{"id":"01d1268e-92b8-4d43-976b-c9ae22272a25","arxiv_id":"2611.04278","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"An estimated 62% of the 1000 largest companies worldwide did not publicly acknowledge any AI risks in 2026, and companies generally failed to acknowledge the risks that experts rated as most severe.","lead":"The paper finds that an estimated 62% of the world's 1000 largest companies do not publicly acknowledge any AI risks, and that companies rarely acknowledge the risks experts judge most severe, such as power centralization or competitive dynamics. The result provides the first systematic measurement of AI risk disclosure across all major sectors and AI roles, establishing a baseline for designing regulatory frameworks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Company-level risk classification relies on a single human reviewer aggregating LLM output, with no reported inter-rater reliability at that level, which could affect the 62% estimate.","rationale":"The reader's weakest_assumption concerns the proxy validity of public disclosure. I identify a separate, more technical concern about measurement reliability at the company‑level aggregation stage. While the paper's methodology is otherwise rigorous, the absence of inter‑rater reliability for the final company‑level classification is a gap that could affect the headline estimate. The concern is not fatal—the authors' careful validation of the LLM step and transparency about limitations are commendable—but it warrants a condition before full acceptance: either provide the missing reliability check or explicitly discuss this as a limitation. Thus the verdict should shift from ACCEPT to CONDITIONAL, requiring documentation of inter‑rater agreement at the company level.","tokens_in":52564,"tokens_out":4952,"duration_ms":50675,"concrete_test":"Have a second independent human reviewer, blind to the first reviewer's decisions, independently determine for a random sample of 40 companies (20 from ranks 1–100, 20 from the random lower sample) whether each of the 24 AI risks was acknowledged, using the same LLM passage outputs and the original source documents (as needed). Compute Cohen's κ for the dichotomous company-level variable \"any risk acknowledged\" and for each risk individually. If κ for \"any risk\" falls below 0.70, the single-reviewer aggregation is a measurable source of uncertainty that should be reported and, if possible, corrected via consensus or a second coder.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's headline estimate—that 62% of the largest 1000 companies did not acknowledge any AI risk—depends on a final step where a single human reviewer examined all LLM-classified passages for each company and decided whether the company had acknowledged each risk. The paper validates the LLM passage-level classifications against independent human coders (Cohen's κ 0.64–0.88), which is good, but does not report any reliability check for the reviewer's company-level aggregation. The Methods section (§\"Data classification\", final paragraph) describes this as: \"A human reviewer considered all available evidence to determine whether a company acknowledged each risk.\" No inter‑rater statistic, not even a sample of doubly‑coded companies, is provided. Because the central claim is a company‑level count, any systematic error or idiosyncrasy in this single reviewer's judgments (e.g., overly conservative interpretation of what counts as an acknowledgement, or overlooking a risk mentioned in an auxiliary document) would propagate directly into the 62% figure. The LLM validation shows the raw material is reliable, but the synthesis step is uncalibrated. This is a methodological gap, not a known error, but it is load‑bearing because the headline is the paper's most widely cited result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"This study measures how many of the largest 1000 companies worldwide publicly acknowledge AI risks, using a manual content analysis of 539 public documents (annual reports, ESG reports, AI governance documents, etc.) collected in early 2026. The central finding is that 62% (95% CI: 54%–70%) of the largest 1000 companies do not acknowledge any AI risk; the top five acknowledged risks are discrimination, weapons & cyberattacks, fraud & scams, loss of privacy, and devaluation of human creativity. The paper also compares company acknowledgements with expert judgments from a companion Delphi study (Saeri et al., 2026) and finds that the risks experts rate as most severe are rarely acknowledged, that companies in more vulnerable sectors are more likely to acknowledge a risk (but still usually do not), and that the actors experts judge most responsible are no more likely to acknowledge the risk than others. The paper argues that current disclosure is insufficient for governance and proposes a three-tier standardized reporting scheme.","tokens_in":52928,"tokens_out":3179,"duration_ms":33848,"significance":"If the findings hold, this paper provides the first large-scale empirical baseline showing that most large companies do not publicly acknowledge any AI risks, and that the risks they do acknowledge are operational (e.g., discrimination) rather than the systemic risks experts consider most severe (e.g., power