{"id":"238dc7a5-4e92-4575-a7db-13433152505c","arxiv_id":"2501.11457","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Creative workers see consent, credit, and compensation as necessary but imperfect tools for governing generative AI, and report a wide gap between the governance they want and what employers, platforms, and governments currently provide.","lead":"Twenty creative workers in design, writing, and programming told researchers what they want from generative AI governance: consent and compensation matter, but credit is sometimes unwelcome and one-time consent can go stale. The findings give regulators and companies a worker-side evidence base for debates about training data and AI regulation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central generalization from 20 young, mostly Global North, network-recruited participants to creative workers generally is the weakest load-bearing assumption; the pilot participants' inclusion compounds the risk that the reported 'nuances' are partly an artifact of question design.","rationale":"The reader's weakest assumption—representativeness of the 20-interview convenience sample—is indeed the most load-bearing point. The paper's headline contribution is a general empirical claim about creative workers' views on consent, credit, and compensation, and a general claim about a governance gap. Both are stated at the level of 'creative workers' in the Introduction and Discussion, but the evidence comes from a small, young, mostly Global North sample with a specific recruitment path. The Section 5.6 limitations are honest, but they do not fully resolve the mismatch between the sample and the scope of the claims. I agree with the reader's conditional verdict rather than moving to reject: qualitative interview studies can support transferable insights, the quotes are illustrative, and the authors do not claim statistical generalization. The additional pilot-inclusion issue is real but probably mild; it is worth testing rather than sufficient to overturn the paper. The recommended two-part test—an internal sensitivity analysis excluding pilot participants and a larger stratified replication—would directly determine whether the concern lands or whether the paper's conclusions are robust. Until such a test is run, the conditional verdict is the right calibration.","tokens_in":26462,"tokens_out":4048,"duration_ms":47463,"concrete_test":"Perform a two-part robustness check. First, re-analyze the anonymized transcripts excluding P6, P7, and P11, and determine whether the key themes—'(credit can cause reputational harm)' and '(compensation does not resolve attachment-based objections)'—are each expressed by at least three independent non-pilot participants. If these themes disappear or reduce to one or two voices, the pilot-inclusion circularity materially affects the central claim. Second, run a pre-registered, quota-based replication with a larger sample (N≥100) stratified by creative domain (visual art/design, writing, programming), employment type (employee/freelancer), career stage, and region, using the same consent/credit/compensation vignettes from the interview script.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that the 3 Cs are 'not as straightforward' as current discourse suggests and that a governance gap exists—is supported almost entirely by 20 semi-structured interviews recruited via snowball sampling from the authors' networks plus Fiverr (Section 3.3.1). The participant pool averages 28.6 years old and 5.23 years of experience, skews Global North, and contains all programmers as company employees while writers/artists are a mix of freelancers and employees (Table 1; Section 5.3). The authors partially concede this in Section 5.6, but the paper's Discussion and Section 5.5 policy recommendations move freely from 'participants in our study' to 'creative workers' generally. For example, Section 4.2.1 reports that 'most participants said the company, publisher, or freelance platform they work with currently has no AI governance strategy'; this becomes the basis for the claim that a governance gap exists, yet it depends on participants' own knowledge of employer policies and on the representativeness of the sample. A second, compounding issue is that the three pilot participants who helped shape the interview questions (P6, P7, P11) were re-invited into the official sample (Section 3.1). Their pre-interview input may have steered the protocol toward their own concerns, making the 'nuanced' findings—especially credit as reputational harm and compensation not resolving attachment-based objections—partly a product of the question design rather than of the broader creative workforce. If the sample is not representative, or if pilot participants disproportionately drive the key themes, the paper's headline findings and the regulatory recommendations in Section 5.5 are not supported at the claimed scope.