{"id":"ee5ae785-34b7-4e09-a36e-8cec32f7adaa","arxiv_id":"2508.20137","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Only 14.1% of Basque Country media professionals have received AI training, mainly through self-learning, while newsroom staff lag behind technical and advertising roles.","lead":"Researchers surveyed 504 media workers in the Basque Country and found that only 14.1% have received any AI training, mostly self-directed. The study is a regional snapshot showing that traditional newsrooms lag behind digital, technical, and advertising roles in preparing for AI, and it argues for stronger in-house and academic training.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Self-selected web survey with no response rate: the 14.1% headline and subgroup gaps assume respondents represent the 5,000 Basque media workers; small-cell subgroup claims like technical staff (n=3) and international outlets (n=23) are also within sampling noise.","rationale":"The reader's weakest assumption identifies the same load-bearing issue: the survey's nonprobability sampling and missing response-rate information make the population generalization insecure. The headline 14.1% figure and the training-gap comparisons are the paper's central claims, and both require the sample to represent roughly 5,000 Basque media professionals. With no response denominator and no benchmarking against the census data cited in Section 3, self-selection could plausibly move the headline estimate by several percentage points, which would alter the paper's main message. The small cell sizes in Tables 1 and 2 reinforce this concern: several subgroup claims rest on single-digit denominators, and the absence of confidence intervals or significance tests means the reader cannot tell which differences are real. These are addressable with additional reporting and reanalysis, not fatal flaws, so the conditional verdict remains appropriate. The paper's qualitative interview component and its descriptive breakdowns remain useful, but the quantitative generalizations need the proposed verification.","tokens_in":12058,"tokens_out":4289,"duration_ms":42512,"concrete_test":"Contact the authors or the Basque Journalists' Association for the number of invitations sent and the size of the Open Communication Guide's professional list; compute a response rate. Then compare respondent distributions (sex, media type, job role, broadcasting scope) to the Basque Government (2022) census totals. If the response rate is below 20% or respondent profiles deviate by more than 10 percentage points on any key variable, reweight the sample and recompute the 14.1% trained rate and the subgroup rates in Tables 1 and 2. If the reweighted overall rate shifts by more than 3 percentage points, or if key subgroup differences lose significance under Fisher exact tests, the paper's quantitative conclusions should be downgraded from generalizable findings to hypothesis-generating observations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central generalization—only 14.1% of Basque media professionals have received AI training—depends on a nonprobability sample. Section 3 reports 504 responses collected via an online panel recruited through the Basque Government's Open Communication Guide and the Basque Journalists' Association, but gives no invitation count, no response rate, and no comparison of respondents to the Basque Government census (Basque Government, 2022) used to justify the ±4.15% margin of error. That margin of error is only valid for a simple random sample; under self-selection, the 14.1% estimate can be biased by an unknown amount. If the sample over-represents digitally active or AI-curious workers, the headline rate is likely an overestimate for the population, and the 'newsroom staff are behind' gap may be misstated. Secondary subgroup claims are additionally fragile: 'technical roles leading' rests on 2 trained respondents out of 3 total; the international-vs-local difference (21.7% vs 12.2%) has a standard error of roughly 9.3 percentage points, so the 'nearly double' pattern is not distinguishable from noise. No significance tests or confidence intervals are reported for any Table 1 or Table 2 comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This mixed-method study surveys 504 media professionals in the Basque Country and interviews six media-technology leaders to assess the current state of AI training in regional newsrooms. The central descriptive finding is that only 14.1% of respondents have received any AI training, with most of that training self-directed. The paper further reports that AI training is more common among professionals in digital, advertising, technical, and managerial roles than among editors and presenters, and more common in international and larger organizations than in local outlets. Qualitative interviews are used to contextualize these patterns, highlighting the need for continuous, ethically grounded training. The authors conclude that urgent investment in AI literacy is required, especially for traditional newsroom staff and local media.","tokens_in":12296,"tokens_out":4826,"duration_ms":46582,"significance":"If the descriptive estimates were representative of the roughly 5,000 media professionals in the Basque Country, this would be a valuable, policy-relevant contribution to an understudied regional media system. The paper's strengths include a comparatively large sample for