centralization, competitive dynamics). This directly challenges a key assumption behind voluntary governance instruments—that companies recognize the risks they create. The paper is methodologically careful: it validates LLM-assisted coding against human coding at the passage level, reports inter-rater statistics, provides confidence intervals for the headline estimate, and makes all data and code publicly available. The comparison with a structured expert Delphi study is a novel design that grounds the 'expectation' side of the disclosure gap. The practical proposal for a calibrated three-tier disclosure framework is specific and actionable, and the paper positions its contribution as a measurement step analogous to early climate disclosure research, which proved influential.","major_comments":[{"comment":"The company-level aggregation step (determining whether each company acknowledged each risk) is performed by a single human reviewer with no reported reliability check. The passage-level LLM validation (κ = 0.64–0.88) shows that the raw material is coded consistently, but the final, load-bearing decision—whether a company as a whole has acknowledged a risk—is not validated. Because the headline estimate (62%) is a company-level count, any idiosyncrasy in the reviewer's interpretation (e.g., overly strict criteria for what counts as an acknowledgement, or difficulty integrating conflicting evidence across documents) could systematically affect the result. The authors should provide inter-rater reliability for a random sample of companies (e.g., a second reviewer independently codes 20–30 companies) and report the agreement, or at minimum describe the decision rules that the single reviewer followed when aggregating passage-level codes. Without this, the claim that the estimate is reliable rests entirely on the passage-level validity.","section":"Methods (§Data classification, final paragraph; §Risk validation)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper defines risk acknowledgement at the document-passage level as a statement in which the company 'recognizes the existence or relevance' of a risk. However, the rules for aggregating multiple passages across a company's entire document set into a binary 'company acknowledged risk X' are not detailed. For example, if one document explicitly names a risk but another document from the same company contradicts or ignores it, is the company coded as acknowledging? What if the risk is mentioned only in a boilerplate risk-factor listing versus in a dedicated AI governance document? The human reviewer's process is described as 'considered all available evidence,' but the criteria for a positive decision are not formalized. Providing a small set of examples of borderline cases and how they were resolved would strengthen the replicability and trust in the company-level variable.","section":"Methods (§Data classification; §Determination of company risk acknowledgements)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The linear probability model estimating that 'every one point of increase in vulnerability was associated with a 7.4 percentage point increase in the probability that a company acknowledged a risk' does not report standard errors that account for clustering of risks within companies (each of the 100 companies contributes up to 24 observations, one per risk). Given that company-level acknowledgement decisions are likely correlated across risks, the current confidence intervals (4.0, 10.8) may be too narrow. The authors should re-estimate with company-level clustered standard errors or a multilevel model and report whether the result remains significant. The same concern applies to the statement about the Spearman correlation between severity and acknowledgement (p. 19).","section":"Results (§AI risks acknowledged by sector; linear probability model, p. 20)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"For the top-1000 estimates, the 95% confidence intervals for several low-prevalence risks (e.g., AI misalignment: 0.2%, 95% CI [0.2, 2.9]) are derived from a design-based weighted estimate. The interval for multi-agent risks (0.1% [0.1, 2.8]) appears suspiciously wide for a point estimate near zero, and the lower bound equals the point estimate, suggesting possible rounding or a computational artifact. The authors should clarify the method used to compute these intervals (e.g., linearized variance, bootstrap, or Wilson score) and check that the intervals for rare events are appropriate. If they used a normal approximation, a logit or exact method might be more accurate for proportions near zero.","section":"Results (§AI risks acknowledged by the largest companies; Table 1)","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The y-axis labels in Figure 1 list risk names; some are quite long and may be hard to read. Consider using abbreviated names and providing a full-name footnote. In Table 1, the 'Description' column duplicates the definition from the MIT AI Risk Repository; a short reference or a footnote would suffice, saving space.","section":"Figure 1 and Table 1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The second bullet says 'Almost no company acknowledged the risks experts judged most severe, including power centralization, competitive dynamics, and dangerous capabilities.' The text later reports that 0% of top-100 companies acknowledged power centralization and competitive dynamics, while 4% acknowledged dangerous capabilities. 