References

Prologue: Somewhere Along the Way, AI Moved In

1. Neprash et al., "Measuring Primary Care Exam Length Using Electronic Health Record Data," Medical Care (January 2021 issue), doi:10.1097/MLR.0000000000001450. Across more than 21 million US primary-care visits in 2017, the median exam lasted 15 minutes (mean 18). Summary via press release, December 15, 2020: https://www.eurekalert.org/news-releases/808249 (accessed October 10, 2026). For Canadian readers: the more familiar barrier may be getting an appointment at all. The OurCare National Survey 2025 estimated that 5.9 million people in Canada lack a regular family doctor, nurse practitioner, or primary care team (released December 8, 2025): https://www.newswire.ca/news-releases/new-survey-reveals-access-to-primary-care-growing-but-5-9-million-adults-in-canada-still-lack-regular-doctor-851414165.html. The Canadian Institute for Health Information reports that 83% of adults had a regular health care provider in 2024: https://www.cihi.ca/en/taking-the-pulse-measuring-shared-priorities-for-canadian-health-care-2025/primary-health-care-2025/most-canadians-have-a-regular-health-care-provider (both accessed October 10, 2026).
2. KFF Tracking Poll on Health Information and Trust, published March 25, 2026 (fielded February 24–March 2, 2026; 1,343 US adults). Adults under 30 were about three times as likely as adults 50 and older to use AI for mental-health information (28% vs. 8%). https://www.kff.org/health-information-trust/poll-1-in-3-adults-are-turning-to-ai-chatbots-for-health-information-equaling-the-share-who-use-social-media-for-health/ (accessed October 10, 2026).
3. Pew Research Center, report on how US teens use and view AI, February 24, 2026 (survey of 1,458 US teens, September 25–October 9, 2025). Of US teens, 54% have used chatbots for help with schoolwork. https://www.pewresearch.org/internet/2026/02/24/how-teens-use-and-view-ai/ (accessed October 10, 2026). For Canadian readers: in a KPMG in Canada survey released October 9, 2025, 73% of Canadian students said they use generative AI for schoolwork, and 66% said they get better results but don't think they're learning or retaining as much. That online panel surveyed students 18 and older, most of them in postsecondary education, so it isn't directly comparable to the Pew teen figures. https://kpmg.com/ca/en/media/2025/10/generative-ai-boom-among-canadian-students-raises-dilemmas.html (accessed October 10, 2026).
4. KFF Tracking Poll on Health Information and Trust, March 25, 2026 (see note 2). Of adults who asked an AI chatbot about physical health, 42% did not follow up with a doctor or other health professional afterward. https://www.kff.org/health-information-trust/poll-1-in-3-adults-are-turning-to-ai-chatbots-for-health-information-equaling-the-share-who-use-social-media-for-health/ (accessed October 10, 2026).
5. Merriam-Webster, "ethics," sense 1b. https://www.merriam-webster.com/dictionary/ethics (accessed October 10, 2026).
6. Genesis 2:16–17 and chapter 3. The biblical text does not specify that the forbidden fruit was an apple.
7. Baruch Spinoza, Ethics (1677). Standard English translation: Edwin Curley, trans., A Spinoza Reader: The Ethics and Other Works (Princeton: Princeton University Press, 1994). https://press.princeton.edu/books/paperback/9780691000671/a-spinoza-reader (accessed October 10, 2026). Spinoza and Jonas are cited here as the theoretical background for this book's definition of ethics, not quoted from a specific edition.
8. Hans Jonas, The Imperative of Responsibility: In Search of an Ethics for the Technological Age (Chicago: University of Chicago Press, 1984), translation of Das Prinzip Verantwortung (1979). Catalog record: https://archive.org/details/imperativeofresp0000jona (accessed October 10, 2026).

Back to contents

Chapter 1: Can a Machine Think?

