{"id":"0b1cee33-8a56-41f2-ae4d-0a4932c6f10b","arxiv_id":"2505.00195","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"When two user collectives manipulate the same ML system simultaneously, their actions interact and can sharply reduce each other's effectiveness, with collective size mattering more than membership homogeneity in recommender tests.","lead":"This paper introduces a framework for studying what happens when two organized groups of users try to manipulate the same machine learning system at the same time. In simulations, a group that succeeds alone often loses much of its influence when a second group acts, sometimes dropping from near 100% success to near 25%.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The advertised 75% drop in RQ1 comes from a configuration where the two supposedly distinct strategies collapse to the same tokenizer token, so the headline interaction may be same-strategy conflict rather than distinct-collective interference.","rationale":"The paper's central contribution is a framework for multiple-collective algorithmic collective action, and the framework itself is plausible and useful. The recommender experiments include robustness checks across clustering choices and metrics, which is genuine independent support. The reader's weakest assumption concerned whether matrix-factorization-based clusters model real collective formation; that is a legitimate external-validity concern, but it is not the most load-bearing issue for the paper's headline claim. The headline quantitative effect in RQ1 depends on the A100/B101 condition, and Section 5.1 itself reports that the tokenizer treats those two characters as identical. Consequently, the 'as much as 75%' drop is not evidence for an interaction between two distinct strategies; it is evidence for a same-signal conflict. This does not destroy the paper: same-strategy conflicts are real, and some distinct-strategy effects appear in Figure 3a. But it does mean the central claim, as quantified in the abstract and conclusion, needs to be either re-run with genuinely distinct tokens or re-scoped to describe same-token conflicts. Since the paper's overall verdict was already CONDITIONAL, this concern reinforces the need for a condition rather than changing the verdict category.","tokens_in":23494,"tokens_out":2847,"duration_ms":30229,"concrete_test":"Run the distilbert-based-uncased tokenizer on the characters mapped to '100' (UTF-2E17) and '101' (UTF-2E18) in Appendix B and compare their token IDs to confirm they are identical. Then rerun the Figure 3d / Figure 4b experimental condition with two never-before-seen characters that tokenize to distinct IDs, matched for frequency and insertion pattern, and measure A100's efficacy at 2% participation. If the drop to roughly 25% persists with genuinely distinct tokens, the interaction claim stands; if efficacy remains near 100%, the advertised 75% drop is an artifact of tokenizer aliasing and the headline should be re-scoped.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim in the abstract and conclusion is that a collective achieving near 100% efficacy alone can drop to nearly 25% when a second collective acts (Abstract; Section 7). The plot behind this number is Figure 3d / Figure 4b, the A100 vs B101 condition, described as different targets and different characters. However, Section 5.1 explicitly states that, for the distilbert-based-uncased tokenizer, the characters mapped to '100' and '101' are treated as identical. This means the model receives the same planted signal from both collectives, so the interaction in this configuration is effectively the same-strategy/different-target conflict already shown in Figures 3b and 3e, not an 'unintentional interaction' between two genuinely distinct strategies. The conclusion that interactions between distinct collectives can produce a 75% drop therefore rests on a tokenizer aliasing artifact. The paper does demonstrate interference in same-strategy conflicts, and Figures 3a/3c show some distinct-strategy effects, but the specific quantitative headline -- the 'as much as 75%' figure -- is not supported as stated. Because the authors themselves identify the aliasing, this is an internal weakness in the interpretation of the headline result rather than an external-validity concern.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a framework for studying algorithmic collective action by two or more collectives that each manipulate a shared data-dependent machine learning system. The framework distinguishes collectives by their objectives, construction, available actions, model access, affected parties, and measurement. The authors instantiate the framework in two empirical settings: (1) fine-tuned DistilBERT classifiers on resume data, where two collectives plant different text signals and attempt to steer classification to their own target classes, and (2) matrix-factorization recommender systems on MovieLens 100k, where collectives promote or demote items and the authors vary collective size, homogeneity (via a sampling propensity over user clusters), and the pairing of promoter/demoter archetypes. The main reported finding is that interactions between simultaneously acting collectives can be large, with a claim in the Abstract and Conclusion that a collective's efficacy can drop from near 100% to roughly 25% when a second collective acts. A secondary finding is that in recommender systems, collective size has a first-order effect on efficacy while homogeneity has a second-order effect, and that neither fully homogeneous nor fully heterogeneous collectives are uniformly most effective.","tokens_in":23858,"tokens_out":5222,"duration_ms":50212,"significance":"The paper addresses a genuinely novel question: prior work on algorithmic collective action has largely considered a single collective, whereas real platforms are likely to host multiple, independently motivated campaigns. The proposed framework is a useful conceptual contribution, and the recommender experiments include