'Almost no company' is acceptable but should be consistent with the precise numbers reported in the main results.","section":"p. 4, Highlights","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The axis labels '<– Mean responsibility' and 'Percent acknowledging –>' are ambiguous because the arrows point in opposite directions. It would be clearer to label the bars directly as 'Mean responsibility' and 'Percentage acknowledging' with a note that bars are sorted by mean responsibility. Also, the figure caption should explain why the blue and red bars share the same x-axis but have different scales (they appear to share the 0% line).","section":"p. 23, Figure 7","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The sentence 'Because all general-purpose AI developers were in the Information sector' refers to a sensitivity check; the wording 'Because all general-purpose AI developers were in the Information sector' could be misinterpreted as a causal statement. Rephrase as 'As a robustness check, we tested whether the higher acknowledgement rate in Information was driven solely by general-purpose AI developers...'","section":"p. 8, last paragraph","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The footnotes repeatedly warn that estimates are based on few companies and should be treated with caution. This is appropriate, but the main text occasionally cites these estimates without the same caveats (e.g., 'Information sector companies acknowledged a mean of 7.1 risks' for top-100 is fine, but the top-1000 estimate for Information should be flagged more prominently in the main text because it is based on only 7 sampled companies from ranks 101–1000).","section":"Supplementary Materials, Table S3 and S4","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Several references to the companion Delphi study (Saeri et al., 2026) appear as 'companion Delphi study' or 'companion study' without a page reference. Since readers may want to verify specific expert judgments, providing the arXiv or DOI number at first mention would be helpful. The reference is listed in the bibliography but should also be cited in the text with the identifier, e.g., 'Saeri et al. (2026, arXiv:2606.04490)'.","section":"General","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"minor_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"The paper does what the climate disclosure literature did a decade ago: it systematically measures what the largest companies publicly acknowledge about AI risks, using a validated 24-risk taxonomy and a companion expert Delphi for comparison. That is genuinely new and useful. The sampling design is transparent (all top 100 plus 100 randomly drawn from ranks 101–1000), the LLM-assisted coding is validated against human coders at the passage level, and the data and code are public. The central finding—62% of the largest 1000 companies do not acknowledge any AI risk—is policy-relevant and well-supported by confidence intervals.\n\nThe soft spot is the one the stress-test picks up. The company-level classification is done by a single human reviewer who aggregates LLM output for each risk across all documents. No inter-rater reliability is reported at that level. Since the headline number is a company-level count, any systematic tendency in that reviewer's judgment (e.g., being conservative about what counts as acknowledgement) propagates directly into the 62% estimate. The paper validates the passage-level coding well (LLM–human agreement beats human–human), but the aggregation step is uncalibrated. This is a methodological gap, not a known error, but it is load-bearing.\n\nThe paper's own limitations section is honest: it notes that disclosure is not practice, that sampling by market cap omits unlisted frontier AI developers, and that the snapshot is from early 2026. These are acknowledged thoughtfully.\n\nFor whom is this paper? Anyone working on AI governance, corporate disclosure policy, or comparative risk regulation. It provides an empirical baseline that the field needs. I would bring it to a reading group and cite it in related work.\n\nThe paper deserves a serious referee. The main issue—company-level reliability—can be fixed with a dual-coding sample and a kappa statistic. Even as is, the finding is robust enough to be informative, but the authors should address this gap before publication.","headline":"First systematic measurement of AI risk disclosure across 200 global companies; methodologically strong but the headline estimate relies on a single human reviewer at the company level without reported reliability.","tokens_in":53428,"tokens_out":1422,"duration_ms":15964},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"Most of the world's largest companies do not publicly acknowledge any AI risks, and those that do focus on operational risks while ignoring the most severe ones.","keywords":["AI risk","corporate disclosure","risk acknowledgement","Delphi study","MIT AI Risk Repository","governance","largest companies","sector vulnerability"],"falsifier":"A secret survey or mandated internal audit of a representative sample of the largest 