R0031/2026-03-27¶
Verification of 14 claims from a plural-voice article covering AI trust surveys, academic integrity, journal policies, film adaptations, and AI attribution frameworks. The majority of claims (10 of 14) are confirmed or confirmed with nuance; two require corrections to attribution or naming; one involves a misattribution; and one is a loose characterization.
Claims¶
C001 — KPMG global AI trust statistics — Very likely (80-95%)
Claim: Public trust in AI sits at 46% globally, with 39% in advanced economies versus 57% in emerging ones (attributed to KPMG/University of Melbourne, 48,340 respondents across 47 countries).
Verdict: Substantially confirmed. The 46% global trust, 47 countries, and 39%/57% split are verified. The respondent count is "over 48,000" in press materials rather than the precise 48,340.
| Hypothesis | Status | Probability |
|---|---|---|
| H1: Exactly accurate | Inconclusive | — |
| H2: Substantially correct with minor precision differences | Supported | 80-95% |
| H3: Materially wrong | Eliminated | — |
Confidence: High · Sources: 1 · Searches: 1
C002 — Ipsos 66% use AI despite distrust — Unlikely (20-45%)
Claim: 66% of people use AI despite the majority not trusting it (attributed to Ipsos, 31-country survey).
Verdict: Misattributed. The 66% usage figure comes from KPMG/Melbourne (47 countries) or Google/Ipsos 2026 (21 countries). The Ipsos 2023 31-country survey's 66% refers to expected life impact, not usage.
| Hypothesis | Status | Probability |
|---|---|---|
| H1: Accurate (Ipsos, 31 countries, 66% use) | Eliminated | — |
| H2: Real phenomenon, wrong attribution | Supported | 20-45% |
| H3: Materially wrong | Eliminated | — |
Confidence: High · Sources: 3 · Searches: 1
C003 — 57% workers hide AI use — Almost certain (95-99%)
Claim: 57% of workers hide their AI use at work (attributed to KPMG/University of Melbourne study of over 48,000 workers across 47 countries).
Verdict: Confirmed. KPMG press release states: "Over half (57%) of employees say they hide their use of AI and present AI-generated work as their own."
| Hypothesis | Status | Probability |
|---|---|---|
| H1: Accurate as stated | Supported | 95-99% |
| H2: Partially correct | Eliminated | — |
| H3: Materially wrong | Eliminated | — |
Confidence: High · Sources: 1 · Searches: 1
C004 — 22% students submit AI content — Likely (55-80%)
Claim: 22% of students admit to submitting AI-generated content as their own work.
Verdict: The 22% figure exists (BestColleges, March 2023, N=1,000) but measures broader AI use on assignments, not exclusively submitting fully AI-generated content.
| Hypothesis | Status | Probability |
|---|---|---|
| H1: Exactly accurate | Inconclusive | — |
| H2: Figure exists but broader than claimed | Supported | 55-80% |
| H3: Materially wrong | Eliminated | — |
Confidence: Medium · Sources: 1 · Searches: 1
C005 — UK 7,000 AI misconduct cases — Very likely (80-95%)
Claim: UK universities reported over 7,000 formal academic misconduct cases involving AI in a single year (2023-2024).
Verdict: Confirmed with nuance. The Guardian FOI investigation (131 of 155 universities) found ~7,000 AI cheating cases. "Formal" may slightly overstate — includes suspected and confirmed cases.
| Hypothesis | Status | Probability |
|---|---|---|
| H1: Exactly accurate | Inconclusive | — |
| H2: ~7,000 confirmed but includes suspected cases | Supported | 80-95% |
| H3: Materially wrong | Eliminated | — |
Confidence: Medium · Sources: 1 · Searches: 1
C006 — Journals prohibit AI as author — Almost certain (95-99%)
Claim: Every major journal and conference — Nature, Science, ACM, IEEE, NeurIPS, and all five major academic publishers — has issued a formal policy prohibiting AI as an author.
Verdict: Confirmed. All named entities have formal AI authorship prohibition policies.
| Hypothesis | Status | Probability |
|---|---|---|
| H1: Accurate as stated | Supported | 95-99% |
| H2: Most but not all | Eliminated | — |
| H3: Materially wrong | Eliminated | — |
Confidence: High · Sources: 5 · Searches: 1
C007 — NeurIPS methodology-only disclosure — Almost certain (95-99%)
Claim: NeurIPS requires AI disclosure only when AI is part of the methodology.