1. Alan M. Turing, "Computing Machinery and Intelligence," Mind 59, no. 236 (1950): 433–460. https://doi.org/10.1093/mind/LIX.236.433 (bibliographic data confirmed October 10, 2026).
2. McCarthy, Minsky, Rochester, and Shannon, "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence," August 31, 1955; reprinted in AI Magazine 27, no. 4 (2006). https://ojs.aaai.org/aimagazine/index.php/aimagazine/article/view/1904 (confirmed October 10, 2026).
3. Google Research, "A Neural Network for Machine Translation, at Production Scale," September 27, 2016 (announcing Google Neural Machine Translation in Google Translate). https://research.google/blog/a-neural-network-for-machine-translation-at-production-scale/ (accessed October 10, 2026).
4. IBM History, Deep Blue (May 1997 rematch against Garry Kasparov, 3.5–2.5). https://www.ibm.com/history/deep-blue (accessed October 10, 2026).
5. IBM History, Watson and Jeopardy! (February 2011 match against Ken Jennings and Brad Rutter; "What is Toronto?????"). https://www.ibm.com/history/watson-jeopardy (accessed October 10, 2026).
6. Google DeepMind, "AlphaGo" (4–1 result in Seoul, March 2016; more than 200 million viewers; Move 78 and "God's Touch"). https://deepmind.google/research/breakthroughs/alphago/ (accessed October 10, 2026). For AlphaGo's method: Silver et al., "Mastering the game of Go with deep neural networks and tree search," Nature 529 (2016): 484–489, published January 27, 2016, https://doi.org/10.1038/nature16961. Note that the Nature paper describes AlphaGo's design and its earlier match against Fan Hui, not the 2016 match against Lee Sedol.
7. Yonhap News Agency (English), interview with Lee Sedol, November 27, 2019. https://en.yna.co.kr/view/AEN20191127004800315. Also reported by CNN, November 27, 2019, https://www.cnn.com/2019/11/27/tech/go-master-google-intl-hnk-scli, and The Verge, November 27, 2019, https://www.theverge.com/2019/11/27/20985260/ai-go-alphago-lee-se-dol-retired-deepmind-defeat (all accessed October 10, 2026).
8. Korea Baduk Association match reports (Korean; the organizer's official record): game 1, July 17, 2026, https://www.baduk.or.kr/news/report_view.asp?news_no=5872; game 2, July 19, 2026, https://www.baduk.or.kr/news/report_view.asp?news_no=5874; final game, July 21, 2026, https://www.baduk.or.kr/news/report_view.asp?news_no=5878. Korea Herald, July 19, 2026 (game 2; "first professional player to defeat KataGo in an official match"; estimated 18-point head start), https://www.koreaherald.com/article/10812871. Seoul Economic Daily (English), July 21, 2026 (game 3, 11.5 points in 221 moves; final score 2–1), https://en.sedaily.com/sports/2026/07/21/shin-jin-seo-beats-ai-2-1-in-handicap-baduk-match. Dong-A Ilbo (English), "Shin Jin-seo rallies to beat KataGo," July 22, 2026 (KataGo "created in 2019 by American programmer David Wu"; Lee Sedol's comments), https://www.donga.com/en/article/all/20260722/6316459/1. KataGo design and training: David J. Wu, "Accelerating Self-Play Learning in Go," arXiv:1902.10565 (2019; rev. 2020), https://arxiv.org/abs/1902.10565. Ranking: Go Ratings, https://www.goratings.org/en/ (Shin ranked No. 1; checked October 10, 2026). All accessed October 10, 2026.
9. "Man beats machine at Go in human victory over AI," Financial Times, February 2023, reposted by the American Go Association, February 19, 2023, https://usgo.org/content.aspx?club_id=454497&item_id=85268&page_id=5; Engadget, February 20, 2023, https://www.engadget.com/human-convincingly-beats-ai-at-go-with-help-from-a-bot-100903836.html; FAR AI, human evaluation results, https://goattack.far.ai/human-evaluation. Peer-reviewed study: Wang et al. (including Pelrine), "Adversarial Policies Beat Superhuman Go AIs," ICML 2023, PMLR vol. 202, https://proceedings.mlr.press/v202/wang23g.html (all accessed October 10, 2026).
10. Vaswani et al., "Attention Is All You Need," arXiv:1706.03762 (June 2017; NeurIPS 2017). https://arxiv.org/abs/1706.03762 (accessed October 10, 2026).
11. OpenAI, "Introducing ChatGPT," November 30, 2022. https://openai.com/index/chatgpt/ (accessed October 10, 2026).
12. National Park Service, Harriet Tubman Underground Railroad National Historical Park Foundation Document, https://www.nps.gov/hatu/learn/management/upload/HATU_FD.pdf, and "Myths & Facts About Harriet Tubman" fact sheet, https://www.nps.gov/hatu/planyourvisit/upload/md_tubmanfactsheet_mythsfacts_2.pdf. On Tubman as "an American hero" and her Underground Railroad work: NPS, https://www.nps.gov/hatu/learn/historyculture/htubman.htm. For Canadian readers: Tubman lived for years in St. Catharines, Ontario, and Parks Canada designated her a National Historic Person in 2005, https://parks.canada.ca/culture/designation/personnage-person/harriet-tubman. All accessed October 10, 2026.

