{"id":"f05f2a16-d2cd-481f-96ca-8e716e7dd605","arxiv_id":"2509.01824","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"89.88% of Basque journalists surveyed believe AI will considerably or significantly increase disinformation risks.","lead":"A survey of 504 journalists in the Basque Country finds that nearly 90 percent believe artificial intelligence will considerably or significantly increase disinformation risks. The study maps which risks worry journalists most and how experience and AI use shape those worries.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Voluntary opt-in sample with no reported response rate leaves the 89.88% prevalence estimate unprotected against nonresponse bias; the reported ±4.15% margin of error assumes simple random sampling and does not cover this.","rationale":"The reader identified representativeness/self-selection as the weakest assumption; the manuscript itself concedes this possibility and gives no response-rate or frame-comparison data. I find this the most load-bearing concern because the headline number is a prevalence estimate, and nonresponse bias directly affects the numerator. The reported margin of error is misleading in this design, reinforcing the concern. I also considered the absence of the full questionnaire and the minor inconsistency in Table 2 co-occurrence percentages, but these are secondary: the questionnaire issue is about reproducibility, and the co-occurrence typo does not affect the central claim. The statistical tests for associations are appropriately adjusted and their effect sizes are small; they are not the main basis for the paper's contribution. Thus the reader's CONDITIONAL verdict is appropriate and no further adjustment is needed.","tokens_in":15351,"tokens_out":9260,"duration_ms":101955,"concrete_test":"Conduct a nonresponse bias analysis using the sampling frame: obtain the number of journalists invited through the Open Communication Guide and the Basque Association, compute the response rate, and compare respondents to the frame on available covariates (gender, media type, province, association membership). Then re-estimate the 89.88% prevalence with inverse-probability weights and under a sensitivity scenario in which nonrespondents are, say, half as likely to choose 'considerably/significantly' as respondents. If the weighted and sensitivity-adjusted lower bound remains above ~75%, the headline 'large majority' is robust; if it dips substantially, the central claim would need to be reported as conditional on unmeasured self-selection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is a population prevalence estimate: 89.88% of Basque journalists think AI will considerably or significantly increase disinformation risk. The data, however, come from a voluntary opt-in survey distributed through official directories and the Basque Association of Journalists. The Methods give no denominator: neither the number of journalists invited nor the response rate is reported. The Limitations section itself states that 'individuals with stronger views on AI and disinformation may have been more motivated to respond.' For a prevalence estimate, that admission is directly load-bearing: if even a modest fraction of the roughly 4,500 non-respondents held weaker risk perceptions, the true value could be materially below 89.88%. The reported margin of error (±4.15%, 95% CL) is derived from a simple random sampling formula and therefore captures only sampling error, not selection bias, nonresponse, or frame undercoverage. The associations by experience and AI use are secondary, but they too are vulnerable if the decision to participate is correlated with AI attitudes. No auxiliary comparison between respondents and the frame is provided beyond total employment (~5,000), so the size of the bias is unbounded in the manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a quantitative survey of 504 journalists in the Basque Country, conducted in May–June 2024, about their perceptions of artificial intelligence (AI) and disinformation risks. The main descriptive finding is that 89.88% of respondents believe AI will 'considerably' or 'significantly' increase disinformation risks. The authors also examine associations between perceived risk and gender, years of professional experience, type of media, and respondents' use of AI; they report significant associations with experience and AI use. A further section ranks perceived risk categories and analyzes co-occurrences of risk pairs. The manuscript concludes that journalists view AI as a serious threat to information integrity, with relevant differences by experience and technology familiarity, and discusses implications for training and AI literacy.","tokens_in":15645,"tokens_out":6089,"duration_ms":64640,"significance":"If interpreted as a description of the surveyed sample, the paper provides useful, localized evidence about journalistic risk perceptions in a region with a distinct media landscape. Strengths include the use of appropriate bivariate statistical tests with effect sizes, adjustment for multiple comparisons, correspondence analysis for visualization, and a limitations section that acknowledges self-selection. The paper's central claim, however, is a population prevalence estimate for approximately 5,000 Basque