{"id":"a052f8dc-6a92-4641-878b-0505959bccc1","arxiv_id":"2508.00723","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A 16-participant interview study across four domains finds that decision-maker background, perceptions of the AI, personal consequences, and stakeholder implications shape whether professionals adopt AI tools.","lead":"The authors interviewed 16 professionals in journalism, law, medicine, and the public sector about why they do or do not use AI tools at work. They identify four common factors and propose a checklist, the AI adoption sheet, to help organizations predict and support uptake of AI.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The legal-domain cell has only 2 participants, yet the Section 5 cross-domain contrasts—especially the 'tailoring communications absent in LEG' claim and the Figure 4 adoption-sheet signs—rest on it, and no inter-coder reliability check is reported for those signs.","rationale":"The reader identified the small, partly insider sample as the weakest assumption, and I agree that the legal cell is the most fragile. My stress-test sharpens this in two ways: first, the specific Section 5.2 inference about law is based on only two participants and an external survey, not on direct evidence; second, the Figure 4 signs that carry the inter-domain comparison have no reported coding-reliability check, so even the existing 16 interviews cannot be independently audited from the paper itself. I do not see internal inconsistency or overclaiming beyond what the authors acknowledge in Section 6.3. The external survey [72] partially corroborates the LEG tailoring absence, and the authors are appropriately tentative in presenting the adoption sheet as a starting point rather than a validated instrument. For these reasons, the concern is about external validity and reproducibility, not about fraud or circularity. If the proposed expanded-sample re-run and inter-coder reliability check succeeded, the framework's cross-domain claim would be substantially stronger. Until then, the reader's CONDITIONAL verdict is the appropriate assessment, so no verdict change is needed.","tokens_in":21443,"tokens_out":4780,"duration_ms":73077,"concrete_test":"Recruit at least 8–10 additional legal practitioners (defense attorneys, prosecutors, in-house counsel, and judges) and re-run the Section 3.2 protocol, probing e-discovery and tailoring communications explicitly. Have two independent coders, blind to the paper's hypotheses, apply the Figure 1 sheet to all transcripts and report chance-corrected agreement (e.g., Cohen's kappa) for each Figure 4 sign. If any legal participant mentions AI for tailoring communications, or if inter-coder agreement falls below a pre-specified threshold (e.g., κ < 0.6), the paper's central cross-domain contrast and the reproducibility of its adoption-sheet signs would be unsupported by the current evidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest empirical claim is that four factors explain decision-maker adoption across domains, and its comparative evidence is concentrated in the Section 5 case studies and Figure 4. That evidence is least secure in the legal cell. Table 1 lists only two LEG participants (P05, P06), and Section 5.2 draws a domain-level conclusion—that tailoring communication to audience backgrounds is absent in law—from a cell of n=2, mitigated only by an external survey [72]. Section 3.3 describes open coding and reflexive thematic analysis but reports no codebook, no inter-coder reliability, and no rule for assigning the +/−/∼ signs in Figure 4. Because the Figure 4 signs are the paper's empirical payload, the framework's cross-domain explanatory power depends on those signs being both representative and reproducible. Additionally, Section 3.1 notes that some participants are researcher-insiders rather than practicing decision-makers; if the legal sample or the coding of the legal transcripts is unrepresentative, the claimed LEG contrast and the framework's generality lose their empirical footing. The authors do acknowledge sample-size limitations in Section 6.3, but the specific small-cell inference in Section 5.2 remains a load-bearing point that the current evidence does not fully secure.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper argues that decision-maker adoption of AI—defined as the voluntary and consistent consultation of AI tools within a workflow—should be distinguished from organizational adoption and from reliance on AI outputs. Based on 16 semi-structured interviews across journalism, law, medicine, and the public sector, the authors induce four factors that shape adoption: the decision-maker's background, perceptions of the AI model, consequences for the decision-maker, and perceived implications for other stakeholders. These factors are organized into an 'AI adoption sheet' and applied to two cross-domain case studies: e-discovery and tailoring communications to audiences of differing backgrounds. The paper claims that the framework explains inter-domain differences in adoption, and it offers recommendations for developers, organizations, and policymakers.","tokens_in":21670,"tokens_out":5929,"duration_ms":71164,"significance":"If its claims were fully supported, the paper would make a useful conceptual contribution to the human-AI interaction literature by shifting attention from reliance to the prior question of whether decision-makers adopt AI at all. The interview protocol, use-case taxonomy in the appendix, and the AI adoption sheet are transparent and potentially reusable, and the authors are explicit about several limitations. The cross-domain framing is valuable and goes