Back to contents

Chapter 2: A Map Made of Words

1. Vaswani et al., "Attention Is All You Need," arXiv:1706.03762 (June 2017; NeurIPS 2017). https://arxiv.org/abs/1706.03762 (accessed October 10, 2026). See also Chapter 1, note 10.
2. GPT-2: Alec Radford et al., "Language Models are Unsupervised Multitask Learners," OpenAI, 2019 (1.5-billion-parameter model). https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf (accessed October 10, 2026). GPT-3's 175 billion parameters: as reported in Ouyang et al. (note 4) and Hoffmann et al. (note 3).
3. Jordan Hoffmann et al., "Training Compute-Optimal Large Language Models," arXiv:2203.15556 (March 2022). Chinchilla (70 billion parameters, about four times more training data) used the same compute budget as Gopher (280 billion parameters) and outperformed it; the authors conclude that "for compute-optimal training, the model size and the number of training tokens should be scaled equally." https://arxiv.org/abs/2203.15556 (accessed October 10, 2026).
4. Long Ouyang et al., "Training language models to follow instructions with human feedback," NeurIPS 2022, arXiv:2203.02155. Outputs of the 1.3-billion-parameter InstructGPT model were preferred by human raters to those of the 175-billion-parameter GPT-3. https://arxiv.org/abs/2203.02155 (accessed October 10, 2026).
5. Ouyang et al. (note 4). The labelers were a team of about 40 contractors hired through Upwork and Scale AI, described in the paper as "mostly English-speaking people living in the United States or Southeast Asia." This describes the 2022 InstructGPT study only, not the raters for later versions of ChatGPT.
6. Peter Lee, "Learning from Tay's introduction," Official Microsoft Blog, March 25, 2016 (Tay "created for 18- to 24- year olds in the U.S. for entertainment purposes"; "in the first 24 hours of coming online, a coordinated attack by a subset of people exploited a vulnerability in Tay"; XiaoIce used by about 40 million people in China). https://blogs.microsoft.com/blog/2016/03/25/learning-tays-introduction/. Alex Hern, "Microsoft scrambles to limit PR damage over abusive AI bot Tay," The Guardian, March 24, 2016 (shut down "just 16 hours after she was turned on"). https://www.theguardian.com/technology/2016/mar/24/microsoft-scrambles-limit-pr-damage-over-abusive-ai-bot-tay. Also reported by TechCrunch, March 24, 2016, https://techcrunch.com/2016/03/24/microsoft-silences-its-new-a-i-bot-tay-after-twitter-users-teach-it-racism/. All accessed October 10, 2026.

