Corrections and updates

Corrections

Page 16, Chapter 1, the practitioner's ledger. The text says "the same three lines fit a regulatory filing". The table above it has four lines, and all four transfer.

Pages 19 and 37, Mata v. Avianca. The book says the lawyer read the brief. The court's findings are narrower. The lawyer who signed it reviewed it for style and flow, and read none of the cases it cited. That makes the point more strongly, not less.

Pages 21 and 33, Chapter 2, the Gillespie survey. The book says that 66% of those surveyed rely on AI output without checking it, and 56% said AI had led them into mistakes. Both figures come from the employees in the survey who use AI at work, not from all 48,000 respondents.

Pages 27, 35, 55 and 59, the reward-hacking argument. Page 55, and the source note on page 59, say the work of Skalse and colleagues "suggests you essentially cannot build a reward measure a determined optimiser cannot game". Pages 27 and 35 call the underlying claim sound. The paper is narrower. Across the set of all stochastic policies, two reward functions can only be unhackable if one of them is constant. Restricted to deterministic policies, or to finite sets of stochastic policies, non-trivial unhackable pairs do exist. The book's advice does not depend on the wider reading: check with outside sources and a second, unrelated model, and never ask a tool to grade its own work.

Pages 29 and 34, Chapter 2, Dahl and colleagues. The book compares these results with the legal research tools, saying general-purpose models "asked the same questions" were wrong between 58% and 88% of the time. Page 34 says "the same kind of legal questions". The questions were not the same. Dahl and colleagues asked specific, verifiable questions about randomly chosen federal court cases, and measured how often the answers were hallucinated. The two studies cannot be compared directly, and neither figure measures how often legal advice is wrong.

Page 33, Chapter 2, what Part II contains. The book says that in Part II "each of the seven types gets a worked example, set out in full". It does not work through each type. The example it names as "the one worked example to carry forward" (page 161) is the Council of Voices, on page 160, and the task chapters reuse its shape.

Pages 37 and 38, Chapter 3, Goh and colleagues. The book reads this study as a test of a person reviewing AI output ("The seat was occupied"). The study tested doctors using an AI model as a tool alongside their usual resources. It shows that the pairing made no significant difference. It is not a direct test of review.

Pages 41 and 47, Chapter 3, Dell'Acqua (2022). The book says "as the tool improved, the watching decayed", and the source note on page 47 says the same in brief. The study had four conditions. The group with the stronger algorithm did worse than the group with the weaker one, but the near-perfect algorithm beat every other group. The effect sits in a middle band, not in a straight line.

Pages 48 and 59, Chapter 4, Sullivan & Cromwell. The book says the filing leaned on cited authorities an AI had generated, which opposing counsel "caught at once". The firm's letter describes inaccurate citations generated by AI. Its schedule of corrections records mistaken page references, volume numbers, quotations and rule references, and on its face names no case that does not exist. The total number is not settled on the public record.

Pages 50 and 59, Chapter 4, Kamoi and colleagues. The book says self-correction shows up "only" when something external is brought in. The survey found no reliable self-correction from prompting alone, except in tasks unusually suited to it. It found two things that work: reliable outside feedback, and large-scale fine-tuning.

Page 85, Chapter 7, the book's own ledger. The book describes "a fabricated reference its checks caught late". As page 27 says, the checks flagged it the first time it appeared. What came late was closing the finding. Whether the reference was invented outright, or was a garbled citation of a real paper, cannot now be established.

Page 93, Chapter 8, Matt Beane. The book gives the "ten to twenty times less" hands-on practice "by his account", but places it in "a 2019 study". The figure is Beane's own estimate, from his public writing about his fieldwork. The 2019 paper is the source for his finding on shadow learning.

Page 94, Chapter 8, Berkeley School of Law. The book says the school "flipped its default". The school's policy, adopted in May 2026 and in force from Summer 2026, is confirmed. What the earlier policy said is not publicly available, so a reversal cannot be shown.

Page 102, Chapter 9, the Replit incident. The book says an agent "at the company Replit" deleted a production database. It was a customer's project running on Replit's platform, in July 2025. The data was later recovered.

Page 149, Part II, Tools, models, and working with the system. The book says a connector "cannot see more than you can, and it often sees less". That is not reliable. Many connectors run under a service account with wider access than yours, so a connector can reach material you are not cleared to see. Add a question to the tool check: under what identity does this connector run?

Updates and clarifications

Pages 11, 18, 73, 76 and 78, Dell'Acqua and colleagues. The book says consultants working past the model's frontier were "19% less likely" to reach a correct answer. That follows the published abstract. The paper's own figures show a fall of about 19 percentage points, against a control group that was right about 84.5% of the time. As a relative fall, that is about 22 to 23%. The source notes on pages 17 and 78 date the paper 2025. The journal issue they cite, Organization Science 37(2), is dated 2026, so the year should be 2026.

Pages 51, 52 and 58, Chapter 4, Panickssery and colleagues. The book says AI judges prefer their own output to work that human raters scored equal. That is what the study reports. A paper by Roytburg and colleagues, first posted in January 2026 and presented at ICML 2026, re-ran four studies of this kind. Its main text finds that the judge's own uncertainty "accounts for an average of 89.6% of measured self-preference", "substantially reducing but not eliminating evidence for bias". The advice to check with a different model stands. The size of the effect is less settled than the book suggests.

Pages 64 and 70, Chapter 5, Ayinde. The book says the leadership duty "was laid on top". The source note on page 70 says the court set out leadership responsibilities "on top of the individual duty". The judgment does not describe a new duty. It says that practical and effective measures must now be taken by those with leadership responsibilities, and by the regulators. It adds that in hearings of this kind the profession can expect the court to ask whether those responsibilities have been met.

Pages 94 and 99, Chapter 8, entry-level hiring. The book says "the alarm can run ahead of the evidence", citing an employer survey by the Strada Institute. Payroll research revised in August 2026 (Brynjolfsson, Chandar and Chen) points the other way. It finds that employment of 22- to 25-year-olds in the most AI-exposed occupations fell about 11% between November 2022 and June 2026, while the least-exposed grew about 10%. That is about 19% lower in relative terms. The authors say this is descriptive, not a causal estimate. On this evidence, the book's reassurance no longer holds as written.

Page 112, Chapter 9, where to keep the three lines. The book advises keeping the three lines "where the system runs". Keep them where your other controls live, away from the system they govern, so that a change to the system cannot quietly change them too.