beyond the mostly single-domain studies in prior work. However, the current evidence base is thin for the comparative claims: the legal cell has only two participants, several participants are researcher-insiders rather than practicing decision-makers, and the Figure 4 signs that carry the case-study analysis are not backed by a documented coding procedure. The contribution is therefore better described as a generative qualitative framework than as a confirmed empirical account of cross-domain adoption.","major_comments":[{"comment":"The cross-domain claims rest on a small and partly second-hand sample. Table 1 lists only two legal participants (P05 and P06) and five of the sixteen participants (P01, P06, P10, P13, P14) are listed with primary area 'Research' rather than domain practice, which Section 3.1 acknowledges by including 'researcher-insiders' as a participant category. Because the central contribution is a comparative analysis across domains, the small cell sizes and the inclusion of proxy voices matter directly: a domain-level conclusion drawn from two interviews, one of whom is a researcher-insider, is not securely grounded. Section 6.3 acknowledges the limited sample, but the acknowledgment does not by itself protect the specific legal-domain contrasts in Section 5.2, which should at minimum be reframed as exploratory.","section":"Section 3.1, Table 1"},{"comment":"The plus/minus/tilde signs in Figure 4 are the empirical payload of the case studies, but the manuscript reports no codebook, no unitizing or aggregation rule, and no inter-coder reliability check for converting open-ended interview content into these signs. Section 3.3 describes open coding and reflexive thematic analysis, which are appropriate for identifying themes, but it does not explain how individual mentions were combined into a single sign per factor per domain. Without a documented rule or agreement measure, the signs cannot be independently audited. The authors could either make the coding materials and aggregation procedure available or explicitly relabel the signs as interpretive summaries rather than coded evidence.","section":"Section 3.3, Figure 4"},{"comment":"The four-factor framework and the case-study explanations are derived from the same 16 interviews, so the case studies are in-sample illustrations rather than independent tests of the framework. For example, the e-discovery analysis in Section 5.1 uses participant quotes from the same interview corpus that generated the factors in Section 4. The paper should avoid wording such as 'highlighting how our factors help explain inter-domain differences' (Section 1) if the explanation is not tested against new data; the external surveys cited in Sections 5.1 and 5.2 provide partial triangulation, but they are invoked selectively and do not validate the framework as a whole.","section":"Sections 4 and 5"},{"comment":"The claim that tailoring communications to audiences of differing backgrounds is 'absent in LEG' is an inference from non-mention in only two legal interviews (P05 and P06), both of whom are listed in Table 1. In open-ended interviews, absence of mention is weak evidence for absence in the domain, and the external survey [72] cannot fully substitute for targeted probing of the legal participants. The paper should either soften this to 'not mentioned by our legal participants' or provide a more systematic elicitation procedure (for example, asking participants about a fixed list of use cases) if the domain-level contrast is to be load-bearing.","section":"Section 5.2, Table 1"}],"minor_comments":[{"comment":"Several typographical errors appear in the appendix, including 'Applicaiions', 'explinability', 'organizatioanl', 'Tehnology', 'availabtility', 'cpartners', and 'lieracy'; a careful proofreading pass is needed.","section":"Appendix A.2, Appendix B"},{"comment":"Recommendation 3 contains 'Oftent' instead of 'Often', and the sentence beginning 'AI systems often demand substantial resources' would benefit from a period or semicolon for readability.","section":"Section 6.1"},{"comment":"The use-case tables contain spelling errors such as 'reviwed', 'identifiy', 'pneumoonia', and 'radiolgists'; these should be corrected before publication.","section":"Tables 2-5"},{"comment":"The abstract's phrase 'the perceived implication of AI adoption' should be 'perceived implications' to match the factor name used in Section 4.4.","section":"Abstract and Section 4"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the paper has a clear conceptual contribution and unusually detailed supplementary materials, but the load-bearing comparative claims are not yet fully secured by the evidence. I would like the authors to either strengthen the legal cell and the coding documentation or substantially reframe the claims as exploratory and illustrative. If the authors can do that within the manuscript's scope, it could become suitable for publication in a human-centered AI venue; as it stands, the empirical overreach is the main obstacle."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's the short version: this is a well-written qualitative study that offers a plausible four-factor framework for why decision-makers adopt or reject AI tools, plus a practical 'AI adoption sheet.' The cross-domain framing is genuinely new relative to single-domain studies in medicine or law. The problem is that the empirical base is thin—16 interviews, only 2 from law—and the paper's comparative case studies, especially the legal contrast, lean on that weak cell. The Figure 4 signs are presented as the payoff but no inter-coder reliability check is reported. That's not a fatal flaw for a qualitative study, but it means the strongest claims are not fully secured.