a regional study, detailed disaggregation by media type, role, experience, and scope, transparent reporting of raw cell counts that lets readers judge small-cell fragility, and qualitative triangulation through interviews with innovation leaders. The main limitation is that the survey uses a nonprobability online panel with no response rate or representativeness check, so the headline figure and subgroup gaps should be treated as sample statistics rather than population estimates. The absence of inferential statistics (confidence intervals or significance tests) further weakens the comparative claims. As an exploratory study the paper is useful, but the framing overstates the evidentiary basis.","major_comments":[{"comment":"The reported margin of error of ±4.15% is justified only under simple random sampling from a target population of about 5,000, but participants were recruited through an online panel advertised via the Open Communication Guide and the Basque Journalists' Association, with no response rate, invitation count, or comparison to the Basque Government census used to define the population. This margin of error does not quantify the uncertainty around the 14.1% headline figure under self-selection bias, which could shift the estimate substantially. Please either provide demographic or other evidence that the sample is representative of the ~5,000 media workers, or explicitly reframe the results as describing the 504 respondents only and remove the margin-of-error claim.","section":"Section 3, Materials and Methods"},{"comment":"The paper repeatedly claims 'significant differences' (e.g., 'Significant differences in AI training are observed across different types of media' and 'international outlets reporting nearly double the training rate (21.7%) of local outlets (12.2%)') without any significance tests, confidence intervals, or effect sizes. For the international-vs-local comparison, the approximate standard error of the difference is about 9.3 percentage points, so the observed 9.5-point gap is within sampling noise; a similar problem affects several small cells, such as technical workers (2 of 3 trained) and presenters (6 trained in total). The authors should either add appropriate inferential statistics (including confidence intervals for proportions) or soften all comparative claims to avoid overstating differences that the sample cannot support.","section":"Section 4.1, Tables 1 and 2"},{"comment":"The claim that leadership-role holders receive 'twice as much training' (19.9% vs. 9.4%) and the subsequent inference that 'employees with longer tenures' are more likely to be trained are based on bivariate cross-tabulations only. Tenure, responsibility role, age, and media type are correlated, so the independent contribution of each cannot be assessed without a multivariate model or at least stratified analysis. Please add such an analysis or explicitly caveat that these are unadjusted associations and avoid causal phrasing such as 'likely factors influencing this trend.'","section":"Section 4.1, Table 2 and discussion of leadership roles"}],"minor_comments":[{"comment":"The text states that among trained respondents, '52.1% did so through both personal and company efforts, 39.4% relied solely on self-directed learning, and only 8.5% received training exclusively from their employer,' but these percentages cannot be directly derived from Table 5, which reports percentages of all respondents rather than of the 71 trained participants. Please reconcile the text and table by stating the denominator explicitly and, if possible, adding a row for 'any training' with mutually exclusive categories.","section":"Section 4.2"},{"comment":"The text refers to 'Table 4' in the sentence 'This trend aligns with the types of roles that reported higher training levels (Table 4),' but no Table 4 appears; the following table is numbered Table 5. Please renumber the tables sequentially.","section":"Section 4.2"},{"comment":"Decimal separators are used inconsistently: some percentages use commas (e.g., '12,2') and others use periods (e.g., '6.3'). Please standardize throughout, especially within table cells.","section":"Tables 1-3 and 5"},{"comment":"The survey instrument's 'self-study' category is undefined; specify whether it includes reading books, online tutorials, webinars, or other informal learning formats, since this affects interpretation of the self-directed learning finding.","section":"Section 4.1"},{"comment":"The selection criteria for the six interviewed media leaders are not explained beyond their being 'senior figures' at 'prominent' outlets. Please state how these outlets span the regional media landscape (e.g., ownership, language, reach, medium) and whether saturation was considered in the analysis.","section":"Section 3, Interviews"},{"comment":"The paper lacks a dedicated limitations subsection. In addition to the sampling and inference issues raised above, the data were collected in May-June 2024, a period of extremely rapid AI tool development; the authors should acknowledge that the reported training rates may already be outdated and that the situation is likely evolving.","section":"Section 5, Discussion"}],"recommendation":"major_revision","confidential_remarks":"This is a post-publication version of a Frontiers in Communication article, so the revision request is advisory. The paper's empirical contribution is useful, but the framing—especially the margin-of-error claim and the repeated use of 'significant' without statistical tests—overstates the representativeness and robustness of the findings. A revised version that adds a limitations section, reports confidence intervals or significance tests for the key comparisons, and tempers the population-level generalizations would be much stronger. The small-cell issues (e.g., technical workers, international outlets) should be explicitly flagged in the text, not only in tables."