journalists, and that inference is not supported by the reported recruitment and response information. The co-occurrence analysis, though secondary, also contains reporting inconsistencies. The topic is timely and the dataset could be a valuable contribution, but the manuscript needs to either supply more sampling information or substantially qualify its prevalence claims.","major_comments":[{"comment":"The headline result—89.88% of journalists believing AI will considerably or significantly increase disinformation risk—is presented as a population prevalence (e.g., 'The majority of professionals perceive...'). However, the Methods do not report the number of journalists invited or the response rate; recruitment is described only as identification through the Open Communication Guide and collaboration with the Basque Association of Journalists. The Limitations section itself acknowledges that voluntary participation may attract individuals with stronger views. The reported ±4.15% margin of error is a simple-random-sampling formula and does not quantify nonresponse or self-selection bias. Because the central claim is a prevalence estimate, this is load-bearing: a modest difference between respondents and non-respondents could materially move the 89.88% figure. Please report the invitatio","section":"Sections 2 and 3.1"},{"comment":"The co-occurrence table contains internal inconsistencies that undermine the quantitative summary of RQ3. The counts listed sum to 501, not 504, despite the note stating they are based on 504 respondents. Several percentages do not match the counts: 17/504 = 3.37%, not 3.99%; 69/504 = 13.69%, not 13.77%. The text states 'Among the journalists who selected two risks, 50.5% (n = 253) ...', but 253/504 = 50.2%, and the denominator of 'journalists who selected two risks' is not reported. Please clarify how respondents who selected only one risk, or only 'none other', are handled, and recompute all percentages with a precisely defined denominator. The top-pair ranking is likely unaffected, but the current presentation is not reproducible.","section":"Section 3.2, Table 2"},{"comment":"The discussion suggests that 'direct experience and familiarity with AI may foster a more nuanced understanding of both its potential and its limitations.' This causal interpretation goes beyond the cross-sectional association reported in Section 3.1. The inverse relationship between AI use and risk perception is equally compatible with selection effects (journalists who perceive lower risk are more likely to adopt AI) or with unmeasured confounding (age, digital skills, newsroom resources). Please temper the causal language and explicitly acknowledge the direction-of-causation problem. This issue is load-bearing for the secondary claim about AI use and risk perception.","section":"Section 4 (Discussion)"}],"minor_comments":[{"comment":"Typo: 'inr oads' should be 'inroads'.","section":"Introduction"},{"comment":"The corresponding author email is given as 'urko.pena@ehu.eus@ehu.eus'; the duplicate domain is likely a typo.","section":"Author affiliations"},{"comment":"The binary grouping of 'considerably' and 'significantly' into a high-risk category is used for the headline 89.88%. Please clarify whether this aggregation was planned a priori or decided after data collection, and consider noting that the scale anchor 'significantly' may be read ambiguously by some readers.","section":"Section 3.1"},{"comment":"The text says 'The heatmaps presented in this study were generated using ChatGPT 4.1', but the figures appear to be percentage distributions and correspondence-analysis plots, not heatmaps. Please correct or clarify this statement.","section":"Figure 1 and Methods"},{"comment":"The citation 'Peña-Fernández et al. (2023)' is used for two different works (Peña Fernández, Meso Ayerdi, Larrondo Ureta, & Díaz Noci; and Peña-Fernández, Peña-Alonso, & Eizmendi-Iraola). Please disambiguate.","section":"References"},{"comment":"Some online references have no access date (e.g., Basque Government Open Communication Guide, Brundage et al. 2018). The Methods should state when the survey link was active and when the frame was last accessed.","section":"References and Methods"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a published author-accepted version, and the referee report addresses the arXiv manuscript. The central problem is the gap between the sample statistic and the population-level inference; this is fixable by adding response-rate information, comparing the sample to the frame, and/or softening the claims. The co-occurrence table also needs a careful numerical correction. No concerns about research misconduct are raised."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read.\n\nThe headline number—89.88% of Basque journalists saying AI will considerably or significantly increase disinformation risk—is the paper's core, and it's a fine descriptive result for the 504 people who answered. What's actually new is that no one had quantified this for the Basque media system, and the associations with years of experience and with AI use are genuine additions to the literature, even if the effects are modest. The co-occurrence analysis of perceived risk pairs is a useful small step.