{"id":"f7a38d20-2ac5-47d3-9be2-bea26ae41e94","arxiv_id":"2506.20695","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"TikTok's sound infrastructure lets far-right actors keep xenophobic audio memes online by cloaking them in benign trends, and even explicitly hateful posts disappear at low rates.","lead":"This study shows how far-right TikTok users embed xenophobic messages in popular sounds and meme templates, often with hateful lyrics hidden in intros or remixes. The result matters for content moderation because these posts survive largely unmoderated, even when they violate platform rules.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Moderation-trace comparison conflates platform takedowns with user deletions and ignores account-level clustering; the 85%/75% vs 66%/53% gap is not yet evidence that text hate is moderated more than audio hate.","rationale":"The reader's weakest assumption is correct and is the load-bearing point. I agree with the CONDITIONAL verdict. The central descriptive finding about sound-based cloaking is well supported by the walkthrough and network examples, but the quantitative trace comparison is over-interpreted. The authors' own caveat shows they know the measure is noisy; since the conclusion about platform moderation depends on distinguishing who removed content, the study needs additional checks before the numeric comparison is cited. Account-level clustering is a further reason the post-level percentages may mislead. These concerns are not fatal to the descriptive contribution, so no stronger verdict is warranted.","tokens_in":23275,"tokens_out":7503,"duration_ms":92148,"concrete_test":"Run an account-level, matched-control audit: recheck a new sample of 'Musikgeschmack ist wichtig' posts from the same dates, split into (a) benign songs, (b) Türke intro with cloaked memes, and (c) Türke intro with explicit textual hate, matched on plays, followers, and posting date, and fit a mixed-effects logistic model of URL availability with random intercepts for accounts. If the explicit-text coefficient is not significant after controlling for account, the moderation-difference claim is an artifact of account deletion; if the benign control disappears at the same rate, the claim that text hate is specifically moderated fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Figure 6.4's headline numbers—85%/75% of cloaked meme posts online vs 66%/53% of explicit-text posts online—are the quantitative support for the claim that textual hate predicts disappearance better than audio hate. The problem is that 'unavailable' or 'account gone' bundles three different mechanisms: platform removal, user-led deletion, and privacy setting changes. The authors explicitly acknowledge this limitation, yet the policy conclusion ('scarcely moderated', 'textual hate is a greater predictor') requires the platform-removal interpretation. The comparison is also computed at the post level although posts are nested in accounts; a single deleted account carrying several explicit-text posts can move the 66%/53% figures without any systematic content-level moderation. No control group of comparable benign posts from the same 'Musikgeschmack ist wichtig' trend is reported, so there is no baseline natural-attrition rate. The qualitative walkthrough (search results, sound pages, remix dendrogram) does support the descriptive claim that xenophobic audio memes are visible and persist; what the trace cannot support is the attribution of that persistence to weak platform moderation of audio.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies how TikTok's sound infrastructure, particularly the 'use this sound' button and sound pages, enables far-right actors in Germany to spread xenophobic content in forms that are difficult to moderate. Based on persona-based feed observations, hashtag-based collections, sound search walkthroughs, and repeated URL status checks of posts using the 'Türke' and 'Zecken' sounds, the paper argues that explicit textual hate is more likely to disappear than audio-based hate embedded in cloaked memes, and that moderation of such content is minimal. It also highlights DSA transparency limits that prevent researchers from attributing takedowns to platform moderation versus user deletion.","tokens_in":23374,"tokens_out":5282,"duration_ms":59345,"significance":"The study's qualitative and network findings are a useful addition: they concretely show how benign-sounding audio trends, such as motorbike videos and the 'music taste is important' meme, become carriers of racist content, and how the sound page infrastructure creates visible entry points even for banned search terms. The authors are appropriately transparent about not being able to verify platform removal at the level of individual posts. If the moderation-attribution claims are set aside, the descriptive account of persistence is credible and policy-relevant. The quantitative trace conclusions, however, are currently overstated and require revision before the claims about TikTok's moderation of audio hate can be accepted.","major_comments":[{"comment":"The trace analysis cannot distinguish platform moderation from user-led deletion