# ALMOST RIGHT — Publication Claims Audit
## Audit date: August 31, 2026
## Audited manuscript: ALMOST-RIGHT-publication-claims-audited-manuscript.md

> **Restructuring note — August 31, 2026 (evening).** After this audit was completed, Part V was rewritten for voice and consolidated from eleven chapters to six; the book is now 24 chapters. Old chapters 16/23 became Chapter 16; 17/24 became Chapter 17; 21/22 became Chapter 18; 19/25 became Chapter 19; 18/20 became Chapter 20; 26 became Chapter 21; and old 27/28/29 became Chapters 22/23/24. No audited factual claim, number, or qualifier was changed in the rewrite. Chapter references below have been updated to the new numbering.

## Publication status

**Claims-audited candidate.** The manuscript's load-bearing factual claims, quantitative claims, policy/legal claims, medical evidence, labor-market evidence, security examples, and newly added research sections have been checked and the known publication-blocking errors found in this pass have been corrected in the manuscript.

This does **not** mean every interpretive sentence is a provable fact. The book contains argument, synthesis, predictions, and the author's firsthand experiences. The edit now distinguishes those categories more carefully from externally verifiable claims. Vendor benchmarks are identified as vendor benchmarks; observational studies are described as observational; developing 2026 policy and labor-market claims are dated to the relevant snapshot.

The remaining production gate is a conventional copyedit/proofread and final source-formatting pass, not another substantive fact rewrite unless new information appears before submission.

## Audit method

The audit prioritized primary sources and authoritative records wherever available: official research papers, government and regulatory documents, institutional survey dashboards, court/legal databases, company statements for company-specific facts, and security researchers' original disclosures. Secondary reporting was retained only where it supplied facts not available in a primary record, and high-risk claims were narrowed when a primary source did not justify the stronger wording.

Author anecdotes were treated as author-supplied testimony rather than independently verified factual reporting.

## Corrections applied

1. Ch4: corrected Stack Overflow trust to 33%, distrust to 46%, and clarified 84% means using or planning to use.
2. Ch4: removed unsupported characterization of all respondents as professional programmers.
3. Ch4: updated Charlotin database from July snapshot to Aug. 28, 2026 and corrected what the database counts.
4. Ch4: changed Charlotin 'filings' to accurate database characterization.
5. Ch23: removed false claim that Charlotin database contains 1,500 lawyers.
6. Ch4 summary: replaced reductive 'scale alone ended AI winters' claim with historically defensible formulation.
7. Ch6: removed unsupported causal link between Anthropic's Public First Action donation and pro-Bores election spending.
8. Ch6: removed unverified 'largest technology-sector political spending ever' superlative.
9. Ch3: removed stale/brittle $725B 2026 capex aggregate and company-by-company guidance.
10. Medicine: corrected cancer/causal wording around Budzyń et al. observational colonoscopy study.
11. Medicine: corrected cancer/causal wording around Budzyń et al. observational colonoscopy study.
12. Medicine: corrected cancer/causal wording around Budzyń et al. observational colonoscopy study.
13. Medicine: corrected cancer/causal wording around Budzyń et al. observational colonoscopy study.
14. Ch22: corrected Tea breach description (13,000 verification images incl. IDs/selfies; ~72,000 images total).
15. Ch23: corrected RLS/client-key explanation and removed false linkage between Lovable RLS issue and Tea breach.
16. Ch23: separated security review from usability testing and narrowed author anecdote claim.
17. Ch23: removed claim that each of five controls would have prevented a specific documented breach.
18. Sources: corrected Escape vendor benchmark wording to >5,600 apps / >2,000 vulnerabilities / 400+ secrets / 175 PII instances.
19. Ch18: removed future-dated Sept. 2026 Guardian example and replaced it with Aug. 28 Charlotin database evidence.
20. Ch22: removed unsupported 'safest transportation ever' and 'automation nearly ate it' causal superlatives.
21. Ch16: qualified Veracode 45% result as a vendor benchmark rather than universal AI-code failure rate.
22. Ch23: replaced vague 'Stanford' confidence claim with properly scoped descriptions of METR and security-study evidence.
23. Ch5/6: removed unsupported 'one billion weekly' wording; retained confirmed hundreds-of-millions scale.
24. Ch5: narrowed adoption-speed claim to the researchers' PC/internet comparisons.
25. Ch24: narrowed cross-study synthesis to avoid claiming one causal mechanism across heterogeneous evidence.
26. Ch4: narrowed the hallucination-paper theorem to its formal assumptions rather than universal '2x' rule.