robustness checks across four combinations of clustering algorithm (k-means vs. k-medoids) and distance metric (L2 vs. cosine), as well as across centroid selection strategies. The constructiveness score is a sensible operationalization of between-collective influence. The finding that interactions between collectives can be substantial, if it survives the concerns below, would be a valuable and falsifiable contribution for both researchers and platform designers. The paper does not ship code or machine-checked proofs, so its contribution rests on the clarity of the framework and the soundness of the empirical analysis.","major_comments":[{"comment":"The headline claim of a drop from near 100% efficacy to roughly 25% (a 75% drop) is based on the A100 vs B101 condition in Figures 3d and 4b, but Section 5.1 explicitly states that for the distilbert-based-uncased tokenizer, the characters mapped to '100' and '101' are treated as identical. This means the model receives the same planted signal from both collectives, so the observed interaction is effectively a same-strategy conflict rather than an unintentional interaction between two genuinely distinct strategies. The Abstract and Conclusion do not carry the qualification that appears in the results section, and the caption of Figure 4b even calls the strategies distinct. The authors should either rerun the condition with tokens that remain distinct after tokenization, or revise the Abstract and Conclusion to attribute the headline drop to a same-strategy conflict and temper the 'distinct strategies' framing.","section":"Abstract; Section 7; Section 5.1"},{"comment":"The RQ1 language-model results are reported as point estimates only. Figure 3 shows no error bars, and Figure 4's caption says the heatmaps are averaged across 5 trials, while Appendix B states that each experimental condition was run 10 times; no variance or confidence interval is reported anywhere. Because the paper makes precise quantitative claims (e.g., a drop to nearly 25% efficacy, or that 'B101 requires nearly 2x larger participation than A100'), the absence of uncertainty reporting makes it impossible to assess whether the observed differences are meaningful relative to run-to-run noise. Please report per-condition means with standard deviations or confidence intervals, and reconcile the trial-count discrepancy.","section":"Section 5.1; Figures 3 and 4; Appendix B"},{"comment":"The homogeneity manipulation is confounded with target-item selection. After sampling collective members with propensity p, the targeted items are chosen as the collectively highest-rated items of those members, so varying p changes not only member homogeneity but also the coherence, popularity, and overlap of the item sets being promoted or demoted. This confound undercuts the conclusion that homogeneity has only a secondary influence on efficacy: the observed differences could be driven by the properties of the items selected under different p values. Please hold the target item set fixed across homogeneity levels, or provide an analysis that separates member homogeneity from item-set properties.","section":"Section 4.2 (Collective Formation and Item Selection) and Section 5.2 (Figures 5 and 6)"}],"minor_comments":[{"comment":"The sentence 'However, it is the opposite' in the discussion of the green lines is confusing; the preceding sentence says the promoter wants a positive score and the demoter wants a negative score, and the next sentence clarifies that both are hindered. Please rewrite to state directly that both collectives receive the opposite sign from what they desire.","section":"Section 5.2 (Figure 6 discussion)"},{"comment":"The caption states that 'the strategy and the target used by each collective is distinct,' which is inconsistent with the paper's own observation in Section 5.1 that the tokenizer treats the '100' and '101' characters as identical. Please correct the caption to reflect the tokenizer aliasing.","section":"Figure 4b caption"},{"comment":"Appendix B says each experimental condition was run 10 times, while the Figure 4 caption says the results are averaged across 5 trials. Please reconcile this inconsistency.","section":"Appendix B vs. Figure 4 caption"},{"comment":"The caption says 'solid lines represent demoting groups while solid lines represent promoting groups'; the second phrase should presumably be 'dashed lines represent promoting groups.'","section":"Figure 11 caption"},{"comment":"References [12] and [13] appear to cite the same paper (Etter and Albu 2021) with the same title, authors, and publication; please merge or differentiate them.","section":"References [12] and [13]"}],"recommendation":"major_revision","confidential_remarks":"The paper proposes a useful framework and an interesting empirical setup, but the headline quantitative claim is currently not supported because of the tokenizer aliasing in the A100 vs B101 condition. The recommender-system homogeneity conclusion is also weakened by a confound between sampling propensity and target-item selection. Both issues are fixable in a revision: re-running the language-model condition with tokens that survive tokenization, and holding target items fixed in the recommender experiments. I would not reject the paper, but the abstract and conclusion need to be brought in line with what the experiments actually show."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the thing to know: the paper is worth engaging, but its headline number is not what it appears. The advertised 75% drop in LM efficacy comes from the A100 vs B101 condition, and the authors themselves state in Section 5.1 that the distilbert tokenizer treats '100' and '101' as identical. So the two 'distinct' strategies are the same signal to the model. The stress-test note has this right: it's same-strategy conflict, not distinct-collective interference.