{"id":"9ab6f47d-a141-4d0a-a3e0-d57413d04391","arxiv_id":"2505.08385","paper_version":2,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"TikTok's prewritten search recommendations create governance, transparency, and research challenges that the paper documents and frames as a research agenda.","lead":"This position paper examines TikTok's 'search recommendations': prewritten search queries placed on videos that push users toward specific topics. It argues these recommendations create transparency and accountability gaps that regulators and researchers have not yet addressed.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'neutral aggregator' charge may overread TikTok's documentation, but the governance agenda does not depend on it; no verdict change needed.","rationale":"The reader's weakest assumption identifies the reliability of TikTok's documentation and the representativeness of the Dutch sample. My concern overlaps with the documentation-reading part but is sharper: the quoted Creator Academy text, as reproduced in the paper, already says 'AI' and 'a variety of factors,' so the paper's stronger claim that TikTok 'simply aggregates' and presents itself as a 'neutral intermediary' is not directly implied by that source. This matters because the abstract uses the 'simply aggregates' framing to argue that TikTok sidesteps responsibility. The concern is real but not decisive. The paper is explicitly a position paper; its central contribution is a research agenda and a set of governance challenges, both of which are supported by the interface examples, media reporting, and the acknowledged opacity of the feature. The 'Nikki as a boy' example and the reporting on coordinated commenting show that the feature can surface contextually harmful or manipulated queries even if TikTok's documentation is read conservatively. The small sample is presented as preliminary and illustrative, and the computational challenges section does not claim to have solved data-access problems. Therefore, the correct response is to flag the over-reading as a point to verify and possibly soften, but not to reject or conditionally accept the paper. The concrete test above settles whether TikTok's own words support the neutrality framing. If the framing fails, the paper's core claims about transparency, control, and research access remain intact, so the verdict is unchanged.","tokens_in":8313,"tokens_out":5911,"duration_ms":65796,"concrete_test":"Retrieve archived and current versions of TikTok Creator Academy 'search' and support 'How TikTok recommends content' pages (e.g., via Internet Archive, 2023-2025, English and Dutch) and extract all sentences describing how search suggestions/recommendations are generated. Code each sentence for explicit mentions of AI, 'variety of factors', comments, user searches, and exclusivity markers ('only', 'solely', 'simply'). If the archived pages say 'AI' and 'variety of factors' without exclusivity markers, the paper's 'simple aggregation/neutral aggregator' framing is not directly supported and should be revised to 'TikTok discloses only two user-derived factors while remaining silent on the rest.' If exclusivity markers appear, the original framing is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2 grounds the charge that TikTok 'simply aggregates comments and common searches' in the Creator Academy statement that recommendations are generated using AI based on a 'variety of factors,' two of which are user comments and searches. That quotation does not say 'simply' or 'solely'; it explicitly names AI and a broader factor set. The paper's inference that TikTok 'positions itself as a neutral intermediary' is therefore an interpretive leap rather than a direct quotation, and the abstract's causal claim ('By suggesting ... it sidesteps responsibility') inherits that weakness. The same texts still support the stronger, better-evidenced governance point: TikTok discloses almost nothing about the 'variety of factors,' gives users no way to disable the feature, and offers no policy-level detail on moderation. The Dutch-only, three-day sample is explicitly preliminary and used for illustration; since the paper is a position paper, the unrepresentativeness of that sample is less load-bearing than the documentation reading. If the documentation reading is wrong, the paper can drop or soften the neutrality framing and retain its central claim about transparency, control, and research access.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This position paper analyzes TikTok's search recommendation feature, which presents preformulated search queries on or under videos and in comments sections. The authors argue that this feature raises governance challenges—limited transparency about generation and moderation, weak user control, and potential contextual harms—and that these challenges are not adequately addressed by TikTok's current documentation, interface design, or data access provisions. The paper grounds this argument in an interpretive reading of TikTok's Creator Academy and support pages, journalistic reports, and a small preliminary dataset collected from Dutch influencers' videos. It then discusses the European Digital Services Act as a relevant regulatory framework