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper reports findings from 20 semi-structured interviews with creative professionals in visual art/design, writing, and programming, conducted to understand their perspectives on generative AI governance. The authors investigate whether the '3 Cs' framework (consent, credit, compensation) adequately addresses creative workers' concerns, document a perceived gap between existing governance structures and what workers want, and offer policy and practice recommendations. The central claims are that the 3 Cs are more nuanced than current discourse suggests—e.g., consent is constrained by employment power asymmetries, credit can be unwanted due to reputational risk, and compensation does not resolve attachment-based objections—and that current company, platform, and governmental governance is insufficient. The paper includes a description of the interview protocol, a codebook referenced in supplementary materials, and a candid limitations section.","tokens_in":26592,"tokens_out":3128,"duration_ms":37666,"significance":"If the findings hold, the paper makes a useful empirical contribution to HCI and AI governance by grounding the popular '3 Cs' discourse in worker perspectives and by articulating concrete governance gaps. The study is well-designed for its qualitative scope: the semi-structured protocol is included, the two-author iterative thematic analysis is documented with 121 codes and 13 themes, and the authors are transparent about sample limitations. The nuanced findings—particularly that credit can harm workers and that consent must be ongoing and collective—are potentially actionable for companies and regulators. However, the strength of these contributions is bounded by the sample's narrowness (20 participants, mostly young, Global North, network-recruited) and by one specific methodological choice involving pilot participants, which I detail below.","major_comments":[{"comment":"The pilot-participant contamination issue is load-bearing because it directly affects the validity of the study's main qualitative findings.","section":"§3.1"},{"comment":"The overgeneralization concern is central because the paper's second and third contributions are framed as general claims about creative workers and recommended governance structures.","section":"§5.6 and §5.5"},{"comment":"The reliance on participant awareness is a legitimate limitation that should be disclosed in the Results or Discussion, not only in the general limitations section.","section":"§4.2.1"}],"minor_comments":[{"comment":"There is a typo in the title and abstract: 'Cred it' should be 'Credit'.","section":"Title/Abstract"},{"comment":"Reference [40] for the IEEE AIS initiative appears to have an incorrect URL; it points to a DeepMind page rather than the intended IEEE page. Please verify and correct the reference.","section":"§2.2, Reference [40]"},{"comment":"Participant P4 is listed as 'Trading associate' with 'Under 10' employees—it is unclear how this role fits the creative domains of art/design, writing, or programming described in the paper. Please clarify the participant's creative role or exclude if it is not within the stated inclusion criteria.","section":"Table 1"},{"comment":"In the interview script, 'I will be conducing an interview' should be 'I will be conducting an interview.'","section":"Appendix A"},{"comment":"The sentence 'cf. e.g. Art. 7(3) GDPR, Sentence 2' is awkward because Article 7(3) GDPR does not contain a named 'Sentence 2.' Consider rephrasing to 'cf. the second sentence of Art. 7(3) GDPR.'","section":"§5.5"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid empirical core and the topic is timely for CHI. The main barrier to acceptance in my view is the pilot-participant contamination issue (§3.1), which is fixable by re-analysis excluding those three participants or by a convincing methodological argument. The generalizability issue is also concerning but is partially addressed in the limitations; still, the abstract and recommendations overstate the scope. These are not fatal flaws, but they require careful revision before the paper should appear. I would be comfortable with acceptance after the authors address the pilot-participant issue and temper the general claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nYou should know this is a careful, well-scoped qualitative study, and its main result is more interesting than the title suggests. The paper gives creative workers' own words to show that the '3 Cs'—consent, credit, compensation—are not the simple fixes they often appear to be in advocacy. Some workers don't want credit because they don't want to be associated with model output; consent decisions shift as people anticipate future capabilities; work output and personal projects elicit different feelings. Those specifics are new and worth having.\n\nThe method is solid for what it is: 20 semi-structured interviews, protocol in the appendix, codebook referenced, two-author thematic analysis, and a limitations section that names self-selection and snowball recruitment. The quotes are well chosen and the findings are mostly reported as participant views, not as population estimates.\n\nThe soft spots are real but not disqualifying. The biggest is the gap between the sample and the generalizing language. The 20 participants average 28.6 years, skew Global North and network-recruited, and the paper's recommendations in Section 5.5 are written for creative workers at large. When the paper says 'most participants said their employer has no AI governance strategy,' that is a statement about 20 people's knowledge, not a verified fact about the industry. The authors concede some of this in 5.6 but still move freely to 'creative workers' in the Discussion. That's a framing fix, not a fatal flaw.\n\nA smaller but real issue: the three pilot participants who helped shape the interview questions were re-invited into the official sample (P6, P7, P11) without the paper noting this as a potential bias. Their input likely steered the protocol toward their own concerns, which could make the 'nuanced' findings partly an artifact of question design. I'd want that acknowledged, though the effect is probably minor given the small number.