or privacy changes, and the manuscript explicitly concedes this. In this situation, the 85%/75% versus 66%/53% gap for Türke and the analogous Zecken figures are not evidence that textual hate is moderated more than audio hate: they are raw non-availability rates. The comparison is computed at the post level even though posts are nested in accounts, so a single deleted account hosting several explicit-text posts can move the 66%/53% figures without any content-level enforcement. Moreover, no baseline is reported for natural attrition of comparable benign posts from the same 'Musikgeschmack ist wichtig' trend. The descriptive persistence claim survives, but the load-bearing claim that textual hate is a greater predictor and the phrase 'scarcely moderated' need to be re-expressed as bounds or subjected to a sensitivity analysis, for example by excluding account-deleted posts, reporting account-level clustered rates, and comparing against a benign control.","section":"Moderation traces: minimal ‘disappearances’ (Figures 6.4 and 6.5)"},{"comment":"The statement that deleted accounts 'point to instances of deplatforming rather than users voluntarily deleting their social capital' is an unsupported assumption that is load-bearing for any platform-enforcement interpretation. If account deletion is treated as evidence of deplatforming, the paper should show that these accounts were the target of enforcement, for instance by matching deletion timing with TikTok's DSA statement-of-reasons data or by examining whether deletion coincides with reported enforcement waves. In its current form, the sentence should be removed or downgraded to a hypothesis.","section":"Moderation traces: minimal ‘disappearances’"},{"comment":"The abstract states categorically that songs with hateful lyrics 'are not eligible for personalized feeds', but the quoted community guideline says content 'may be ineligible' for the For You Feed when it 'indirectly demeans protected groups'. Because eligibility is not an observed enforcement outcome, the categorical phrasing overstates what the paper can conclude. Use the hedged formulation throughout, including in the conclusion.","section":"Findings / TikTok community guidelines quote"}],"minor_comments":[{"comment":"In the paragraph reporting Zecken results, the parenthetical '(left diagram in figure 6.5)' is inconsistent with the preceding sentence, which assigns the entire dataset to the left and textual-hate posts to the right; please correct the reference.","section":"Moderation traces: minimal ‘disappearances’ (Figure 6.5)"},{"comment":"The sound queried in this section is called 'Anotha Europe' earlier and 'Another Europe' in the Figure 6.3 caption; use one spelling consistently.","section":"Soundscapes as dispersed problematic niches"},{"comment":"The persona-based feed analysis is reported only as counts in Figure 6.1; the text would benefit from stating the number of feed sessions, the duration of scrolling, and the date range, so that the comparison across the UK, Germany, and the Netherlands is interpretable.","section":"Methods"},{"comment":"In Figure 6.4, the left diagram's 'hampelmänner' category is said to appear 'better moderated', but the manuscript correctly notes this could reflect account deletion driven by other posts on those accounts; this caveat should also be applied to the overall 85%/75% figure.","section":"Moderation traces: minimal ‘disappearances’"}],"recommendation":"major_revision","confidential_remarks":"The editor may want to ask the authors to add a scope note to the final version clarifying that 'disappearance' is not synonymous with 'moderation', and to consider whether the trace analysis can be re-run with account-level clustering and a benign control. The chapter fits the edited volume, but the categorical abstract claims are likely to be quoted out of context if left uncorrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful part of this paper is the descriptive infrastructure work. The authors show concretely how TikTok's 'use this sound' button and sound pages let xenophobic audio memes persist under benign trends, and they name three 'cloaking while amplifying' strategies. The German election case (Türke, Zecken, the #1161 network) is fresh, and the walkthrough observations are credible. This is a solid qualitative contribution to platform governance debates, especially the DSA transparency point: TikTok's Statements of Reasons don't let researchers link takedowns to specific posts.\n\nThe soft spot is the moderation trace analysis. The 85%/75% versus 66%/53% gap is presented as evidence that explicit textual hate disappears more than cloaked audio hate, and the authors lean on this to say niche hateful soundscapes are 'scarcely moderated.' But they admit they can't distinguish platform removal from user deletion or privacy changes, and they give no baseline for natural attrition of comparable benign posts. Posts are nested in accounts, so one deleted account with several explicit-text posts can move the percentages. That doesn't kill the paper, but it means the quantitative trace can't carry the moderation-attribution claim. The paper would be stronger if it treated 'disappearance' as a descriptive outcome and explicitly refrained from calling it moderation except where a deleted account strongly suggests deplatforming.