and proposes four computational research challenges: modeling how recommendations are generated from comments, detecting coordinated behavior, identifying context-dependent harms, and classifying recommendation topics. The paper is explicitly agenda-setting rather than empirical.","tokens_in":8505,"tokens_out":7483,"duration_ms":80626,"significance":"The paper addresses a timely and understudied feature that sits at the intersection of recommender-system governance, search engine studies, and platform accountability. Its strength is in identifying a concrete, evolving product and connecting it to both user-facing harms (misgendering, speculation, coordinated amplification) and regulatory obligations under the DSA. The authors are to be credited for grounding their argument in public documentation and news reports, for clearly separating the four research challenges, and for appropriately hedging most of their claims as open questions. If revised, the paper could serve as a useful intellectual agenda for platform governance and computational social science research. However, the current version's significance is weakened by an overreading of TikTok's documentation and by the thinness of the empirical component; the central governance argument survives after those issues are corrected.","major_comments":[{"comment":"The claim that TikTok “simply aggregates comments and common searches” and that comments and searches are “the only factors TikTok attributes to shaping recommendations” is not supported by the quoted Creator Academy text, which states that recommendations are “generated using AI based on a variety of factors,” two of which are user comments and searches. The quote explicitly leaves room for additional factors and algorithmic mediation. Because the “neutral intermediary” and “sidesteps responsibility” framing appears in the abstract, introduction, and conclusion, this is a load-bearing interpretive step. Please revise to distinguish what TikTok literally discloses from the authors' critical argument that the disclosed explanation is vague and downplays TikTok's editorial role; the latter is well supported and does not require the “only factors” reading.","section":"Abstract and Section 2"},{"comment":"The preliminary qualitative study is described too thinly for the role it plays. The authors report selecting the 10 most-followed Dutch influencers, collecting recommendations over three consecutive days, and obtaining 92 recommendations under videos and 167 in comments, followed by a “qualitative thematic analysis,” but they provide no coding scheme, no sampling procedure, no handling of duplicates, no intercoder reliability, and no discussion of ethical considerations for using public influencer data. The exact count (“18 out of 23 recommended queries”) is presented as empirical evidence. Since the paper is a position paper, this can be fixed by expanding the methods paragraph and clearly labeling the data as illustrative; in the current form, the empirical component carries more weight than its description supports.","section":"Section 2, dataset paragraph"},{"comment":"The legal analysis is asserted rather than developed. The paper moves from “arguably includes these search recommendations” to “TikTok is required to provide transparency on recommendation factors under DSA Article 27(1),” and similarly asserts that Article 25(1) dark-patterns concerns apply, without discussing the DSA's definitions of “recommender system” or “online platform,” or explaining whether search recommendations count as “information” for Article 16 purposes. The hedged phrasing (“arguably,” “raises a compliance question”) is welcome, but the abstract's “despite requirements under regulatory frameworks like the DSA” makes this a central claim. Please add a short doctrinal justification or explicitly frame the DSA discussion as open questions that require legal interpretation.","section":"Section 3, DSA discussion"}],"minor_comments":[{"comment":"The quote contains “FYF,” which appears to be a typo for “FYP” (For You Page); please verify against the original source and correct if needed.","section":"Section 2, Community Guidelines quote"},{"comment":"The figures are referenced in the text but no captions or images are included in the provided version; the final submission should include the screenshots with dates and appropriate permissions or citations.","section":"Figures 1 and 2"},{"comment":"The checklist response says limitations are “NA directly, as this is a position paper,” but the paper presents original empirical observations and should include a dedicated limitations and data-availability statement; this inconsistency should be corrected.","section":"Paper checklist, item 1(e)"},{"comment":"The challenge title “Detecting when coordinated behaviour affects recommendations” is slightly inaccurate for the example, which describes coordinated comments causing a recommendation; consider “Detecting coordinated behaviour that drives recommendations.”","section":"Section 4, Challenge 2"},{"comment":"The phrase “According to an Adobe Express survey” would be more precise as “According to a survey commissioned by Adobe Express” if that is the case, and the citation should indicate the survey's methodology