\n\nI don't share the stress-test's concern that this is load-bearing. The central contribution—that the 3 Cs are more complicated than the slogan suggests—would survive even with a different sample, because it is about documenting previously unreported reasoning, not about estimating population frequencies. The paper would be stronger if it consistently said 'our participants' instead of 'creative workers,' but the underlying finding stands.\n\nThis paper is for people working on AI governance, HCI, and labor policy. It gives them a useful empirical baseline and a set of testable hypotheses. It deserves a serious referee. I'd engage with it, and I'd recommend it go to peer review with a request for tighter framing around generalizability.\n\nBest.","headline":"A careful interview study that adds real nuance to the 3 Cs debate; its policy reach exceeds its 20-participant sample, but that is fixable framing, not a fatal flaw.","tokens_in":27329,"tokens_out":3565,"would_cite":true,"duration_ms":36346,"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":"Consent, credit, and compensation are not enough to protect creative workers from generative AI, a 20-interview study argues.","keywords":["AI governance","generative AI","creative workers","consent","credit","compensation","qualitative interviews","labor rights"],"falsifier":"A representative, preregistered survey of several hundred creative workers across visual art, writing, and programming that found a large majority would accept one-time consent, mandatory attribution, and standardized compensation as adequate would directly contradict the paper's core claim. So would observational evidence that companies already offering ongoing re-consent and credit opt-outs see no improvement in worker trust or retention.","tokens_in":26114,"feed_emoji":"🎨","tokens_out":4599,"duration_ms":47203,"temperature":0.7,"pith_summary":"The paper argues that the popular \"3 Cs\" remedy for harms to creative workers from generative AI—obtaining consent before training on their work, giving credit, and paying compensation—is far less straightforward than advocates assume. Drawing on 20 interviews with visual artists, designers, writers, and programmers, it finds that consent is often meaningless under employer power imbalances, that credit can expose workers to reputational harm for outputs they do not control, and that compensation does not resolve workers' objections to training on personal, attachment-laden projects. It also documents a gap between the governance workers say they want and what companies, publishers, platforms, and governments currently provide. The authors conclude that governance must be ongoing rather than one-time, collective rather than purely individual, and extended to training-data pipelines, not just model outputs.","feed_headline":"Consent, credit, pay not enough to shield creatives from AI","feed_subtitle":"Interviews with 20 artists, writers, and programmers show why the '3 Cs' need ongoing, collective, and pipeline-level governance.","key_machinery":"The object that carries the argument is the \"3 Cs\" framework (consent, credit, and compensation), a term coined in 2017 for protecting Indigenous cultural property and since adopted as a rallying cry in generative AI disputes. The paper treats it not as a fixed standard but as a hypothesis to be tested against workers' accounts, and the test produces a set of specification conditions: consent must be informed, genuinely voluntary, and renegotiated over time; credit must account for anonymity preferences and liability concerns; compensation must be contextualized by employment status and personal attachment. The interview instrument, with its hypothetical scenarios about training models on work before and after leaving an employer, is the mechanism that surfaces these conditions.","core_discovery":"The central finding is that the 3 Cs framework—consent, credit, and compensation—does not function as a simple protective triad in creative work. Almost all interviewees wanted to be asked for consent, but employees often felt they could not refuse because their company owns their work and holds power over their livelihoods; freelancers similarly feared economic retaliation. Some workers explicitly did not want credit, either because they saw their employer as the rightful author, because they were not well-known enough for attribution to matter, or because being named would tie them to AI-generated outputs they might find offensive or harmful. All wanted compensation, but many distinguished between work-for-hire output, which they were willing to see trained on, and personal projects, where money did not repair the sense of violation. The paper reads these patterns as evidence that consent must be re-asked as technology develops and working conditions change, credit must be optional and informed by reputational risk, and compensation must be paired with ongoing consent and ownership structures.","pith_inferences":["A natural extension the authors leave implicit is that individual notice-and-choice consent mechanisms, as under data-protection law, are structurally ill-suited to this setting; collective consent through works councils or unions would be a more faithful implementation of their findings.","The age and experience profile of the sample (average 28.6 years, 5.23 years of experience) suggests the paper may understate the strength of