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read: this is a straightforward descriptive survey, not a high-stakes claim. The new thing is the dataset: 504 Basque media professionals, the first region-specific numbers on AI training. The 14.1% headline is plausible and probably in the right ballpark, though the absolute figure may be optimistic given the self-selected sample. The paper does a decent job of breaking down training by media type, role, experience, and organizational scope, and the six interviews add useful context. The writing is clear and the literature review is adequate.\n\nThe soft spot is the sampling and the inference. Section 3 says the margin of error is ±4.15% based on the 5,000 population, but this is an online opt-in panel recruited through a government guide and a professional association. No response rate, no comparison to the census, no weighting. The margin-of-error claim is only valid for a simple random sample; under self-selection, the 14.1% could be over- or understated by an unknown amount. The authors also say \"significant differences\" but report no significance tests or confidence intervals. Several subgroup cells are tiny: technical roles have n=3 (2/3 trained), presenters n=6, copywriters n=1. The claim that international outlets have \"nearly double\" the training rate of local outlets rests on 5/23 vs 10/82, and the standard error difference is large. That pattern may be real, but the paper can't support it as stated.\n\nThese issues are addressable. The paper would be stronger with a limitations paragraph, no margin-of-error claim, explicit confidence intervals or at least a caveat, and a clear statement that these are sample-specific descriptive frequencies. The central descriptive point—that AI training among Basque media workers is low and largely self-directed—is likely directionally right even if the exact percentage is imprecise.\n\nFor a reader: if you work on journalism education or media policy, this is a useful baseline for one region. It doesn't test an intervention and the regional scope limits generalizability. It deserves a serious referee because the data is new and the flaws are fixable, but I'd send it back for statistical reporting and an honest limitations section.","headline":"A genuinely new regional dataset on AI training for media workers, but the headline 14.1% and the subgroup comparisons outrun what the self-selected sample can support.","tokens_in":12791,"tokens_out":2319,"would_cite":false,"duration_ms":20179,"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":"Basque media professionals are largely untrained in AI, and the gap is sharpest inside newsrooms.","keywords":["artificial intelligence training","journalism education","media professionals","newsroom skills gap","AI ethics","digital transformation","Basque media","self-directed learning"],"falsifier":"A stratified random sample (or full census) of Basque media workers drawn from employer payrolls rather than voluntary online lists that finds the trained share materially higher than 14.1%, or company training logs showing that more than 8.5% of employees have completed employer-provided AI training, would overturn the paper's headline gap.","tokens_in":11910,"feed_emoji":"🎓","tokens_out":6871,"duration_ms":56867,"temperature":0.7,"pith_summary":"Drawing on a survey of 504 media professionals and interviews with six innovation leaders in the Basque media market, this paper sets out to measure how prepared journalists are for artificial intelligence. It reports that only 14.1% of respondents have received any AI training, that most of that training was self-directed rather than employer-provided, and that training is concentrated in technical, managerial, advertising, and corporate-communication roles while newsroom editors and presenters lag behind. The paper argues that this pattern, combined with higher training rates at larger and internationally focused outlets, risks widening an existing technological gap between major and local media, and it concludes that sustained in-house and university training covering both technical and ethical dimensions is urgent. If the finding holds, the AI transition in journalism is currently being driven by the least-trained part of the workforce.","feed_headline":"Only 14.1% of media professionals have AI training","feed_subtitle":"A survey of 504 Basque media workers finds AI training is rare, mostly self-taught, and weakest in newsrooms.","key_machinery":"The analytical engine is a three-way crosstabulation of self-reported AI training by type of media (press, radio, television, digital, press office, advertising agency), broadcasting scope (local, regional, national, international), and job role (editor, presenter, technical, management, and others), computed from the 504 survey responses and interpreted through six semi-structured interviews with innovation and technology leaders. The main organising distinction is