or sample if available.","section":"Section 2, Adobe Express survey"},{"comment":"The sentence “Namely, it is users who shape recommendations through comments and search queries” is ambiguous; please clarify that this is TikTok's claimed framing, not the authors' conclusion.","section":"Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper is timely and likely of interest to the readership, and the proposed research agenda is useful. However, the overreading of TikTok's documentation and the thin empirical section need to be addressed before publication. I would not reject, but the authors should be asked to revise. Please also consider whether the venue expects position papers or original empirical contributions; if the latter, the data section will need to be substantially strengthened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWorth a look. This is the first thing I've seen that treats TikTok's search recommendations as a governance problem rather than just a product quirk. The authors are right that the feature is new, under-documented, and likely to spread—YouTube Shorts already has it, Instagram has signaled intent. They do a solid job laying out the transparency, moderation, and user-control gaps and connecting them to DSA provisions. The examples (\"Nikkie as a boy\" under a trans creator's video, coordinated Super Bowl comments reshaping a search term) are vivid and do real work in showing why context matters.\n\nThe soft spots are real but not fatal. The qualitative study is genuinely preliminary: 10 Dutch influencers, three consecutive days, no coding scheme, no dataset or collection protocol. For a position paper that's fine as illustration, and the authors don't oversell it—the checklist marks most empirical sections N/A. More concerning is the \"neutral aggregator\" charge. The Creator Academy quote says recommendations are generated using AI based on a \"variety of factors,\" two of which are comments and searches. The paper turns that into \"simply aggregates comments and common searches\" and \"positions itself as a neutral intermediary.\" That's an interpretive leap, and the stress-test note has it right: the documentation doesn't say \"simply\" and explicitly names AI. The good news is the governance argument doesn't depend on the leap. The stronger point—TikTok discloses almost nothing about the \"variety of factors,\" users can't disable the feature, creators get no policy-level transparency—stands on its own with the same evidence.\n\nThe DSA analysis is asserted more than argued. Saying Article 27 \"arguably\" covers search recommendations is honest, but the paper doesn't dig into whether a preformulated query string counts as part of a recommender system's inputs, outputs, or neither. That's a section for the authors to strengthen, not a reason to reject.\n\nCitation pattern looks fine; they use journalism and a few academic sources, and the self-citation to their own dataset paper is legitimate since the dataset is the base for the sample.\n\nBottom line: it's a good research agenda paper, modest in its claims, with one overreach you can fix by softening language. I'd send it to peer review. The reader's ACCEPT verdict is about right, with the neutrality framing as the main revision point.\n\nRecommendation: engage with it; the agenda is useful and the feature deserves scrutiny.","headline":"First paper to treat TikTok search recommendations as a governance object; the empirical base is thin and the 'neutral aggregator' framing overreads the documentation, but the research agenda stands.","tokens_in":8992,"tokens_out":1578,"would_cite":true,"duration_ms":14974,"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 steers users toward prewritten searches while denying authorship.","keywords":["search recommendations","TikTok","platform governance","transparency","content moderation","recommender systems","Digital Services Act","social search"],"falsifier":"A large-scale audit could test the paper's core premise: take videos with recommendations, block or delete all comments and suppress post-view searches on a controlled set, and see whether recommendations persist; if they continue unchanged, the claim that TikTok merely aggregates user comments and searches would be false, and if a predictive model built from comments can match recommendations with high accuracy while no contextually harmful queries appear, the paper's transparency and harm claims would be weakened.","tokens_in":2009,"feed_emoji":"🔍","tokens_out":1967,"duration_ms":71319,"temperature":0.7,"pith_summary":"TikTok now prompts users to search by overlaying preformulated query terms on videos and in comment sections. This position paper argues that the feature is a genuinely new form of recommendation: the platform decides what is searchable, inverts the open-ended nature of search, yet claims to be nothing more than a neutral aggregator of user comments and common searches. The paper maintains that this 'neutral aggregator' framing is a strategic opacity that shields TikTok from responsibility for contextually harmful suggestions, such as a transphobic query on a trans creator's videos, while leaving users and creators with no way to disable, contest, or fully report the recommendations. Because