objections among older, more established workers whose reputational capital and career investment are larger.","A testable corollary is that workers who are offered an anonymity-preserving credit option plus ongoing re-consent will report higher trust in an AI training arrangement than workers offered one-time credit and compensation; this could be examined in a vignette experiment.","The attachment distinction could be operationalized into policy by requiring separate consent tracks for commissioned versus personally initiated work, though the authors do not develop that mechanism."],"forward_implications":["If the paper is right, any governance regime built on a single up-front consent checkbox will fail a substantial share of creative workers, because consent loses meaning under employment power asymmetries and because technology changes over time.","Mandatory credit requirements would backfire for workers who want to dissociate from model outputs; credit should be an option with default anonymity protections built in.","Compensation schemes that ignore the attachment distinction between work and personal projects will leave the strongest objections unresolved.","Regulators should extend oversight from AI outputs to the data pipelines and employment and contractual settings in which training data is collected, not just to the final system.","Companies, publishers, and platforms currently lack AI governance strategies; the paper's recommendations imply a proactive, worker-consultative standard rather than reactive banning."],"supporting_citations":[{"why":"Introduces the 3 Cs (consent, credit, compensation) as a framework the paper tests against workers' experiences.","marker":"[43]"},{"why":"Popularizes the 3 Cs as a demand in the generative AI context, the discourse the paper complicates.","marker":"[54]"},{"why":"Shows workers cannot meaningfully consent under workplace power dynamics, which the paper uses to interpret interviewees' consent limitations.","marker":"[13]"},{"why":"Provides experimental evidence that workers value attribution even over payment, which the paper complicates with reputational concerns.","marker":"[77]"},{"why":"Documents the law and norms of attribution, used to frame credit as an enforceable right and to identify its limits.","marker":"[29]"},{"why":"Documents career harms of generative AI on artists, the background problem the paper addresses.","marker":"[44]"},{"why":"Shows users often ignore security notifications, cited to argue consent updates must be delivered through accessible, repeated channels.","marker":"[92]"},{"why":"Argues legal frameworks for ownership and originality in AI must be reframed, which the paper endorses for moral rights in the AI pipeline.","marker":"[61]"}],"fun_headline_variants":["Consent, credit, pay aren't enough: power skews AI training deals","Creatives want more than credit: they want ongoing control over AI use","For creators, AI consent is a trap when bosses own the work","AI training needs collective, not just individual, creator consent"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's conclusions rest on 20 interviews recruited through the researchers' own networks and an online freelancing platform, so the findings stand or fall with whether that small, self-selected, mostly Global North group represents creative workers in general.","fun_headline_variants_meta":{"raw":{"variants":["Consent, credit, pay aren't enough: power skews AI training deals","Creatives want more than credit: they want ongoing control over AI use","For creators, AI consent is a trap when bosses own the work","AI training needs collective, not just individual, creator consent"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000496,"raw_usage":{"total_tokens":2402,"prompt_tokens":883,"completion_tokens":1519,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":1442}},"tokens_in":499,"tokens_out":1519,"duration_ms":12428,"temperature":1.0,"reasoning_tokens":1442,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:15:16.881287+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A representative, preregistered survey of several hundred creative workers across visual art, writing, and programming that found a large majority would accept one-time consent, mandatory attribution, and standardized compensation as adequate would directly contradict the paper's core claim. So would observational evidence that companies already offering ongoing re-consent and credit opt-outs see no improvement in worker trust or retention.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the 3 Cs (consent, credit, compensation) as a framework the paper tests against workers' experiences."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Popularizes the 3 Cs as a demand in the generative AI context, the discourse the paper complicates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides experimental evidence that workers value attribution even over payment, which the paper complicates with reputational concerns."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the law and norms of attribution, used to frame credit as an enforceable right and to identify its limits."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents career harms of generative AI on artists, the background problem the paper addresses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Argues legal frameworks for ownership and originality in AI must be reframed, which the paper endorses for moral rights in the AI pipeline."}],"review_version":1}