between core journalistic tasks and technical, managerial, or persuasive roles: the closer a role is to traditional journalism, the lower the training rate. A second distinction, between employer-provided and self-directed training, shows that company programmes reach only a small minority and mostly those already motivated to learn.","core_discovery":"The central discovery is an inverse relationship between AI training and proximity to core journalistic work. Editors (13.8%) and presenters (6.3%) show the lowest training rates, while technical staff (66.7%), management (25.0%), public relations (24.0%), and advertising (44.4%) show the highest; only 8.5% of all respondents received employer-provided training, compared with 12.9% who trained on their own. Internationally based outlets report nearly double the training rate of local outlets (21.7% vs. 12.2%), and digital-native outlets (25.8%) far outpace television (5.8%), radio (10.8%), and print (9.6%). The authors read this as evidence that AI adoption will deepen the digital divide between large and small outlets unless training becomes systematic and includes ethical as well as technical content.","pith_inferences":["If the role gradient is not specific to the Basque market, similar newsroom-versus-technical divides should appear in other mid-sized media markets; a multi-region replication survey would test this directly.","The dominance of self-directed learning suggests a selection effect, so simply adding courses may not reach editors and presenters unless training is embedded in daily workflows.","The international-outlet advantage could reflect either more resources or greater exposure to AI-driven platforms; distinguishing those causes would change where policy should intervene.","Company training records, rather than self-reports, could independently verify whether the 8.5% employer-provided-training figure understates actual provision."],"forward_implications":["As AI tools enter editing and presenting workflows, the people making publication decisions will be the least trained to detect errors, bias, or fabrication, so the risk of unchecked AI output rises.","Digital-native and international outlets will pull further ahead of local and traditional media, widening a technology gap that already exists from earlier digitalisation.","Because only 8.5% of respondents had employer-provided training, current workplace policies alone are unlikely to close the gap.","Ethics-focused AI training needs to be embedded in formal journalism education, not left to self-study, which has reached only a small minority."],"supporting_citations":[{"why":"Supplies the AI-literacy concept that frames the paper's call for journalism training.","marker":"Deuze and Beckett, 2022"},{"why":"Documents journalists' concerns about generative AI, used to explain newsroom scepticism and the training lag.","marker":"Peña-Fernández et al., 2023a"},{"why":"Presents expert and academic perceptions that AI training is a recognised need in journalism.","marker":"Noain Sánchez, 2022"},{"why":"Prior work on training journalists for the AI era, grounding the call for formal education pathways.","marker":"Larrondo-Ureta and Peña-Fernández, 2024"},{"why":"Reviews teaching experiences with AI in journalism, supporting the university-training recommendation.","marker":"Gómez-Diago, 2022"},{"why":"Contextualises the urgency of training through an analysis of ChatGPT use and risks in journalism.","marker":"Gutiérrez-Caneda et al., 2023"},{"why":"Supports the ethical dimension of the training need by discussing trustworthy AI in journalism.","marker":"Opdahl et al., 2023"}],"fun_headline_variants":["Only 14.1% of media workers have AI training","AI training gap: newsrooms worst off in Basque media","Self-taught majority: AI skills sparse among journalists","Digital divide in AI training hits local outlets","Ethics absent from rare AI training for journalists"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey sample is assumed to stand in for the roughly 5,000 media workers in the Basque Country, but respondents were recruited through an open government listing and a professional association without a documented response rate, so a self-selection bias toward digitally active or AI-curious workers could make the 14.1% trained figure an overestimate.","fun_headline_variants_meta":{"raw":{"variants":["Only 14.1% of media workers have AI training","AI training gap: newsrooms worst off in Basque media","Self-taught majority: AI skills sparse among journalists","Digital divide in AI training hits local outlets","Ethics absent from rare AI training for journalists"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000607,"raw_usage":{"total_tokens":2792,"prompt_tokens":875,"completion_tokens":1917,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":1841}},"tokens_in":491,"tokens_out":1917,"duration_ms":13014,"temperature":1.0,"reasoning_tokens":1841,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:51:17.262691+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A stratified random sample (or full census) of Basque media workers drawn from employer payrolls rather than voluntary online lists that finds the trained share materially higher than 14.1%, or company training logs showing that more than 8.5% of employees have completed employer-provided AI training, would overturn the paper's headline gap.","supporting_citations":[],"review_version":2}