1000 companies could reveal whether risk management practices are as absent as the public record suggests. If most companies turn out to have internal risk registers and mitigation plans for the 'missing' risks, the paper's central inference that governance instruments are failing would be weakened. Conversely, if internal audits confirm the pattern seen in public documents—especially the absence of any engagement with power centralization, competitive dynamics, or dangerous capabilities—the paper's conclusions would be strongly supported.","tokens_in":54078,"feed_emoji":"⚠️","tokens_out":2729,"duration_ms":28168,"temperature":0.7,"pith_summary":"The paper systematically examines whether the world's largest companies publicly acknowledge risks from artificial intelligence. By coding public documents from 200 companies against a comprehensive taxonomy of 24 AI risks, the authors find that an estimated 62% of the largest 1000 companies do not acknowledge any AI risk at all. Among the largest 100 companies, 32% are silent on AI risks. The risks that are acknowledged cluster around discrimination, weapons and cyberattacks, fraud, privacy, and devaluation of human creativity, while risks that experts rate as most severe—such as power centralization, competitive dynamics, and dangerous capabilities—are almost never mentioned. Even companies in sectors judged extremely vulnerable to a risk, and companies judged most responsible for addressing a risk, are no more likely to acknowledge that risk than others. The findings challenge a key assumption behind many AI governance instruments: that companies recognize the risks they face and act on them.","feed_headline":"62% of world's largest firms ignore AI risks entirely","feed_subtitle":"A systematic review of public documents from 200 companies finds no mention of AI risk in most firms, while the few that do focus on…","key_machinery":"The study pairs a systematic content analysis of corporate public documents with a structured expert elicitation (Delphi) to compare what companies say with what experts think they should be saying. The content analysis scans up to 80 search results per company (plus annual reports, LLM-agent searches, and company feedback) and codes each passage for the presence of any of the 24 AI risks defined by the MIT AI Risk Repository (Slattery et al., 2026). A large language model performs initial classification, validated by human review against a random sample. The expert Delphi study (Saeri et al., 2026) provides median ratings of each risk's expected severity, each sector's vulnerability, and each AI ecosystem role's responsibility. The comparison is then made across three dimensions: severity, vulnerability, and responsibility, using weighted estimates for the top 1000 companies and direct counts for the top 100.","core_discovery":"The paper establishes that public acknowledgement of AI risks by the world's largest companies is both sparse and misaligned with expert priorities. Of the largest 100 companies by market capitalization, 32% did not publicly acknowledge any AI risk; for the largest 1000, the estimated share is 62% (95% CI: 54%, 70%). When companies do acknowledge risks, over half of all mentions fall into five categories: discrimination, weapons and cyberattacks, fraud and scams, loss of privacy, and devaluation of human creativity. Meanwhile, risks that a companion Delphi study of 272 experts identified as most severe over 2025–2030—power centralization, competitive dynamics, dangerous capabilities, AI misalignment, and multi-agent risks—are virtually absent from corporate disclosures. A quantitative comparison shows no correlation between expert-judged severity and company acknowledgement (Spearman's rho = –0.16). Sector vulnerability is weakly predictive: companies in more vulnerable sectors are more likely to acknowledge a risk, but even among 'extremely vulnerable' sectors, 77% of risk–company combinations go unacknowledged. Role responsibility is not predictive: companies judged most responsible for addressing a risk are no more likely to acknowledge it than others. For example, only 15% of general-purpose AI developers acknowledge AI misalignment, whereas 85% acknowledge discrimination.","pith_inferences":["The gap between what companies acknowledge and what experts consider severe may reflect a structural misalignment: corporate risk management is incentivized by shareholder value, not by public harm, so risks that externalize costs are systematically underdiscussed.","The finding that general-purpose AI developers acknowledge misalignment much less than discrimination suggests that even technically sophisticated firms avoid the hardest governance questions, possibly because acknowledging misalignment carries reputational or regulatory risk.","The methodology could be extended to other documents (e.g., patent filings, hiring postings, internal memos leaked or subpoenaed) to test whether non-public recognition matches or diverges from public acknowledgement, providing a stronger test of the paper's proxy assumption.","A natural experiment would be to track disclosure rates before and after a major AI incident or regulatory fine; an increase in acknowledgements of specific risks after the event would support the paper's implication that current