\n\nThe paper does several things right. It reports effect sizes, adjusts p-values for multiple comparisons, and acknowledges the voluntary-sample problem in the limitations. The methods are described well enough to know what was done, though the questionnaire itself and the raw data are not provided.\n\nThe soft spots are real, and they center on the gap between the sample and the population. The authors give no denominator: we don't know how many journalists were invited, so there is no response rate. The ±4.15% margin of error comes from a simple random sampling formula and does not cover nonresponse or self-selection. The paper's own limitation paragraph admits people with stronger views may have been more likely to respond. For a prevalence claim like 89.88%, that matters. If even 10% of the roughly 4,500 non-respondents had weaker risk perceptions, the true number could be meaningfully lower. So the stress-test note is on target, and I'd add that the AI-use association is interpreted as 'familiarity breeds nuance,' but the data are cross-sectional and could just as easily reflect that people who use AI are different in other ways.\n\nNone of this sinks the paper. The descriptive claim about the sample is supported, and the associations are presented cautiously. It's a regional study, not a field-shaping one, but it's competent and useful for anyone working on journalist attitudes toward AI. I'd send it to review (it's already published, but if this were a submission, it deserves referee time, not a desk reject), with the main request being a response rate, a comparison of respondents to the frame, and robustness checks on the grouping decisions.\n\nI'd bring it to a reading group for a methods discussion on survey representativeness.","headline":"A solid regional survey whose headline prevalence number is credible for the sample but not fully protected against nonresponse bias.","tokens_in":16048,"tokens_out":2289,"would_cite":true,"duration_ms":25595,"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 survey of 504 Basque Country journalists finds that 89.88% believe artificial intelligence will considerably or significantly increase the risk of disinformation.","keywords":["artificial intelligence","disinformation","risk perception","journalism","news media","survey","deepfakes","Basque Country"],"falsifier":"A random probability-based survey of Basque journalists—using a complete employment list with quotas rather than voluntary response—that finds either a substantially lower majority (e.g., below 75%) reporting AI will considerably/significantly increase disinformation, or no monotonic increase in that perception with years of experience, would directly contradict the paper's central claims.","tokens_in":15330,"feed_emoji":"⚠️","tokens_out":7078,"duration_ms":68207,"temperature":0.7,"pith_summary":"This paper tries to establish, with quantitative evidence, how journalists in the Basque Country perceive the impact of artificial intelligence on disinformation. Based on a structured survey of 504 journalists conducted in 2024, it reports that 89.88% of respondents believe AI will considerably or significantly increase disinformation risks. The perception is consistent across gender and media type, but statistically significantly stronger among journalists with more professional experience and weaker among those who use AI more frequently. The two most frequently cited risks are difficulty detecting false content and deepfakes (37.9%) and obtaining inaccurate data (33.33%), and about half of those selecting two risks paired these two concerns. The paper matters because understanding professional risk perception is a precondition for designing training, ethical guidelines, and regulation as AI becomes embedded in newsrooms.","feed_headline":"89.9% of journalists see AI as disinformation risk","feed_subtitle":"Survey of 504 Basque Country journalists finds risk perception grows with experience, dips among heavy AI users.","key_machinery":"The empirical machinery is an ad hoc structured survey administered to 504 Basque Country journalists between May and June 2024, analyzed with Fisher's exact tests, Cramér's V effect sizes, and correspondence analysis (CA). The decisive analytic device for the risk-hierarchy claim is a 4x4 co-occurrence matrix that records which pairs of risks each respondent selected (up to two), revealing that detection difficulty and data inaccuracy co-occur in over half of all paired responses. The CA biplots and the p-values after Holm/Benjamini-Hochberg corrections carry the statistical argument that experience and usage are associated with risk perception.","core_discovery":"On its own terms, the paper's discovery is that a large majority of working journalists (89.88%, n=453 of 504) classify the impact of AI on disinformation as 'considerably' or 'significantly' increasing, with only 10.12% choosing 'minimally' or 'not at all.' Fisher's exact tests show no significant association with gender or media type, but a significant positive association with years of professional experience (Cramér's V = 0.144, small-to-medium) and a significant negative association with frequency of AI use (Cramér's V = 0.120): heavy users perceive less risk. Correspondence analysis supports both patterns, with experience aligned toward high-risk categories and non-users