{"id":"0fdee524-5422-4bf7-97c7-759479f4daf9","arxiv_id":"2501.13836","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Automated content moderation for low-resource languages fails not just from data scarcity but from systemic inequities including data monopolies, underinvestment, and English-centric design.","lead":"Researchers interviewed 22 AI experts who build automated moderation tools for Tamil, Swahili, Maghrebi Arabic, and Quechua. They found that beyond data scarcity, data monopolies, underinvestment, and English-centric design create systemic moderation failures for low-resource languages.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that moderation failures are 'rooted in colonial suppression' is an author-imposed interpretive frame rather than an empirical finding, and the 'beyond data scarcity' counterfactual is asserted, not tested.","rationale":"The reader identified the sample's reliability and the interpretive leap to colonialism as the weakest assumptions. I agree that the colonial-suppression causal claim is the most load-bearing part of the paper's headline contribution, but I would locate the problem more precisely: the interview quotes in Section 4 support claims about corporate incentives, data gatekeeping, and technical design, yet no participant is quoted as attributing these to colonial history. The colonial lens is introduced in Section 5.1 entirely through theoretical literature, making the abstract's 'rooted in colonial suppression' a conclusion the data cannot bear without additional evidence. The paper is valuable as a qualitative account of barriers and as a demonstration that coloniality is a useful interpretive frame, and it already includes a positionality statement that partially addresses researcher stance. However, the causal framing in the abstract and conclusion overstates what the method can establish. The proposed concrete test—a blinded re-coding of the transcripts—would directly reveal whether the colonial attribution is grounded in participant discourse. I recommend no change to the reader's CONDITIONAL verdict: the empirical core is solid, but the central claim needs reframing or additional evidentiary support. I marked agreement as 'partial' because the reader emphasized sample-size limitations and Western affiliations, while my concern targets the causal inference as the more decisive weakness; the sample limitation matters for generalizability, but even a larger sample would not justify the colonial causal claim without stronger evidence.","tokens_in":20074,"tokens_out":5258,"duration_ms":50817,"concrete_test":"Run a blinded thematic analysis of the 22 interview transcripts with the coloniality framework withheld from coders: code every participant-stated causal attribution and count how many invoke colonial history or suppression. If participant-attributed colonial causes are zero or negligible, the conclusion that limitations are 'rooted in colonial suppression' is not grounded in the data. Complement this with a matched-data-size hate-speech benchmark (e.g., Tamil vs. English at 5k/10k/20k training examples using the same mBERT head); if the performance gap vanishes at matched sizes, the 'beyond data scarcity' counterfactual weakens.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and conclusion assert that moderation failures in low-resource languages are 'not just technical gaps caused by data scarcity' but are 'rooted in colonial suppression of non-Western languages.' The interview data in Section 4 richly document barriers—data access restrictions, annotation underfunding, tokenizer/stemmer failures, and model biases—but no participant in the quoted evidence attributes these problems to colonial history. The coloniality framework enters only in Section 5.1 as the authors' theoretical lens, citing Quijano, Kwet, and others. This does two things. First, it repackages an analytic frame as a finding, so the strongest causal claim is author-imposed rather than empirically grounded. Second, the 'even if more data were available' counterfactual is supported only by anecdotal examples (e.g., Mulaicchu→Mulai in §4.3) and participants' beliefs, not by any controlled comparison of performance at matched data sizes. Readers cannot distinguish observed barriers from the counterfactual claim that these barriers would persist with abundant data. The paper would be defensible if it proposed coloniality as one possible interpretation, but as written it asserts causality.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a qualitative interview study of 22 AI researchers and practitioners who work on harmful content detection in four low-resource languages: Tamil, Swahili, Maghrebi Arabic, and Quechua. Using semi-structured interviews and reflexive thematic analysis, the authors document barriers across data curation, annotation, preprocessing, and model training, and argue that these barriers are not reducible to data scarcity. They interpret the findings through a coloniality lens, claiming that moderation failures are rooted in structural inequities and colonial suppression of non-Western languages, and they propose multi-stakeholder recommendations for improving moderation. The paper contributes an empirical account of systemic pipeline-level challenges and connects them to decolonial theory.","tokens_in":20233,"tokens_out":3270,"duration_ms":32571,"significance":"If the findings hold, the paper makes a useful contribution by moving the discussion of low-resource content moderation beyond a narrow data-scarcity framing. Its strengths include rich interview material with concrete examples across four typologically diverse languages, a