\n\nWhat the paper does well: it clearly distinguishes organizational adoption, decision-maker adoption, and reliance—a useful conceptual cleanup. The four factors (background, model perception, consequences, stakeholder implications) synthesize prior constructs well and map naturally onto existing technology-adoption literature. The adoption sheet is a concrete, actionable checklist that could genuinely help model developers and policymakers think through adoption barriers before deployment. The authors are transparent about sample limitations and the exploratory nature of the work. The interview excerpts are rich and illustrate the factors effectively.\n\nWhere it sags: the n=2 legal cell carries more weight than it should. Section 5.2's claim that tailoring communications is absent in law rests on two participants and an external survey. The in-sample case studies are illustrations, not independent validations. And the Figure 4 signs are essentially the authors' judgment calls without a documented coding process; a reader can't tell if another coder would assign the same signs. The abstract's word 'impact' suggests causality that a 16-person interview study can't support. These are proportional concerns—they don't sink the paper, but they do limit how much weight the cross-domain conclusions can bear.\n\nThe paper is honest, clearly argued, and cites relevant prior work without obvious gaps. It's a useful contribution for the human-AI interaction and responsible-AI community. I'd send it to peer review, but I'd expect the reviewers to ask for tempered claims and either more evidence or a clearer statement that the cases are hypothesis-generating. I'd cite it as a framework reference. Yes, bring it to reading group.","headline":"A plausible but thinly evidenced framework for AI adoption; worth refereeing, but the cross-domain claims need tempering.","tokens_in":22187,"tokens_out":3491,"would_cite":true,"duration_ms":36513,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Decision-maker adoption of AI is governed by four cross-domain factors—background, model perception, personal consequences, and stakeholder implications—and can be analyzed with an adoption sheet.","keywords":["decision-maker adoption","AI adoption","human-AI decision-making","semi-structured interviews","cross-domain analysis","AI adoption sheet","medicine","public sector"],"falsifier":"Run a larger, pre-registered survey of clinicians, lawyers, journalists, and public-sector workers that codes their stated adoption reasons against the four factors; if a substantial share of decisions is driven by reasons outside the four categories, or if the factors do not predict actual consultation behavior in a longitudinal deployment, the framework's completeness claim fails. A simpler disconfirmation would be finding that e-discovery tools are already routinely adopted by clinicians in a broad sample, since the medicine case study predicts the opposite.","tokens_in":21247,"feed_emoji":"🤖","tokens_out":4900,"duration_ms":57291,"temperature":0.7,"pith_summary":"Decision-maker adoption—the voluntary, consistent choice to consult an AI tool—is the missing layer between an organization buying AI and a human relying on its output. The paper claims this choice is governed by four cross-domain factors: the decision-maker's own background, their perception of the model, the consequences they personally face, and what they believe other stakeholders gain or lose. Based on 16 interviews across journalism, law, medicine, and the public sector, the authors propose an AI adoption sheet that turns these factors into diagnostic questions, and show through two case studies why similar tools are adopted in some domains and not others. The value is practical: if the framework holds, developers and policymakers can anticipate uptake before deployment rather than after.","feed_headline":"Four factors explain why decision makers adopt AI","feed_subtitle":"Interviews across medicine, law, journalism and the public sector yield a checklist for predicting uptake.","key_machinery":"The AI adoption sheet is the paper's central instrument: a short checklist of questions under the four factors (decision-maker background, model perception, consequences for the decision-maker, perceived implications for other stakeholders). It functions as an analytic lens for comparing use cases across domains and as a prospective tool for developers and policymakers to ask why a given decision-maker would or would not use a tool before it is built or deployed. The case-study charts turn the sheet into per-domain ratings—plus, minus, or tilde—showing how the same use case can get opposite adoption signals in different professions.","core_discovery":"The paper's central claim is that before any question of reliance—how often a decision-maker defers to AI—there is a prior choice of adoption, and that choice is systematically shaped by four factors that recur across domains. In journalism, law, medicine, and the public sector, decision-makers weigh their professional experience and personal biases, what they believe about the model's flaws, transparency, and capabilities, the legal, professional, political, and workload consequences they would bear, and the perceived impact on their organization and on the people their decisions affect. The authors use these factors to explain two concrete puzzles: e-discovery tools are adopted for research and analysis in journalism, law, and the public sector but rarely in medicine, while tools that tailor communication to different audiences are used in