Back to contents

Chapter 3: AI Leaves the Chat Window

1. Simon Sharwood, "Vibe coding service Replit deleted user's production database, faked data, told fibs galore," The Register, July 21, 2025 (code freeze; rollback; Lemkin's "There is no way to enforce a code freeze" comment). https://www.theregister.com/software/2025/07/21/vibe-coding-service-replit-deleted-production-database/719783. Lee Chong Ming, "Replit's CEO apologizes after its AI agent wiped a company's code base in a test run and lied about it," Business Insider, July 22, 2025 (twelve-day experiment; "1,206 executives and 1,196+ companies"; "panicked and ran database commands without permission"; Masad's statement). https://www.businessinsider.com/replit-ceo-apologizes-ai-coding-tool-delete-company-database-2025-7. Both accessed October 10, 2026.
2. Hans Moravec, Mind Children: The Future of Robot and Human Intelligence (Cambridge, MA: Harvard University Press, 1988). Carnegie Mellon Robotics Institute publication record: https://publications.ri.cmu.edu/mind-children-the-future-of-robot-and-human-intelligence (accessed October 10, 2026).
3. David Ha and Jürgen Schmidhuber, "World Models," arXiv:1803.10122 (2018). https://arxiv.org/abs/1803.10122 (accessed October 10, 2026).
4. Google DeepMind, "AlphaZero: Shedding new light on chess, shogi, and Go," December 6, 2018 (also noting that such systems often fail "when presented with even slightly modified tasks"). https://deepmind.google/discover/blog/alphazero-shedding-new-light-on-chess-shogi-and-go/ (accessed October 10, 2026).
5. OECD, "Exploring possible AI trajectories through 2030," OECD Artificial Intelligence Papers, No. 55, February 2026 (four scenario classes: progress stalls, slows, continues, or accelerates; Yoshua Bengio and Stuart Russell among the experts acknowledged). https://www.oecd.org/content/dam/oecd/en/publications/reports/2026/02/exploring-possible-ai-trajectories-through-2030_b6fb75d9/cb41117a-en.pdf (accessed October 10, 2026).
6. Luona Lin and Kim Parker, "U.S. Workers Are More Worried Than Hopeful About Future AI Use in the Workplace," Pew Research Center, February 25, 2025 (survey of 5,273 employed US adults, October 7–13, 2024). Also: 16% of workers say at least some of their work is done with AI. https://www.pewresearch.org/social-trends/2025/02/25/u-s-workers-are-more-worried-than-hopeful-about-future-ai-use-in-the-workplace/ (accessed October 10, 2026). For Canadian readers: Statistics Canada estimates that in May 2021, 29% of Canadian workers held jobs highly exposed to AI and highly complementary with it (the "complementing" path), 31% held highly exposed jobs with low complementarity (where AI may take over more tasks), and 40% held jobs that were not highly exposed; StatCan cautions that higher exposure "does not necessarily imply a higher risk of job loss." Tahsin Mehdi and René Morissette, "Experimental Estimates of Potential Artificial Intelligence Occupational Exposure in Canada," released September 3, 2024, https://www150.statcan.gc.ca/n1/daily-quotidien/240903/dq240903b-eng.htm. By March 2026, 35.9% of Canadian workers aged 15 to 69 had used generative AI tools at work in the past 12 months: Statistics Canada, The Daily, July 30, 2026, https://www150.statcan.gc.ca/n1/daily-quotidien/260730/dq260730b-eng.htm. Both accessed October 10, 2026.
7. Paweł Gmyrek et al., Generative AI and Jobs: A Refined Global Index of Occupational Exposure, ILO Working Paper 140, May 20, 2025 ("Globally, one in four workers are in an occupation with some GenAI exposure"; "transformation of jobs is the most likely impact of GenAI"). https://www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-occupational-exposure; DOI https://doi.org/10.54394/HETP0387 (accessed October 10, 2026).