Verdict: Confirmed. NeurIPS 2025 policy requires disclosure when LLM use is "an important, original, or non-standard component" of the method. Grammar/editing tools exempt.
| Hypothesis | Status | Probability |
|---|---|---|
| H1: Accurate as stated | Supported | 95-99% |
| H2: Broader requirements | Eliminated | — |
| H3: No requirement | Eliminated | — |
Confidence: High · Sources: 1 · Searches: 1
C008 — Science requires full prompts — Almost certain (95-99%)
Claim: Science requires full prompts to be included in the methods section.
Verdict: Confirmed. Science editorial policy: "The full prompt used in the production of the work, as well as the AI tool and its version, should be disclosed" in the methods section.
| Hypothesis | Status | Probability |
|---|---|---|
| H1: Accurate as stated | Supported | 95-99% |
| H2: Partial requirement | Eliminated | — |
| H3: No such requirement | Eliminated | — |
Confidence: High · Sources: 1 · Searches: 1
C009 — ACM/IEEE acknowledgment requirements — Almost certain (95-99%)
Claim: ACM and IEEE require acknowledgment of AI use with varying specificity.
Verdict: Confirmed. ACM requires Acknowledgments disclosure (footnote for small amounts). IEEE requires more specific disclosure: AI system identification, specific sections, and usage description.
| Hypothesis | Status | Probability |
|---|---|---|
| H1: Accurate as stated | Supported | 95-99% |
| H2: Only one requires it | Eliminated | — |
| H3: Neither requires it | Eliminated | — |
Confidence: High · Sources: 2 · Searches: 1
C010 — Kurosawa Shakespeare adaptations — Likely (55-80%)
Claim: Kurosawa's Throne of Blood (1957) is based on Macbeth, The Bad Sleep Well (1960) is based on Hamlet, and Ran (1985) is based on King Lear.
Verdict: Substantially correct. Throne of Blood/Macbeth and Ran/King Lear are unambiguous. The Bad Sleep Well draws on Hamlet but scholars describe it as a loose adaptation, not a direct one.
| Hypothesis | Status | Probability |
|---|---|---|
| H1: All three accurately attributed | Inconclusive | — |
| H2: Two direct, one loose adaptation | Supported | 55-80% |
| H3: One or more wrong | Eliminated | — |
Confidence: High · Sources: 3 · Searches: 1
C011 — IBM AI Attribution Toolkit — Almost certain (95-99%)
Claim: IBM published an AI Attribution Toolkit in 2025 that captures contribution type, amount, and review process, and describes itself as "a first pass" at a voluntary standard.
Verdict: Confirmed. Published May 13, 2025. Captures contribution balance, AI role, and review process. IBM describes it as "a first pass at formulating what a voluntary reporting standard might look like."
| Hypothesis | Status | Probability |
|---|---|---|
| H1: Accurate as stated | Supported | 95-99% |
| H2: Features differ | Eliminated | — |
| H3: Does not exist | Eliminated | — |
Confidence: High · Sources: 1 · Searches: 1
C012 — AIA icon system — Unlikely (20-45%)
Claim: The AIA icon system was proposed by Avery, Abril, and del Riego with graduated visual indicators for AI involvement levels: Generated, Edited, Suggested (published in Journal of Technology, Innovation & Practice, 2024).
Verdict: Paper exists but journal name is wrong. Published in Northwestern Journal of Technology and Intellectual Property, not "Journal of Technology, Innovation & Practice." Icon labels "Generated, Edited, Suggested" unverified.
| Hypothesis | Status | Probability |
|---|---|---|
| H1: Accurate including journal | Eliminated | — |
| H2: Paper exists, journal name wrong | Supported | 20-45% |
| H3: No such paper | Eliminated | — |
Confidence: Medium · Sources: 1 · Searches: 1
C013 — CHI 2025 AI credit study — Almost certain (95-99%)
Claim: Research presented at CHI 2025 (He et al.) found that AI receives less credit than humans for equivalent work.
Verdict: Confirmed. He, Houde, and Weisz (IBM Research) presented at CHI 2025. N=155. "Across nearly all natures of contribution, participants assigned AI partners less authorship credit than human partners for equivalent contributions."
| Hypothesis | Status | Probability |
|---|---|---|
| H1: Accurate as stated | Supported | 95-99% |
| H2: Findings differ | Eliminated | — |
| H3: No such paper | Eliminated | — |
Confidence: High · Sources: 1 · Searches: 1
C014 — CRediT taxonomy no AI provision — Almost certain (95-99%)
Claim: The CRediT taxonomy (NISO Z39.104-2022) covers 14 types of human contribution but has no provision for AI.