journalism, medicine, and the public sector but not in law. The explanation, summarized in plus/minus/tilde charts, is that the same factor pushes different ways in different domains—for example, liability concerns and distrust of opaque models weigh heavily against adoption in medicine, while easy verification makes e-discovery attractive in journalism.","pith_inferences":["A natural test is to convert the adoption sheet into a structured survey and administer it to a large sample across new domains such as education or finance; if the four factors fail to predict stated adoption, the framework needs revision.","The case studies imply that adoption barriers are often about the decision-maker's liability and the inability to verify outputs, not about the raw quality of the model; this suggests that improving verifiability and indemnification may matter more than accuracy gains.","The framework could be made dynamic: tracking the same decision-makers over time would test whether exposure, organizational support, and regulatory clarity shift adoption in the direction the interviews suggest.","For high-stakes domains, the sheet could be used as a pre-deployment checklist in procurement, turning a research finding into a governance artifact, though the paper itself leaves that empirical validation to future work."],"forward_implications":["Reliance studies that measure how often people defer to AI need to separate adoption from reliance, since non-adoption can masquerade as low reliance.","Model developers can use the adoption sheet before deployment to identify domains where a tool will stall because of liability, distrust, or stakeholder concerns.","Organizational adoption—a hospital buying a system or a firm mandating a tool—does not imply decision-maker adoption, and the two can be driven by different factors.","The same AI capability can face opposite adoption outcomes in different professions, so domain-blind deployment strategies are likely to fail.","Because perceptions change with exposure, adoption is not static; early negative impressions can shift as tools improve and organizations support them."],"supporting_citations":[{"why":"The prior legal-domain interview study the paper extends to show how role-based attitudes shape AI adoption.","marker":"[28]"},{"why":"The healthcare survey identifying seven constructs that influence professionals' adoption intentions; the paper's factors absorb and reorganize these.","marker":"[30]"},{"why":"The medical onboarding study that motivates the adoption gap by showing practitioners need more than accuracy to engage with AI.","marker":"[16]"},{"why":"A deployed clinical AI study showing peer endorsement and workflow matter for actual use, cited as evidence that adoption precedes reliance.","marker":"[32]"},{"why":"The hospital procurement survey that represents the organizational-adoption perspective the paper distinguishes from decision-maker adoption.","marker":"[50]"}],"fun_headline_variants":["Four factors decode decision makers' AI adoption","Behind every AI adoption choice: four key factors","Why doctors avoid AI but journalists don't: new study","AI adoption checklist: background, perceptions, consequences, stakeholders","Cross-domain AI adoption: what tips the scale for experts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework rests on 16 interviews—two in law and several with researcher-insiders instead of practicing decision-makers—being representative enough to support conclusions that generalize across entire professions.","fun_headline_variants_meta":{"raw":{"variants":["Four factors decode decision makers' AI adoption","Behind every AI adoption choice: four key factors","Why doctors avoid AI but journalists don't: new study","AI adoption checklist: background, perceptions, consequences, stakeholders","Cross-domain AI adoption: what tips the scale for experts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000899,"raw_usage":{"total_tokens":3854,"prompt_tokens":913,"completion_tokens":2941,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":2865}},"tokens_in":529,"tokens_out":2941,"duration_ms":25107,"temperature":1.0,"reasoning_tokens":2865,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:57:47.409663+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a larger, pre-registered survey of clinicians, lawyers, journalists, and public-sector workers that codes their stated adoption reasons against the four factors; if a substantial share of decisions is driven by reasons outside the four categories, or if the factors do not predict actual consultation behavior in a longitudinal deployment, the framework's completeness claim fails. A simpler disconfirmation would be finding that e-discovery tools are already routinely adopted by clinicians in a broad sample, since the medicine case study predicts the opposite.","supporting_citations":[{"cited_title":"AI Am Here to Represent You","cited_arxiv_id":null,"evidence_quote":"The prior legal-domain interview study the paper extends to show how role-based attitudes shape AI adoption."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The healthcare survey identifying seven constructs that influence professionals' adoption intentions; the paper's factors absorb and reorganize these."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A deployed clinical AI study showing peer endorsement and workflow matter for actual use, cited as evidence that adoption precedes reliance."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The hospital procurement survey that represents the organizational-adoption perspective the paper distinguishes from decision-maker adoption."}],"review_version":1}