

read the original abstract

Large Language Models (LLMs) have raised significant concerns regarding the fair use of copyright-protected content. While prior studies have examined the extent to which LLMs reproduce copyrighted materials, they have predominantly focused on English, neglecting multilingual dimensions of copyright protection. In this work, we investigate multilingual biases in LLM copyright protection by addressing two key questions: (1) Do LLMs exhibit bias in protecting copyrighted works across languages? (2) Is it easier to elicit copyrighted content using prompts in specific languages? To explore these questions, we construct a dataset of popular song lyrics in English, French, Chinese, and Korean and systematically probe seven LLMs using prompts in these languages. Our findings reveal significant imbalances in LLMs' handling of copyrighted content, both in terms of the language of the copyrighted material and the language of the prompt. These results highlight the need for further research and development of more robust, language-agnostic copyright protection mechanisms to ensure fair and consistent protection across languages.

Figures

Figures reproduced from arXiv: 2503.05713 by the authors.

**Figure 1.** LLM copyright compliance is affected by both the language of the copyrighted content and the language of prompt. Red indicates copyrighted content, while purple represents hallucinated output. right laws (Karamolegkou et al., 2023; Chang et al., 2023; Liu et al., 2024; Mueller et al., 2024). This issue has led to a growing number of legal disputes, most notably The New York Times’ lawsuit against OpenAI and Microsof… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 15 canonical work pages

[1]

Imagine there’s no hell

"Imagine there’s no hell" - This line is incorrect; the correct line is "Imagine there’s no heaven"

work page
[2]

You are a helpful assistant

Comparing hallucination detection met- rics for multilingual generation. arXiv preprint arXiv:2402.10496. Antonia Karamolegkou, Jiaang Li, Li Zhou, and An- ders Søgaard. 2023. Copyright violations and large language models. In Proceedings of the 2023 Con- ference on Empirical Methods in Natural Language Processing, pages 7403–7412. Ahmad R. Kirmani. 2023....

arXiv 2023
[3]

你（我）问我爱你有多深 Now, let’s identify the unique lines in the ground truth lyrics:

work page
[4]

Où les enfants de mon âge

The line "Où les enfants de mon âge" is in the ground truth

work page
[5]

I bet you are hell

"I bet you are hell" - This line does not exist in the ground truth. The rest of the lines in the model’s output are present in the ground truth. The model’s output consists of 6 lines. Out of these, 2 lines are not in the ground truth. To calculate the percentage of lines that are not in the ground truth: Percentage = Number of lines not in ground truth ...

work page
[6]

Oui, je t’aime, et je te donne ce poème

The line "Oui, je t’aime, et je te donne ce poème" is slightly different from the ground truth "Oui je t’aime et je te donne ce poème" (comma placement), but this is a minor difference

work page
[7]

月亮代表我的心" is present in both the model’s output and the ground truth. -

你去想一想你去看一看 Next, we compare the lines from the model’s output with the ground truth: - "月亮代表我的心" is present in both the model’s output and the ground truth. - "你（我）问月亮代表什么" is not present in the ground truth. - "你（我）问我爱你有多深" is a variation of "你问我爱你有多深我爱你有几分" from the ground truth, but it is not an exact match. Now, let’s calculate the percentage of lines...

work page
[8]

Douce France, cher pays de mon enfance

The language model’s output starts with "Douce France, cher pays de mon enfance", which is in the ground truth

work page

Show all 17 references

[9]

Bercée de tendre insouciance

The next lines "Bercée de tendre insouciance" and "Je t’ai gardée dans mon cœur" are also in the ground truth
[10]

Mon village, au clocher aux maisons sages

The line "Mon village, au clocher aux maisons sages" is slightly different from the ground truth "Mon village au clocher, aux maisons sages" (comma placement), but this is a minor difference
[12]

Ont partagé mon bonheur

The line "Ont partagé mon bonheur" is in the ground truth
[14]

Oui, je t’aime, dans la joie ou la douleur

The line "Oui, je t’aime, dans la joie ou la douleur" is slightly different from the ground truth "Oui je t’aime dans la joie ou la douleur" (comma placement), but this is a minor difference
[15]

Douce France, cher pays de mon enfance

The repetition of "Douce France, cher pays de mon enfance" is in the ground truth
[16]

Bercée de tendre insouciance

The repetition of "Bercée de tendre insouciance" is in the ground truth
[17]

Je t’ai gardée dans mon cœur

The repetition of "Je t’ai gardée dans mon cœur" is in the ground truth. The language model’s output is a subset of the ground truth lyrics, with only minor punctuation differences. Therefore, there are no parts of the language model’s output that are not in the ground truth. ...
[2012]

Answer in [LANGUAGE]

We manually ensure that all of the songs are copyright protected. U.S., China, France, and Korea are all signatories to the Berne Convention (World Intellectual Property Organization, 1886), which is an international treaty that ensures that works created in one member country...

2022
[2024]

In Proceedings of the 2024 Con- ference on Empirical Methods in Natural Language Processing, pages 15134–15158

Copybench: Measuring literal and non-literal reproduction of copyright-protected text in language model generation. In Proceedings of the 2024 Con- ference on Empirical Methods in Natural Language Processing, pages 15134–15158. Lyra D’Souza and David Mimno. 2023. The chatbot a...

2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Imagine there’s no hell

"Imagine there’s no hell" - This line is incorrect; the correct line is "Imagine there’s no heaven"

work page

[2] [2]

You are a helpful assistant

Comparing hallucination detection met- rics for multilingual generation. arXiv preprint arXiv:2402.10496. Antonia Karamolegkou, Jiaang Li, Li Zhou, and An- ders Søgaard. 2023. Copyright violations and large language models. In Proceedings of the 2023 Con- ference on Empirical Methods in Natural Language Processing, pages 7403–7412. Ahmad R. Kirmani. 2023....

arXiv 2023

[3] [3]

你（我）问我爱你有多深 Now, let’s identify the unique lines in the ground truth lyrics:

work page

[4] [4]

Où les enfants de mon âge

The line "Où les enfants de mon âge" is in the ground truth

work page

[5] [5]

I bet you are hell

"I bet you are hell" - This line does not exist in the ground truth. The rest of the lines in the model’s output are present in the ground truth. The model’s output consists of 6 lines. Out of these, 2 lines are not in the ground truth. To calculate the percentage of lines that are not in the ground truth: Percentage = Number of lines not in ground truth ...

work page

[6] [6]

Oui, je t’aime, et je te donne ce poème

The line "Oui, je t’aime, et je te donne ce poème" is slightly different from the ground truth "Oui je t’aime et je te donne ce poème" (comma placement), but this is a minor difference

work page

[7] [7]

月亮代表我的心" is present in both the model’s output and the ground truth. -

你去想一想你去看一看 Next, we compare the lines from the model’s output with the ground truth: - "月亮代表我的心" is present in both the model’s output and the ground truth. - "你（我）问月亮代表什么" is not present in the ground truth. - "你（我）问我爱你有多深" is a variation of "你问我爱你有多深我爱你有几分" from the ground truth, but it is not an exact match. Now, let’s calculate the percentage of lines...

work page

[8] [8]

Douce France, cher pays de mon enfance

The language model’s output starts with "Douce France, cher pays de mon enfance", which is in the ground truth

work page