\n\nMinor: sample sizes are small (65 and 38 for the two Türke groups), and no dataset or codebook is released, which limits reproducibility. Some figures are not fully inspectable from the text.\n\nOverall, the descriptive claims about sound infrastructure and cloaking are supported. The moderation trace is suggestive, not confirmatory. The paper deserves a serious referee; it's a useful case study for people working on platform moderation and far-right audio memes, and the DSA transparency angle is timely. I'd ask for a revision that either gets platform-side data (SoR linkage) or softens the moderation language and adds a disappearance baseline.","headline":"Solid descriptive infrastructure analysis of far-right audio memes, but the moderation-trace comparison overreaches beyond what the data can distinguish.","tokens_in":23975,"tokens_out":1768,"would_cite":true,"duration_ms":19111,"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":"TikTok's one-tap sound button keeps far-right hate in memes alive.","keywords":["TikTok","audio memes","content moderation","far-right extremism","cloaked hate","soundscapes","moderation trace analysis","algospeak"],"falsifier":"Re-run the same URL trace on a matched set of benign sound pages with similar creator sizes and engagement levels; if those benign posts disappear at the same or higher rates, the claim that hateful audio is lightly moderated would no longer be supported.","tokens_in":22983,"feed_emoji":"🎵","tokens_out":9574,"duration_ms":91633,"temperature":0.7,"pith_summary":"This paper argues that TikTok's sound infrastructure, especially user-uploaded 'original sounds' and the one-tap 'use this sound' button, gives far-right actors a durable channel for xenophobic content that largely escapes moderation. Studying posts connected to the 2024 German state elections, the authors traced whether videos and accounts disappeared from public view after two and four months. They found that cloaked hate survived: 85 percent in October and 75 percent in December of posts using a hateful song's intro inside an ambiguous meme template remained online, while only 66 percent and 53 percent of posts with explicit textual hate did. Even posts whose audio carried explicit lyrics about killing immigrants stayed online in 89 and 86 percent of cases overall. The result matters because it identifies a structural blind spot: moderation that keys on text will miss hate that travels through sound and memetic association.","feed_headline":"TikTok's one-tap sound button keeps far-right hate in memes alive","feed_subtitle":"Cloaked audio hate stayed online while text hate vanished at higher rates during Germany's 2024 elections.","key_machinery":"The load-bearing mechanism is the 'original sound' plus 'use this sound' button pairing. Every upload of a user-generated audio registers as an 'Original Sound' attributed to that user, any other user can replicate it instantly, and all posts using that sound collect on a hyperlinked sound page. The authors treat these pages as soundscapes, distributed, affectively held-together environments where communities convene through sound replication. This mechanism carries the argument because it explains both amplification, since a hateful or hijacked sound can spread faster than moderation can track it, and obfuscation, since the same sound reappears in remixes, speed-ups, slowdowns, or under new names after a deplatforming.","core_discovery":"The central claim is that TikTok's sound architecture does not merely host far-right extremism; it actively affords its proliferation and obfuscation. A sound uploaded once becomes an 'original sound' that any user can replicate with one click, and each replicated post is indexed on a sound page that acts as a gathering place for a community. This lets extremists attach racist lyrics to a beloved club track, play only the intro of a hateful song inside a benign trend, or remix an innocent folk tune into a song comparing immigrants to ticks. The authors' moderation traces show that explicit hate in text predicts disappearance better than explicit hate in audio: posts in the Türke soundscape with textual hate dropped to 66 and 53 percent availability, whereas posts with the same audio but hate cloaked in memes stayed at 85 and 75 percent. They also found that the platform rarely intervenes in niche soundscapes, partly because such posts are ineligible for personalized feeds, attract low engagement, and thus are almost never flagged. The conclusion is that audio-based hate is minimally moderated and continues to sustain extremist communities in the undercurrents of the platform.","pith_inferences":["Because the authors cannot distinguish user deletions from platform removals, their 'moderation traces' conflate the two; a platform-side dataset with post-level identifiers would be needed to settle how much of the disappearance is actually TikTok acting.","A direct testable extension would be to run the same four-month URL trace on a matched set of benign sound pages; if