\n\nWhat is genuinely new: they extend algorithmic collective action from a single collective to two, with a framework that separates objectives, strategies, size, and homogeneity. The recommender experiments are the stronger half: across four clustering/centroid choices, a promoter and a demoter with maximally separated seed clusters still interfere, and the constructiveness score is a clean way to quantify that. The 100-simulation runs with shaded error bands give me reasonable confidence in that result. The same-strategy LM conflicts are also informative.\n\nSoft spots, in proportion. Figures 3 and 4 have no error bars or confidence intervals despite a stated 10 trials per condition; that should be fixed. The selective emphasis on the A100/B101 drop overstates distinct-strategy interference; the paper would be more accurate if it framed that case as a tokenization artifact. The recommender formation model—clustering users by rating similarity and sampling with propensity p—is an assumption about how real collectives form that the authors acknowledge but don't test; I'd call that minor given the paper's scope. No code is released, which limits reproducibility of the exact experiments.\n\nCitation pattern is fine: it builds on Hardt et al. and Baumann and Mendler-Dünner, and distinguishes itself from multi-objective adversarial work. The acknowledgments suggest honest engagement.\n\nWho this is for: people studying algorithmic collective action, data activism, or targeted poisoning. It's a useful stepping stone, not the last word. A serious editor should send it to peer review, with the clear expectation that the headline claim gets re-examined and uncertainty estimates get added. The core finding—multi-collective interactions can be large—survives the tokenizer critique because the recommender experiments establish it independently.","headline":"The recommender half is solid and the framework is a useful extension, but the LM headline 75% drop is undermined by the authors' own admission that '100' and '101' are the same token to the model.","tokens_in":24263,"tokens_out":2795,"would_cite":true,"duration_ms":26914,"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":"A collective that can steer a language model alone can lose up to 75% of its efficacy when a second collective acts simultaneously, even when they target different classes.","keywords":["Algorithmic Collective Action","Social Computing","Data Campaigns","Recommender Systems","Language Models","Collective Efficacy","Constructiveness Score","Multi-collective interaction"],"falsifier":"Run the same two-collective language model experiment with a tokenizer verified to assign distinct token IDs to the two planted characters (e.g., '100' and '101'); if the ~75% efficacy drop in the '100' versus '101' condition does not occur, then that specific interference is an artifact of tokenizer conflation rather than a general property of multi-collective action.","tokens_in":23274,"feed_emoji":"🎯","tokens_out":5597,"duration_ms":54667,"temperature":0.7,"pith_summary":"The paper builds the first framework for studying what happens when two or more organized groups ('collectives') each try to steer the same data-driven algorithm toward their own goal at the same time. It argues that these groups are likely to differ in objectives, strategies, size, and internal similarity, and that such differences produce unintended interactions. Across experiments with a resume-classifying language model and a movie recommender, the paper finds that a group that can achieve near-perfect efficacy alone can lose as much as 75% of that efficacy when a second group acts simultaneously, even when the two groups target different classes or items. In recommender systems, group size matters more than internal homogeneity, and neither fully similar nor fully dissimilar groups are the most effective. The work matters because it suggests that collective-action organizers cannot predict success from their own group alone, and that platform developers need to account for competing data campaigns.","feed_headline":"Two groups gaming an AI can cut each other's power by 75%","feed_subtitle":"First framework for simultaneous data campaigns shows size beats similarity and promoters clash with demoters.","key_machinery":"The central mechanism is the two-collective experimental framework built on (1) a signal-planting strategy for language models—inserting a particular character every 20 words and relabeling those resumes to a target class—and (2) a cluster-based collective construction for recommender systems, where users are clustered by matrix-factorization vectors and members are sampled with propensity $p$ from a seed cluster to control homogeneity. The paper also introduces the constructiveness score $\\mathrm{CT}(c_i, c_j)$, which measures how much collective $j$ helps or hurts collective $i$'s objective compared with $i$ acting alone; this score is the quantitative lens through which interference and synergy are detected.","core_discovery":"The central claim is that simultaneous collective action by two distinct groups produces substantial, often unintended, interactions: a 'targeted promoter' or 'targeted demoter' group that is highly effective in isolation can be sharply hampered by the presence of a second group with its own objective. In the language-model experiments, two collectives each plant a distinct character signal in resumes and relabel their training data; when both act, a collective that reached near-100% top-one accuracy alone can fall to about 25%, depending on whether the model's tokenizer conflates the two signals and on the relative sizes of the groups. In the recommender experiments, the paper defines a 'constructiveness score' that compares a group's hit-ratio gain when acting with a partner versus alone, and finds that two promoting groups help each other, two demoting groups help each other, but a promoter and a demoter actively interfere, even when their seed clusters are maximally far apart. It also reports that collective size has first-order influence on efficacy, while homogeneity—controlled by a sampling propensity from a seed cluster—plays a secondary role, with the