the same feature is now spreading to YouTube Shorts and Instagram, the paper argues that its transparency, moderation, and data-access gaps are a pressing governance problem rather than a niche product quirk. It contributes a preliminary qualitative analysis of Dutch influencers and four computational research challenges for studying the feature at scale.","feed_headline":"TikTok steers users toward prewritten searches while denying authorship","feed_subtitle":"These AI-generated prompts can spread rumors and harm creators, with few ways for users to appeal.","key_machinery":"The central object is the search recommendation itself: a preformulated query phrase TikTok places on or under a video and at the top of the comments section, which a user can tap to launch a search. The device that carries the argument is what the paper treats as TikTok's 'selective opacity'—a public-facing explanation that recommendations are generated by AI from only two user-driven factors (comments on the video and searches made after watching), offered through the Creator Academy, combined with the near-total absence of technical or policy detail in user-facing documents. The paper contrasts this thin account with the interface's asymmetrical affordances: some recommendations can be reported, others cannot; reporting targets results rather than the query; and creators' only lever is comment filtering, which is acknowledged by TikTok spokespeople but missing from official documentation. This gap between the claimed neutral-aggregator mechanism and the observable, context-dependent behavior of recommendations is what the paper uses to motivate both its governance analysis and its research agenda.","core_discovery":"The paper's central claim is that TikTok search recommendations are a new recommendation product that preformulates user queries, and that TikTok uses its own documentation to present itself as a neutral conduit: suggestions are said to be generated purely from user comments and what other users search after watching a video. The authors argue this framing is misleading and strategically opaque because the actual indexing of video descriptions, comments, or transcripts is unknown, the feature cannot be modified or disabled by users, creators are not notified of the recommendations attached to their content, and reporting options vary by interface location. Using a qualitative sample of search recommendations on videos of the ten most-followed Dutch influencers, the paper shows that recommendations can be contextually problematic regardless of surface innocuousness, citing 'Nikkie as a boy' on a trans creator's videos and relationship-speculation prompts on videos that never mention those relationships. The paper concludes that the feature poses unresolved governance challenges under the EU Digital Services Act—transparency of recommendation factors, dark-pattern concerns, and limited reporting of objectionable content—and lays out a four-part computational research agenda to study generation, coordinated manipulation, contextual harm, and topic classification.","pith_inferences":["A direct causal test of the 'neutral aggregator' story is possible: if comments and post-view searches were fully blocked or filtered on a test video and recommendations persisted, the Creator Academy explanation would be falsified; conversely, if recommendations vanish, the transparency objection narrows to moderation and reporting gaps.","Because coordinated comments can surface a query across unrelated videos, search recommendations may be a cheaper amplification channel than ads or follower manipulation, one that bypasses ad-disclosure rules and leaves no advertiser paper trail.","The contextual-harm examples imply that automated detection of harmful recommendations will need creator-identity-aware and discourse-aware classifiers rather than keyword filters, since 'Nikkie as a boy' is only harmful given who Nikkie is.","Cross-platform comparison with YouTube Shorts and Instagram is the most direct way to test whether the transparency gap is intrinsic to preformulated queries or a specific TikTok choice."],"forward_implications":["If TikTok's own documentation is the only public account, the feature likely falls short of the Digital Services Act's recommender-transparency requirements under Article 27, the anti-manipulation standard under Article 25, and the notice-and-action obligations under Article 16.","Users and creators would have no reliable way to contest a harmful or wrong recommendation, since the feature cannot be disabled, creators are not notified of attached queries, and reporting routes differ by placement.","Coordinated commenting can steer recommendations onto unrelated videos, as the 'Bisan and Motaz NFL' case shows, making the feature a plausible vector for organized political or commercial influence.","The four research challenges—modelling recommendation generation from comments, detecting coordinated behavior, classifying contextual harm, and categorizing topics—define a concrete computational agenda that can proceed with scraped public data even though the Research API excludes recommendations.","Adoption of similar preformulated-query features by YouTube Shorts and Instagram would carry the same