benign posts disappear at comparable rates, the low disappearance of hateful audio would reflect general platform dynamics rather than a specific tolerance for hate.","The findings suggest transparency reports under the DSA should expose post-level identifiers; without them, researchers cannot audit whether taken-down posts match platform statements, and the paper's central comparison remains an indirect proxy.","As multimodal moderation improves, the same cloaking logic may shift further into latent audio cues such as melody, rhythm, or remix structure, where semantic hate is even harder to isolate; the paper's soundscape concept gives a unit for studying that shift."],"forward_implications":["Platform moderation that searches text and visible overlays will systematically under-detect hate that lives in audio, because audio hate in the traced soundscapes disappeared at lower rates than textual hate.","Even when an original uploader is banned, the sound they posted can keep circulating through replicate posts and re-uploads, so removing one account does not dismantle the soundscape.","Users searching for benign trends, such as the 'music taste is important' meme, can surface cloaked far-right posts among the first results, because hateful intros ride along with popular templates.","Low-engagement niche hate posts are unlikely to be flagged, which means TikTok's reliance on user reporting leaves audio hate in a moderation gap.","Songs with hateful lyrics remain online and can appear in search results when they intersect with benign meme trends, even if they are ineligible for personalized feeds."],"supporting_citations":[{"why":"Defines TikTok as an imitation public built on technological mimesis, grounding the claim that sound replication is the platform's core affordance.","marker":"Zulli & Zulli (2022)"},{"why":"Introduces audio memes and templatability, providing the vocabulary for how sounds spread as memetic units.","marker":"Abidin (2021)"},{"why":"Shows how algospeak and multimodal communication circumvent moderation, supporting the cloaking mechanisms documented here.","marker":"Steen et al. (2023)"},{"why":"Supplies the moderation trace analysis method used to measure whether posts and accounts disappear over time.","marker":"De Keulenaar et al. (2023)"},{"why":"Supplies the scraping tool used to collect post metadata and sound pages from TikTok.","marker":"Peeters (2021)"},{"why":"Develops the research persona method used to compare far-right content in personalized feeds across countries.","marker":"Bounegru & Weltevrede (2022)"},{"why":"Defines soundscapes as affective assembly environments, the central unit for analyzing sound-based communities.","marker":"Geboers & Pilipets (2024)"},{"why":"Documents how original sounds are deleted and reposted to conceal orchestration, underpinning the re-emergence finding.","marker":"Bösch & Divon (2024)"},{"why":"Reviews multimodal and large-model hate speech moderation, framing the limits of audio detection that the paper exposes.","marker":"Hee et al. (2024)"},{"why":"Provides the engagement-over-well-being explanation for why platforms under-moderate content that drives metrics.","marker":"Christin et al. (2024)"}],"fun_headline_variants":["One-click sound button hides far-right hate in benign memes","Audio hate survives moderation better than text on TikTok","TikTok's use-this-sound button lets far-right memes hide in plain sight","One-tap sound keeps extremist hate online when text gets removed","Audio-based hate stays, text hate goes: TikTok's sound loop"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central argument treats a post or account disappearing from public view as evidence about TikTok's moderation, even though users can delete their own content, and the paper offers no baseline for how often comparable benign posts vanish.","fun_headline_variants_meta":{"raw":{"variants":["One-click sound button hides far-right hate in benign memes","Audio hate survives moderation better than text on TikTok","TikTok's use-this-sound button lets far-right memes hide in plain sight","One-tap sound keeps extremist hate online when text gets removed","Audio-based hate stays, text hate goes: TikTok's sound loop"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000637,"raw_usage":{"total_tokens":2959,"prompt_tokens":995,"completion_tokens":1964,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":1873}},"tokens_in":611,"tokens_out":1964,"duration_ms":15715,"temperature":1.0,"reasoning_tokens":1873,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:51:59.811130+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same URL trace on a matched set of benign sound pages with similar creator sizes and engagement levels; if those benign posts disappear at the same or higher rates, the claim that hateful audio is lightly moderated would no longer be supported.","supporting_citations":[{"cited_title":"Recent Advances in Hate Speech Moderation: Multimodality and the Role of Large Models","cited_arxiv_id":"2401.16727","evidence_quote":"Reviews multimodal and large-model hate speech moderation, framing the limits of audio detection that the paper exposes."}],"review_version":1}