## Load-bearing claims now supported

### Central developer-survey evidence
The 2025 Stack Overflow survey reports that 84% of respondents were using or planning to use AI tools in their development process; 66% cited “AI solutions that are almost right, but not quite” as the largest frustration; 45% cited more time-consuming debugging; and 33% trusted AI accuracy versus 46% who distrusted it. The manuscript now uses those primary-dashboard figures rather than conflating “somewhat trust” with total trust.

### Legal hallucinations
Damien Charlotin's database said it had identified 1,981 cases as of August 28, 2026. The database tracks legal decisions where courts or tribunals addressed alleged or established hallucinated AI material; it is not a count of all filings, all fake citations, or all lawyers. The manuscript now describes it that way.

### Labor market
The August 12, 2026 revision of Stanford Digital Economy Lab's *Canaries in the Coal Mine?* finds no widespread economy-wide AI displacement, but reports that workers ages 22–25 in highly AI-exposed occupations stood 19% below the employment path they would have followed had they kept pace with less-exposed same-age peers; experienced workers showed no comparable gap. The paper also reports that the adjustment is primarily through reduced hiring. The manuscript now uses the revised wording.

### Medical deskilling evidence
Budzyń et al. is a multicentre **observational** study. Standard non-AI adenoma detection rate fell from 28.4% before AI introduction to 22.4% after exposure in the study periods. The manuscript no longer calls adenomas “cancers,” no longer turns the association into proof of individual skill loss, and no longer attributes causation the study design cannot establish.

### Security evidence
Veracode's 45% finding is now explicitly described as a vendor benchmark of tested generation tasks, not a universal percentage of AI-generated code. Escape's research is stated as more than 5,600 publicly available apps, more than 2,000 vulnerabilities, 400+ exposed secrets, and 175 PII exposures. The Lovable/CVE row-level-security issue is no longer incorrectly presented as the cause of Tea's breach. Tea is described separately as an exposure of about 72,000 images, including roughly 13,000 verification images containing photo IDs and selfies, plus a later separate exposure of private messages.

### U.S. federal AI policy
The manuscript now reflects the July 23, 2026 introduction of the FRONTIER Act, developed from the broader Great American AI Act framework. It no longer says the framework had never been formally introduced in any form, and it no longer claims literally that no federal law of any kind touches AI. The narrower claim is that the United States lacks a comprehensive federal AI law imposing a general predeployment verification regime across consumer AI systems.

### European Union
The manuscript now distinguishes the EU AI Act's general application from the Digital Omnibus delay affecting specified high-risk obligations: December 2, 2027 for Article 6(2)/Annex III systems and August 2, 2028 for Article 6(1)/Annex I systems. It no longer implies the entire Act was postponed.

### Political spending
The manuscript no longer treats Anthropic's $20 million February 2026 donation to Public First Action as money used to elect a particular federal candidate. Anthropic explicitly said the donation was restricted to public education/policy work and could not be used to influence candidate elections. An unsupported “largest technology-sector campaign ever” superlative was also removed.

### Aviation analogy
FAA guidance supports the claim that airlines and pilots should maintain manual-flight proficiency in highly automated operations. The manuscript retains aviation as an analogy and institutional model but removes the unsourced claim that commercial aviation is literally “the safest form of transportation ever devised” and the causal flourish that it became safe only after automation “nearly ate it.”