transparency and governance problems onto additional platforms."],"supporting_citations":[{"why":"Creator Academy documentation is the only place TikTok details the two generation factors, comments and post-view searches, and therefore anchors the paper's reading of the platform's neutral-aggregator framing.","marker":"TikTok 2025c"},{"why":"Community Guidelines acknowledge search suggestions in vague terms, grounding the paper's claim that user-facing documentation lacks technical or policy-relevant detail on search recommendations.","marker":"TikTok 2024b"},{"why":"This prior curated dataset supplies the list of the ten most-followed Dutch influencers whose videos formed the paper's qualitative sample.","marker":"Gui et al. 2024"},{"why":"News reporting on TikTok search suggestions manufacturing influencer drama provides one of the paper's core governance-harm examples and contextualizes the moderation gaps.","marker":"Lorenz 2024"},{"why":"The documented coordinated commenting that produced the 'Bisan and Motaz NFL' recommendation is the paper's key evidence that search recommendations can be manipulated.","marker":"DiBenedetto 2024"},{"why":"Supplies the framing of search's open-ended nature and the parallel with Google's vague autocomplete discourse, which the paper uses to characterize TikTok's selective opacity.","marker":"Graham 2023"},{"why":"Grounds the critique that search systems can reinforce stereotyping and discrimination, which the paper aligns with contextually harmful recommendations like 'Nikkie as a boy'.","marker":"Noble 2018"},{"why":"Provides the computational content-moderation lens of contextual harm that the paper extends to the unique case of search recommendations.","marker":"Vidgen and Derczynski 2020"},{"why":"The Digital Services Act is the regulatory benchmark the paper uses to argue that TikTok's documentation, interface design, and reporting mechanisms raise compliance questions.","marker":"Regulation 2022/2065, 'DSA'"}],"fun_headline_variants":["TikTok hides authors of prewritten search prompts","TikTok's search suggestions: more than just comments","Who authors TikTok's search suggestions? Not just users","TikTok claims neutrality but curates search prompts","TikTok's search recommendations: a transparency problem"],"cache_read_input_tokens":11136,"weakest_assumption_plain":"The strongest load-bearing premise is that TikTok's official documentation—particularly the Creator Academy's statement that recommendations come only from comments and post-video searches—can be read as a truthful description of how the feature actually works; the empirical examples also rest on a small, Dutch-only, three-day sample of the ten most-followed Dutch influencers.","fun_headline_variants_meta":{"raw":{"variants":["TikTok hides authors of prewritten search prompts","TikTok's search suggestions: more than just comments","Who authors TikTok's search suggestions? Not just users","TikTok claims neutrality but curates search prompts","TikTok's search recommendations: a transparency problem"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000701,"raw_usage":{"total_tokens":3152,"prompt_tokens":917,"completion_tokens":2235,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":2158}},"tokens_in":533,"tokens_out":2235,"duration_ms":14768,"temperature":1.0,"reasoning_tokens":2158,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:55:28.260122+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A large-scale audit could test the paper's core premise: take videos with recommendations, block or delete all comments and suppress post-view searches on a controlled set, and see whether recommendations persist; if they continue unchanged, the claim that TikTok merely aggregates user comments and searches would be false, and if a predictive model built from comments can match recommendations with high accuracy while no contextually harmful queries appear, the paper's transparency and harm claims would be weakened.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This prior curated dataset supplies the list of the ten most-followed Dutch influencers whose videos formed the paper's qualitative sample."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"News reporting on TikTok search suggestions manufacturing influencer drama provides one of the paper's core governance-harm examples and contextualizes the moderation gaps."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The documented coordinated commenting that produced the 'Bisan and Motaz NFL' recommendation is the paper's key evidence that search recommendations can be manipulated."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the framing of search's open-ended nature and the parallel with Google's vague autocomplete discourse, which the paper uses to characterize TikTok's selective opacity."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Grounds the critique that search systems can reinforce stereotyping and discrimination, which the paper aligns with contextually harmful recommendations like 'Nikkie as a boy'."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the computational content-moderation lens of contextual harm that the paper extends to the unique case of search recommendations."}],"review_version":1}