aligned toward","pith_inferences":["If the usage-risk relationship is causal (familiarity reduces alarm), then the perceived disinformation threat may decline as AI becomes standard practice; a longitudinal replication could test whether the 89.88% figure drops over time.","Because the survey used official directories and a professional association for recruitment, the sample may over-represent established professionals; an independent probability sample could reveal whether the experience gradient is a genuine generational effect or a recruitment artifact.","The Basque Country's distinctive multilingual media landscape may shape results; extending the same instrument to other peripheral or multilingual regions would clarify whether the observed consensus is regional or general.","The co-occurrence pattern suggests a testable extension: experimentally providing journalists with deepfake-detection training should, if the perceived linkage holds, also reduce their concern about inaccurate data."],"forward_implications":["If nearly nine in ten journalists are right to fear AI-driven disinformation, newsrooms face a pressing need for deepfake detection tools and verification protocols tailored to the profession's stated top risk.","The significant positive association between years of experience and risk perception implies that veteran journalists may be the most cautious voices; training and change-management efforts may need to address their specific concerns.","The significant negative association between AI use and perceived risk suggests that hands-on familiarity with AI tools may make journalists less alarmed, potentially a moderating force as adoption grows.","The co-occurrence of detection difficulty with data inaccuracy indicates that solutions should treat verification and data quality as a single integrated problem rather than separate issues.","The absence of significant gender or media-type differences suggests a broadly shared professional consensus, which could support collective professional norms or policy positions on AI governance."],"supporting_citations":[{"why":"Provides the Open Communication Guide directory used to identify active media outlets and define the sampling frame.","marker":"(Basque Government, n.d.)"},{"why":"Collaborated in distributing the survey and supporting participation, shaping the sample.","marker":"(Basque Association of Journalists—Basque College of Journalists, n.d.)"},{"why":"Prior study of journalists' discourse on generative AI and disinformation; the current findings are framed as consistent with its conclusions.","marker":"(Peña-Fernández et al., 2023)"},{"why":"Establishes the concept of AI as a 'weapon of mass disinformation' that anchors the severity of the perceived risk.","marker":"(Kertysova, 2018)"},{"why":"Supplies evidence that deepfakes are a prominent risk, supporting the highest-ranked risk category.","marker":"(Lundberg & Mozelius, 2025)"},{"why":"Documents how media frame data risks, supporting the second-most-cited risk of inaccurate data.","marker":"(Nguyen, 2023)"},{"why":"Provides the threshold criteria for interpreting Cramér's V effect sizes as small-to-medium.","marker":"(Cohen, 1988)"},{"why":"Prior work linking exposure and familiarity with AI to more nuanced risk perceptions, used to interpret the inverse usage-risk association.","marker":"(Gutiérrez-Caneda et al., 2024)"}],"fun_headline_variants":["Veteran journalists more wary of AI disinformation than newcomers","Heavy AI users among journalists see lower disinformation risk","Journalists who use AI more perceive less disinformation risk","Experience boosts journalists' AI-disinformation fears; AI use dampens them","453 of 504 journalists fear AI will fuel disinformation"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The paper assumes that the 504 journalists who voluntarily responded through official directories and a professional association represent all ~5,000 journalists in the Basque Country; if self-selection attracted those with stronger views on AI, the 89.88% figure could be inflated.","fun_headline_variants_meta":{"raw":{"variants":["Veteran journalists more wary of AI disinformation than newcomers","Heavy AI users among journalists see lower disinformation risk","Journalists who use AI more perceive less disinformation risk","Experience boosts journalists' AI-disinformation fears; AI use dampens them","453 of 504 journalists fear AI will fuel disinformation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001395,"raw_usage":{"total_tokens":5484,"prompt_tokens":753,"completion_tokens":4731,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":4660}},"tokens_in":497,"tokens_out":4731,"duration_ms":35308,"temperature":1.0,"reasoning_tokens":4660,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:07:36.291342+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A random probability-based survey of Basque journalists—using a complete employment list with quotas rather than voluntary response—that finds either a substantially lower majority (e.g., below 75%) reporting AI will considerably/significantly increase disinformation, or no monotonic increase in that perception with years of experience, would directly contradict the paper's central claims.","supporting_citations":[],"review_version":1}