reflexive positionality statement, and an explicit attempt to connect technical pipeline decisions to historical and political contexts. The recommendations in Section 5.2 are concrete and grounded in the participants' accounts. However, the central causal claim that the problems are 'rooted in colonial suppression' is asserted more strongly than the empirical evidence supports, and the sample is small and uneven across the four languages. The paper is publishable after the framing and evidentiary status of that causal claim are clarified.","major_comments":[{"comment":"The abstract and Section 5.1 claim that moderation failures are 'rooted in colonial suppression of non-Western languages,' but the interview evidence in Section 4 documents contemporary structural barriers without any participant attributing those barriers to colonial history, and the coloniality framework is introduced as the authors' analytic lens in Section 5.1 rather than as a theme that emerged from the data. This is load-bearing because it is the paper's central contribution. Please either reframe the causal claim as an interpretive argument grounded in the cited decolonial literature, or provide explicit evidence linking participants' accounts to colonial legacies.","section":"Abstract and §5.1"},{"comment":"The systemic conclusions are drawn across four language communities from a sample that includes only three Quechua experts, with half of the participants affiliated with Western institutions, and the paper does not discuss how this imbalance constrains the cross-language claims. The Quechua-related findings in Sections 4.1 and 4.3 rest on very few voices. Please either narrow the scope of the cross-language conclusions, provide a per-language saturation assessment, or explicitly analyze how the uneven sample limits the generality of the claims.","section":"§3 (Participants) and §5.1"},{"comment":"The 'even if more data were available' counterfactual, which is central to the paper's 'beyond data scarcity' thesis, is supported only by isolated examples such as the Tamil stemming error Mulaicchu→Mulai and by participants' beliefs, not by any controlled comparison of model performance at matched data sizes across languages. Please present this counterfactual explicitly as a hypothesis or as participant perception, or add comparative evidence, so that readers can distinguish observed barriers from the authors' interpretation.","section":"§4.3 and Abstract"}],"minor_comments":[{"comment":"The paper does not include the interview protocol or a description of how saturation was determined; adding these would strengthen reproducibility and trustworthiness.","section":"§3 (Data Collection)"},{"comment":"Participants P7 and P8 are listed as specializing in 'Indic languages' rather than one of the four focal languages, which is inconsistent with the claim that all 22 participants specialize in Tamil, Swahili, Maghrebi Arabic, or Quechua; please clarify.","section":"Table 2"},{"comment":"The Tamil example is spelled 'Mualichhu' in the abstract and 'Mulaicchu' in Section 4.3; please make the transliteration consistent throughout.","section":"Abstract and §4.3"},{"comment":"Interviews were conducted in English even for participants whose native language is one of the focal languages; this should be acknowledged as a potential limitation for capturing in-group linguistic and cultural concepts.","section":"§3 (Data Collection)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a critical AI/ethics venue and the empirical material is valuable. The main issue is that the causal coloniality claim is presented as a finding rather than as an interpretive framing; this is fixable by softening the abstract/conclusion language and clarifying the status of the counterfactual. I do not see a need for rejection, but the revision should be substantive rather than cosmetic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth reading for the empirical material. The authors interviewed 22 practitioners and researchers working on Tamil, Swahili, Maghrebi Arabic, and Quechua, and they report concrete barriers across the whole pipeline—data access, annotation, tokenization, stemming, model training. The Section 4 quotes are illustrative and often specific (the Tamil 'Mulaicchu/Mulai' stemming failure, the Perspective API misses on Swahili, the code-mixed hashtag segmentation). That granularity is the contribution; prior work made the coloniality critique conceptually, but this gives it field-level texture.\n\nThe methods are standard for qualitative HCI: purposive and snowball sampling, reflexive thematic analysis, 441 codes merged into 23 subthemes. They also provide a positionality statement that acknowledges the authors' own institutional privileges and the diversity among participants. That is honest and useful.\n\nThe soft spot is the causal frame, mainly in the abstract and conclusion. The interviews document a long list of barriers, some technical, some socio-economic (data monopolies, underfunding, English-centric design). But no participant is quoted as attributing those barriers to colonial suppression; that link is made by the authors in Section 5.1 when they import Quijano, Kwet, and the rest. Presenting coloniality as the root cause is a reasonable interpretive move, but the paper states it as a finding. It isn't one, at least not on this evidence. The same goes for the 'even if more data were available' counterfactual: there's anecdotal support, but no matched comparison, so it's an assertion about what would happen, not a demonstrated result.