Verdict: Confirmed. CRediT defines 14 roles, all for human contributors. No mention of AI anywhere in the taxonomy. The AID Framework was created to fill this gap.
| Hypothesis | Status | Probability |
|---|---|---|
| H1: Accurate as stated | Supported | 95-99% |
| H2: CRediT addresses AI | Eliminated | — |
| H3: Not 14 roles | Eliminated | — |
Confidence: High · Sources: 2 · Searches: 1
Collection Analysis¶
Cross-Cutting Patterns¶
| Pattern | Claims Affected | Significance |
|---|---|---|
| KPMG/Melbourne study is a nexus source | C001, C002, C003 | Three claims draw from the same 2025 study — source dependency risk |
| Journal AI policies are consistent | C006, C007, C008, C009 | All major publishers/conferences align on prohibiting AI authorship while varying on disclosure requirements |
| Attribution gap is well-documented | C011, C012, C013, C014 | Multiple independent efforts to address AI attribution, confirming a real gap |
| Trust-usage paradox is real but sources are confused | C001, C002, C003 | The phenomenon exists across multiple surveys but claim C002 conflates sources |
Collection Statistics¶
| Metric | Value |
|---|---|
| Claims investigated | 14 |
| Fully confirmed (Almost certain) | 7 (C003, C006, C007, C008, C009, C011, C013, C014) |
| Confirmed with nuance (Very likely) | 2 (C001, C005) |
| Confirmed with caveats (Likely) | 2 (C004, C010) |
| Misattributed or wrong details (Unlikely) | 2 (C002, C012) |
| Materially wrong | 0 |
Source Independence Assessment¶
The evidence base shows moderate independence. Claims C001-C003 all depend on the KPMG/Melbourne 2025 study — a single-source risk. The journal policy claims (C006-C009) each reference independent primary sources (Nature, Science, ACM, IEEE, NeurIPS). The attribution framework claims (C011-C014) reference independent efforts (IBM, Avery et al., He et al., NISO). The Kurosawa claim (C010) draws on well-established film scholarship with multiple independent sources.
The most significant independence concern is the KPMG/Melbourne concentration: if that study were found to have methodological issues, three claims would be affected.
Collection Gaps¶
| Gap | Impact | Mitigation |
|---|---|---|
| KPMG full report PDF not accessible | Cannot verify exact 48,340 respondent count | Press materials confirm "over 48,000" |
| Full text of AIA paper not accessible | Cannot verify specific icon labels | Abstract and metadata confirm paper exists |
| Ipsos survey data behind paywalls | Cannot fully trace 66% figure across surveys | Multiple waves checked via public summaries |
| Self-reported survey data across multiple claims | Social desirability bias may affect trust/usage figures | Acknowledged in source scorecards |
Collection Self-Audit¶
| Domain | Rating | Notes |
|---|---|---|
| Eligibility criteria | Low risk | All claims had clear verification criteria defined before searching |
| Search comprehensiveness | Some concerns | 14 claims in a single run limits depth per claim; primary sources found for all |
| Evaluation consistency | Low risk | Same framework applied across all 14 claims |
| Synthesis fairness | Low risk | Corrections and caveats surfaced for claims that needed them (C002, C004, C005, C010, C012) |
Resources¶
Summary¶
| Metric | Value |
|---|---|
| Claims investigated | 14 |
| Files produced | ~280 |
| Sources scored | 22 |
| Evidence extracts | 22 |
| Results dispositioned | 28 selected + 112 rejected = 140 total |
| Duration (wall clock) | 25m 52s |
| Tool uses (total) | 99 |
Tool Breakdown¶
| Tool | Uses | Purpose |
|---|---|---|
| WebSearch | 14 | Search queries across all claims |
| WebFetch | 10 | Page content retrieval for primary sources |
| Write | ~30 | File creation (core files) |
| Read | 2 | Methodology and output format reading |
| Edit | 0 | No file modifications |
| Bash | ~25 | Directory creation, bulk file writing |
Token Distribution¶
| Category | Tokens |
|---|---|
| Input (context) | ~150,000 |
| Output (generation) | ~80,000 |
| Total | ~230,000 |