## Chapter-by-chapter audit status

- **Chapter 1, The Bet at Dartmouth:** Verified with historical qualification. Dartmouth and early-AI history retained; broad teleology treated as interpretation.
- **Chapter 2, Two Winters:** Verified with qualification. AI-winter history retained; avoid monocausal explanations.
- **Chapter 3, The Real Reason:** Verified with current-figure cleanup. OpenAI governance/equity structure supported; capex section made less brittle.
- **Chapter 4, Almost Right, By Design:** Corrected and verified. Stack Overflow, Charlotin database, hallucination paper, and legal examples tightened.
- **Chapter 5, Faster Than the Internet:** Verified with adoption-denominator corrections. Weekly vs monthly user figures kept distinct.
- **Chapter 6, Nobody's Guarding the Door:** Corrected for current law/politics. FRONTIER Act, EU delay scope, and Anthropic political-spending language updated.
- **Chapter 7, The Sellers:** Verified with source-attribution caveats for executive/company claims.
- **Chapter 8, The Two Countries:** Verified with geographic/adoption caveats; source-derived inequalities remain observational/descriptive.
- **Chapter 9, What Actually Works:** Verified. Peer-reviewed customer-support and productivity evidence retained with population/task limits.
- **Chapter 10, Vibe Coding:** Verified with benchmark/company-source caveats. Vibe-coding claims remain time- and platform-specific.
- **Chapter 11, Nobody Hacked Them:** Corrected. Security examples now separate Tea, Lovable/CVE, Base44, Replit, Escape, and their distinct failure modes.
- **Chapter 12, The Middlemen:** Verified as synthesis; middleman/verification argument is interpretive rather than a measured universal law.
- **Chapter 13, Cognitive Debt:** Verified with study-design limits. Cognitive-debt section distinguishes experimental evidence from extrapolation.
- **Chapter 14, The Doctors Got Worse:** Corrected. Colonoscopy evidence now consistently says adenoma detection and observational association.
- **Chapter 15, The Canaries:** Updated to Aug. 12, 2026 Stanford revision; no economy-wide-displacement claim preserved.
- **Chapter 16, When Nobody Knows What Right Looks Like:** Verified with security-benchmark qualification and expertise-boundary argument labeled as synthesis.
- **Chapter 17, The Apprenticeship Problem:** Verified. Apprenticeship mechanism is an argument supported by labor and learning evidence, not a settled causal estimate.
- **Chapter 20, The Verification Economy:** Verified as economic thesis. NIST TEVV evidence supports formalization of verification work; market forecasts remain forecasts.
- **Chapter 19, The Liability Gap:** Verified with date-sensitive legal/insurance reporting; responsibility analysis is normative/synthetic.
- **Chapter 20, The New Professional Class:** Verified as forward-looking argument; security evidence scoped to its underlying benchmarks.
- **Chapter 18, The Factory for Almost Right:** Verified as systems analysis; error-rate examples are illustrative mathematics, not empirical forecasts.
- **Chapter 18, The Evidence Problem:** Corrected. Future-dated September example removed; current legal-database evidence substituted.
- **Chapter 16, The Jagged Frontier:** Verified. BCG jagged-frontier findings retained with task/frontier limits.
- **Chapter 17, The School Before the Workplace:** Verified with study limits. Education evidence does not support blanket anti-AI conclusions; manuscript retains both positive and negative evidence.
- **Chapter 19, Who Verifies the Verifier?:** Verified as governance/safety synthesis; independence, sampling, and red-team recommendations are prescriptive.
- **Chapter 21, The Case for Optimism:** Verified as synthesis/counter-case; productivity evidence is deliberately included to prevent one-sided argument.
- **Chapter 22, What the Pilots Did:** Corrected. FAA analogy retained but aviation superlatives and Tea phrasing narrowed.
- **Chapter 23, Become the Verifier:** Corrected. Practical controls now distinguish database authorization, security review, and real-user testing.
- **Chapter 24, Start Now:** Corrected as final evidence summary; heterogeneous studies no longer claimed to prove one universal causal mechanism.

## Claims that remain inherently non-verifiable or source-dependent

The author's descriptions of his own businesses, advertising spend, software-building experience, conversations, reactions, and firsthand observations are author testimony. A publisher may ask for supporting records, screenshots, receipts, contemporaneous notes, or other documentation during legal review. Those claims should be preserved only to the extent the author can stand behind them personally.

Company press releases and vendor security studies can establish what those organizations reported or measured, but they are not neutral population estimates. The manuscript now attributes them accordingly.

Fast-moving 2026 claims—especially legislation, political spending, model-user counts, and labor-market data—should receive a **date-of-submission refresh** if the manuscript is sent materially after August 31, 2026.

## Editorial conclusion

The manuscript is now in substantially better factual shape than the pre-audit version. The known flat errors and the most important overstatements identified in this publication pass have been corrected directly in the audited manuscript. The argument survives the corrections: the strongest version of *Almost Right* does not require inflated statistics or universal claims. Its core case is stronger when the evidence is kept inside the boundaries of what each source actually measured.

**Recommended next gate:** final copyedit/proofread, source-note normalization, and submission formatting. If submission occurs weeks or months after August 31, refresh the time-sensitive claims before sending.


## Supplemental corrections after automated residual scan

- Ch6 federal-policy section updated to reflect July 23 FRONTIER Act introduction and remove categorical 'no federal law' wording.
- Ch10 Stack Overflow trust figure corrected from 29% to primary-dashboard 33% trust / 46% distrust.
- Ch11 source note updated to Escape's directly supported >5,600 / >2,000 wording.
- Ch24 final evidence stack rewritten to use scoped benchmark, observational, and revised labor-market language.