\n\nSample limits are real but minor for a qualitative study: 22 participants, only three Quechua experts, many Western-affiliated, interviews in English. That tempers how far the results generalize, but it doesn't sink the descriptive findings.\n\nIf I were refereeing, I'd ask for two changes: reframe the coloniality discussion as one possible interpretation rather than the conclusion, and soften the abstract's causal language. The empirical core is solid and the recommendations (DSA data access, data donation, local capacity building) are grounded in the data and the literature.\n\nThis deserves a serious referee. It's a useful contribution to a live debate in platform governance and low-resource NLP, and the flaws are fixable without new data collection.","headline":"A genuinely useful interview study of moderation barriers in four low-resource languages, wrapped in a causal claim about colonial suppression the data don't actually demonstrate.","tokens_in":20814,"tokens_out":2113,"would_cite":true,"duration_ms":19337,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Automated moderation failures in low-resource languages stem from colonial-era inequities, the paper argues, not just missing data.","keywords":["automated content moderation","low-resource languages","coloniality","data scarcity","code-mixing","agglutinative morphology","Global South","qualitative interviews"],"falsifier":"A controlled benchmark would settle the technical core: if, at matched data sizes, a linguistically motivated morphological tokenizer does not improve harmful-content detection over a frequency-based tokenizer for Tamil, Swahili, Maghrebi Arabic, or Quechua, the claim that English-centric preprocessing drives moderation failure in agglutinative languages would be undermined. Separately, if platforms granted vetted researchers full data access in one low-resource language and moderation accuracy did not improve, the data-monopoly link would weaken.","tokens_in":19833,"feed_emoji":"🌐","tokens_out":8165,"duration_ms":65286,"temperature":0.7,"pith_summary":"This paper argues that automated moderation fails for Tamil, Swahili, Maghrebi Arabic, and Quechua not primarily because these languages lack data, but because the entire pipeline—data access, annotation, preprocessing, and model training—is built around English and Western market priorities. Based on semi-structured interviews with 22 AI experts working on harmful-content detection in these languages, it identifies how tech companies' monopoly on user data, weak financial incentives for Global South markets, and reliance on biased machine translation and outdated corpora reproduce historical colonial hierarchies. The paper treats these failures as structural rather than technical, and proposes multi-stakeholder fixes: funding local research capacity, democratizing data access, and adopting language-aware methods such as morphological tokenization. A sympathetic reader would take the central claim as a reframing of 'data scarcity' from a neutral technical condition into a consequence of political and economic choices.","feed_headline":"Low-resource language moderation fails on colonial bias, not just data","feed_subtitle":"22 AI experts trace Tamil, Swahili, Arabic, Quechua moderation failures to data monopolies and English-centric design.","key_machinery":"The carrying object is the automated moderation pipeline, broken into four stages: data curation, annotation, preprocessing, and model training. The load-bearing linguistic mechanism is agglutinative morphology combined with code-mixing: Tamil, Swahili, Maghrebi Arabic, and Quechua build thousands of words from a single root, so frequency-based tokenizers, stemming, and normalization designed for English systematically mangle the forms that carry harmful meaning. The interpretive machinery is the coloniality lens, which converts observed pipeline failures into symptoms of persistent power asymmetries—data monopolies, profit-driven neglect of Global South markets, and English-centric model design—rather than neutral technical gaps.","core_discovery":"The central discovery is that moderation failures in low-resource languages persist even when data volume is not the binding constraint; the binding constraints are who controls the data, who defines harm, and which linguistic features the tools are designed to see. The paper documents how English-centric frequency-based tokenizers split agglutinative words incorrectly—for example, the Tamil word Mulaicchu (meaning 'nipples') can be reduced to Mulai (meaning 'sprout')—so that sexually harassing language evades detection; how language-identification tools mangle code-mixed text; and how toxicity models trained on Western media associate Arabic phrases like 'Allahu Akbar' with terrorism. The authors then read these findings through coloniality: data monopolies, reliance on colonial-era texts for Quechua, underfunded annotation, and the flattening of annotator diversity into a single label are continuous with colonial suppression of non-Western languages. The paper's conclusion is that techno-solutionist fixes that only add more data will not yield equitable moderation; the pipeline itself must be redesigned around local linguistic knowledge and community self-determination.","pith_inferences":["If the paper's diagnosis is right, the same pipeline critique should apply to other morphologically rich low-resource languages beyond the four studied, and would predict that generic multilingual models underperform linguistically motivated tools there too.","The paper's account implies a testable asymmetry: for low-resource languages with simple morphology, data scarcity should dominate failure, whereas for agglutinative languages, architecture mismatches should dominate