most effective groups often not fully homogeneous.","pith_inferences":["If real collectives form through social networks rather than rating-similarity clusters, the homogeneity findings may shift; a natural extension is to build collectives from observed follower or interaction graphs and compare constructiveness scores.","The paper's tokenizer-conflation explanation suggests a testable prediction: using a tokenizer that keeps the two planted characters as distinct tokens should reduce the interference seen in the '100' versus '101' experiment; this could be checked without new data collection.","For three or more collectives, pairwise constructiveness scores may not sum linearly; the paper's suggestion of VCG-style marginal-harm calculations hints that adding a collective could flip the balance in ways pairwise scores miss.","The claim that demoting is easier than promoting implies an asymmetry in content moderation: campaigns seeking to bury content may require less coordination than campaigns seeking to surface content, which platforms could exploit when designing countermeasures."],"forward_implications":["Organizers of a data campaign cannot rely on their group's characteristics alone; the presence of an unrelated campaign can cut efficacy from near 100% to about 25%, so efficacy estimates should account for concurrent actors.","AI and platform developers should expect and monitor for competing data campaigns; the framework gives them a way to compute pairwise constructiveness scores to anticipate which groups interfere.","In recommender systems, demoting items from the top-10 is easier than promoting items into it, so demotivation campaigns can succeed with smaller and less homogeneous groups.","Homogeneity is a secondary factor relative to group size for both promoting and demoting collectives, so recruiting more members matters more than curating a similar membership.","The tokenizer conflation result implies that seemingly different campaign strategies can be treated as the same signal by a model, so 'distinctness' of strategies cannot be assumed from surface-level difference."],"supporting_citations":[{"why":"Supplies the single-collective algorithmic collective action formulation and the resume signal-planting method that this paper extends to two collectives.","marker":"[19]"},{"why":"Provides a prior single-collective recommender campaign study whose setup the recommender experiments build on.","marker":"[4]"},{"why":"Supplies the classical collective-action criteria of size and homogeneity that motivate the recommender research question.","marker":"[39]"},{"why":"Provides DistilBERT, the language model fine-tuned for the resume classification experiments.","marker":"[45]"},{"why":"Supplies the resume dataset used for the language model classification task.","marker":"[28]"},{"why":"Supplies the MovieLens 100k dataset used for the recommender system experiments.","marker":"[21]"},{"why":"Provides the Surprise library used to implement matrix factorization and parameter grid search for the recommender.","marker":"[24]"},{"why":"Supports the expectation that model influence scales with model complexity, which the paper invokes when projecting results to larger models.","marker":"[18]"}],"fun_headline_variants":["Clashing AI collectives slash each other's impact by 75%","When two groups game AI, one can lose 75% of its power","Promoter vs demoter: AI collective clash cuts gains by 75%","Two AI campaigns collide: first group's efficacy plummets 75%","Size beats similarity: dueling AI groups can crush or boost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"In the recommender experiments, the authors assume that matrix factorization user vectors faithfully capture user similarity and that sampling collective members from clusters with propensity $p$ models how real collectives form.","fun_headline_variants_meta":{"raw":{"variants":["Clashing AI collectives slash each other's impact by 75%","When two groups game AI, one can lose 75% of its power","Promoter vs demoter: AI collective clash cuts gains by 75%","Two AI campaigns collide: first group's efficacy plummets 75%","Size beats similarity: dueling AI groups can crush or boost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1450,"prompt_tokens":1064,"completion_tokens":386,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":680,"completion_tokens_details":{"reasoning_tokens":289}},"tokens_in":680,"tokens_out":386,"duration_ms":4082,"temperature":1.0,"reasoning_tokens":289,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:48:02.570442+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same two-collective language model experiment with a tokenizer verified to assign distinct token IDs to the two planted characters (e.g., '100' and '101'); if the ~75% efficacy drop in the '100' versus '101' condition does not occur, then that specific interference is an artifact of tokenizer conflation rather than a general property of multi-collective action.","supporting_citations":[{"cited_title":"Algorithmic Collective Action in Machine Learning","cited_arxiv_id":"2302.04262","evidence_quote":"Supplies the single-collective algorithmic collective action formulation and the resume signal-planting method that this paper extends to two collectives."},{"cited_title":"Algorithmic Collective Action in Recommender Systems: Promoting Songs by Reordering Playlists","cited_arxiv_id":"2404.04269","evidence_quote":"Provides a prior single-collective recommender campaign study whose setup the recommender experiments build on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the classical collective-action criteria of size and homogeneity that motivate the recommender research question."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the resume dataset used for the language model classification task."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the expectation that model influence scales with model complexity, which the paper invokes when projecting results to larger models."}],"review_version":1}