silence is not due to genuine absence of risk."],"forward_implications":["If most large companies do not even acknowledge AI risks, governance instruments that rely on corporate self-recognition and voluntary disclosure are unlikely to be effective without stronger mandates.","The five risks that dominate current disclosures (discrimination, weapons & cyberattacks, fraud & scams, loss of privacy, devaluation of human creativity) are precisely those with direct operational or financial impact on the company, suggesting that firms neglect risks whose costs fall primarily on society.","A tiered reporting framework—common core risks, sector-specific risks, and role-specific risks—could align disclosure with actual exposure and responsibility, analogous to climate reporting (Scope 1/2/3 and sector-specific IFRS S2 standards).","The absence of any acknowledgement of power centralization, competitive dynamics, or dangerous capabilities by the largest companies implies that systemic and catastrophic AI risks are invisible in corporate reporting, making it impossible for outsiders to assess preparedness.","If mandatory disclosure requirements are introduced, the study provides a baseline against which future improvements can be measured, similar to the role of early climate disclosure research."],"supporting_citations":[{"why":"Supplies the 24-risk taxonomy (MIT AI Risk Repository) that the paper uses to classify corporate acknowledgements; the taxonomy is the foundation for all coding and comparison.","marker":"Slattery et al., 2026"},{"why":"Provides the expert-judged severity, sector vulnerability, and role responsibility ratings from a Delphi study of 272 AI risk experts; these ratings are the benchmark against which corporate acknowledgement is compared.","marker":"Saeri et al., 2026"},{"why":"Offers prior evidence that US companies discussing AI risks in SEC 10-K filings focus on legal and competitive concerns, setting the context for the broader, multi-document approach of this paper.","marker":"Uberti-Bona Marin et al., 2025"},{"why":"Documents the landscape of voluntary AI ethics principles, establishing that disclosure on AI risks is currently inconsistent and voluntary, which the paper uses to motivate the need for systematic measurement.","marker":"Jobin et al., 2019"},{"why":"Provides the empirical precedent that mandatory climate disclosure in the UK led to measurable emissions cuts; this is the model the paper invokes for how AI risk disclosure could eventually change corporate behavior.","marker":"Downar et al., 2021"},{"why":"Evaluates the first round of the Hiroshima AI Process framework, finding reports high-level and difficult to compare; this critique supports the paper's call for standardized, calibrated reporting.","marker":"Bogen et al., 2026"}],"fun_headline_variants":["62% of top 1000 firms have zero public AI risk mentions","Corporate silence: most large companies ignore AI risks entirely","Experts' top AI risks absent from nearly all corporate filings","Five AI risks dominate corporate talk; experts' top risks ignored","Top 100 companies: 32% acknowledge zero AI risks in public docs"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The paper assumes that a company's public acknowledgement of AI risks in its documents is a meaningful proxy for whether the company recognizes and manages those risks; if companies manage risks internally without disclosing them, or disclose only symbolically, then the observed silence may overstate the governance failure.","fun_headline_variants_meta":{"raw":{"variants":["62% of top 1000 firms have zero public AI risk mentions","Corporate silence: most large companies ignore AI risks entirely","Experts' top AI risks absent from nearly all corporate filings","Five AI risks dominate corporate talk; experts' top risks ignored","Top 100 companies: 32% acknowledge zero AI risks in public docs"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000107,"raw_usage":{"cost":0.0001069572,"is_byok":false,"cost_details":{"upstream_inference_cost":0.0001069572,"upstream_inference_prompt_cost":5.69772e-05,"upstream_inference_completions_cost":4.998e-05},"total_tokens":1640,"prompt_tokens":1140,"completion_tokens":500,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":0,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":421}},"tokens_in":1140,"tokens_out":500,"cost_source":"openrouter_usage","duration_ms":6602,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":421,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-04T16:20:29.893742+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"A secret survey or mandated internal audit of a representative sample of the largest 1000 companies could reveal whether risk management practices are as absent as the public record suggests. If most companies turn out to have internal risk registers and mitigation plans for the 'missing' risks, the paper's central inference that governance instruments are failing would be weakened. Conversely, if internal audits confirm the pattern seen in public documents—especially the absence of any engagement with power centralization, competitive dynamics, or dangerous capabilities—the paper's conclusions would be strongly supported.","supporting_citations":[],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}