even at matched data sizes.","A concrete policy experiment follows: if platforms opened vetted data access to Global South researchers, hate-speech detection corpora and models should improve faster for those languages than equivalent private investment in more data alone would achieve.","A companion large-scale audit of platform transparency reports, comparing moderation accuracy across languages and regions, could test whether the structural inequities reported by the 22 experts hold beyond their experiences."],"forward_implications":["If the diagnosis is right, adding more labeled data alone will not fix moderation for these languages; data access, annotation incentives, and model architecture must change together.","Language-aware preprocessing—morphological segmenters, rule-based translation, and code-mixed language identification—should outperform generic frequency-based tokenizers and multilingual models on harmful-content detection for agglutinative languages.","Tech companies' API restrictions and shutdown of public research tools will continue to block academic study of evolving hate speech in the Global South unless policy such as the Digital Services Act grants vetted researchers data access.","Regulatory pressure requiring local moderators, locally defined benchmarks, and reporting of recall and precision for low-resource languages would surface failures that accuracy metrics hide.","Investment in grassroots research ecosystems and equitable data-sharing could transfer capacity to local researchers and reduce dependence on Western computational resources."],"supporting_citations":[{"why":"Sets up the received view that low-resource moderation fails from data scarcity and market disinterest, which the paper contests and extends.","marker":"Nicholas and Bhatia 2023"},{"why":"Documents how economic and geopolitical inequalities shape which regions receive moderation resources.","marker":"De Gregorio and Stremlau 2023"},{"why":"Argues that 'low-resource' status is produced by underinvestment, grounding the paper's structural rather than technical frame.","marker":"Nigatu et al. 2024"},{"why":"Provides the digital-colonialism analogy linking platform data control to colonial infrastructure extraction.","marker":"Kwet 2019"},{"why":"Shows how platforms extract Global South user data, supporting the data-monopoly claim.","marker":"Coleman 2018"},{"why":"Establishes that NLP models are not language-independent and encode assumptions favoring English, anchoring the English-centric design critique.","marker":"Bender 2009"},{"why":"Prior decolonial study of moderation that supplies the Western-values-as-global-standards argument the paper extends to the pipeline.","marker":"Shahid and Vashistha 2023"},{"why":"Supplies the coloniality-of-power framework used to interpret pipeline failures as persistent colonial hierarchies.","marker":"Quijano 2007a"}],"fun_headline_variants":["Colonial bias, not data scarcity, drives moderation failures in low-resource languages","Moderation fails for Tamil, Swahili, Arabic, Quechua due to colonial pipeline design","Why low-resource language moderation fails: it's colonial, not just data","English-centric AI tools let harassment slip through in low-resource languages","Tech data monopolies and colonial bias undermine moderation for 4 languages"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on treating the self-reported experiences of 22 purposively sampled AI experts—only three of them Quechua specialists and many based in Western institutions—as reliable evidence for how entire moderation pipelines behave across four language communities, and for the causal interpretation that these failures are rooted in colonial suppression.","fun_headline_variants_meta":{"raw":{"variants":["Colonial bias, not data scarcity, drives moderation failures in low-resource languages","Moderation fails for Tamil, Swahili, Arabic, Quechua due to colonial pipeline design","Why low-resource language moderation fails: it's colonial, not just data","English-centric AI tools let harassment slip through in low-resource languages","Tech data monopolies and colonial bias undermine moderation for 4 languages"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000747,"raw_usage":{"total_tokens":3354,"prompt_tokens":995,"completion_tokens":2359,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":2256}},"tokens_in":611,"tokens_out":2359,"duration_ms":14700,"temperature":1.0,"reasoning_tokens":2256,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:32:47.669581+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled benchmark would settle the technical core: if, at matched data sizes, a linguistically motivated morphological tokenizer does not improve harmful-content detection over a frequency-based tokenizer for Tamil, Swahili, Maghrebi Arabic, or Quechua, the claim that English-centric preprocessing drives moderation failure in agglutinative languages would be undermined. Separately, if platforms granted vetted researchers full data access in one low-resource language and moderation accuracy did not improve, the data-monopoly link would weaken.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Sets up the received view that low-resource moderation fails from data scarcity and market disinterest, which the paper contests and extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the digital-colonialism analogy linking platform data control to colonial infrastructure extraction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior decolonial study of moderation that supplies the Western-values-as-global-standards argument the paper extends to the pipeline."}],"review_version":1}