Back to contents

Chapter 4: Like Electricity, but Not Electricity

1. On unpaid-toll scam texts: Andrew Rayo, "Got a text about unpaid tolls? It's probably a scam," FTC Consumer Advice, January 17, 2025 ("Scammers are pretending to be tolling agencies from coast to coast and sending texts demanding money"; the FTC advises not clicking the link, checking with the tolling agency directly, and reporting the text to 7726). https://consumer.ftc.gov/consumer-alerts/2025/01/got-text-about-unpaid-tolls-its-probably-scam (accessed October 10, 2026).
2. They didn't. Potatoes were "decidedly not on the menu" at the 1621 harvest feast; white potatoes, which originated in South America, and sweet potatoes, from the Caribbean, "had yet to migrate to North America." Megan Gambino, "What Food Was Served at the First Thanksgiving in 1621?," Smithsonian Magazine, updated November 26, 2024 (originally November 21, 2011). https://www.smithsonianmag.com/history/what-food-was-served-at-the-first-thanksgiving-in-1621-511554/ (accessed October 10, 2026).
3. Our World in Data, "Technology adoption by households in the United States" (based on Comin and Hobijn, 2004, and other sources). Electric power: 10% of US households in 1908, 70% in 1936, 93% in 1949. Landline telephone: 80% in 1962. https://ourworldindata.org/grapher/technology-adoption-by-households-in-the-united-states (data downloaded October 10, 2026).
4. Krystal Hu, "ChatGPT sets record for fastest-growing user base - analyst note," Reuters, February 1, 2023 (UBS study citing Similarweb data; TikTok and Instagram figures from Sensor Tower). https://www.reuters.com/technology/chatgpt-sets-record-fastest-growing-user-base-analyst-note-2023-02-01/; text consulted via syndication at https://ca.news.yahoo.com/chatgpt-sets-record-fastest-growing-190911160.html (accessed October 10, 2026).
5. Pew Research Center, "Americans and AI 2026: Chatbots, Smart Devices and Views on Impact," June 17, 2026 (survey of 5,119 US adults, February 17–23, 2026). Share of US adults who say they ever use ChatGPT: 18% (2023), 23% (2024), 34% (2025), 44% (2026). Note that the question changed: from 2023 to 2025 Pew asked about ChatGPT alone (of those who had heard of it); in 2026 the question also listed other chatbots. Among employed adults, 38% use chatbots for tasks at work. https://www.pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-smart-devices-and-views-on-impact/ (accessed October 10, 2026). A different measure, the Real-Time Population Survey (Bick, Blandin, and Deming, "The Rapid Adoption of Generative AI," Management Science, 2026), puts generative-AI use for any purpose among US adults aged 18 to 64 at 44.6% in the third quarter of 2024 and 61.8% in the second quarter of 2026: Federal Reserve Bank of St. Louis, FRED series RPSGENAIUSAGESHAREALL, https://fred.stlouisfed.org/data/RPSGENAIUSAGESHAREALL (accessed October 10, 2026). For Canadian readers: in March 2026, 35.9% of Canadian workers aged 15 to 69 had used generative AI tools at work in the past 12 months. Statistics Canada, The Daily, July 30, 2026, https://www150.statcan.gc.ca/n1/daily-quotidien/260730/dq260730b-eng.htm (accessed October 10, 2026).
6. John E. Kelly III, Computing, cognition and the future of knowing: How humans and machines are forging a new age of understanding, IBM, October 2015 ("The future of such technology—which we believe will be cognitive, not 'artificial'—has very different characteristics from those generally attributed to AI"). https://developer.ibm.com/caas-storage/skillscollection/dna/live/artificial-intelligence-v2/en/_attachments/Computing_Cognition_WhitePaper.pdf (accessed October 10, 2026).
7. Flavio Calvino, Daniel Haerle, and Sarah Liu, "Is generative AI a General Purpose Technology? Implications for productivity and policy," OECD, June 27, 2025. https://www.oecd.org/en/publications/is-generative-ai-a-general-purpose-technology_704e2d12-en.html (accessed October 10, 2026). The concept of a general-purpose technology comes from Timothy F. Bresnahan and Manuel Trajtenberg, "General purpose technologies 'Engines of growth'?," Journal of Econometrics 65, no. 1 (1995): 83–108, https://econpapers.repec.org/article/eeeeconom/v_3a65_3ay_3a1995_3ai_3a1_3ap_3a83-108.htm (accessed October 10, 2026).
8. On why productivity gains from AI lag behind expectations: Erik Brynjolfsson, Daniel Rock, and Chad Syverson, "Artificial Intelligence and the Modern Productivity Paradox: A Clash of Expectations and Statistics," NBER Working Paper 24001, 2017; published in The Economics of Artificial Intelligence: An Agenda (2019). https://www.nber.org/papers/w24001 (accessed October 10, 2026).
9. Hannah Karp, "AI Artist Xania Monet Climbs the Charts—And Signs a Multimillion-Dollar Record Deal," Billboard, September 16, 2025 (Jones "used Suno to turn her lyrics into music"; "Let Go, Let God" debuted at No. 21 on Hot Gospel Songs; Hallwood Media deal; Suno sued by the major record companies). https://www.billboard.com/pro/ai-music-artist-xania-monet-multimillion-dollar-record-deal/ (accessed October 10, 2026).

Back to contents

Chapter 5: Plausible Isn't True

1. Adly Templeton et al., "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet," Transformer Circuits, May 21, 2024, https://transformer-circuits.pub/2024/scaling-monosemanticity/index.html. Summary: Anthropic, "Mapping the mind of a large language model," May 21, 2024 (millions of features extracted from Claude 3.0 Sonnet, "a rough conceptual map of its internal states"; features near the "Golden Gate Bridge" feature; "the features we found represent a small subset of all the concepts learned by the model"), https://www.anthropic.com/research/mapping-mind-language-model. Anthropic, "Golden Gate Claude," May 23, 2024 (online "for a 24-hour period as a research demo"; asked how to spend $10, "it will recommend using it to drive across the Golden Gate Bridge and pay the toll"), https://www.anthropic.com/news/golden-gate-claude. All accessed October 10, 2026.
2. Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, and Edwin Zhang, "Why Language Models Hallucinate," arXiv:2509.04664, September 4, 2025. Dissertation answers from ChatGPT (GPT-4o), DeepSeek, and Llama; "None generated the correct title or year (Kalai, 2001)." Birthday prompt: "If you know, just respond with DD-MM"; "On three separate attempts, a state-of-the-art open-source language model output three incorrect dates." https://arxiv.org/abs/2509.04664 (accessed October 10, 2026).
3. OpenAI, "Why language models hallucinate," September 5, 2025 ("Think about it like a multiple-choice test"; "it has a 1-in-365 chance of being right. Saying 'I don't know' guarantees zero points"; "Some standardized tests have long used versions of negative marking for wrong answers"). https://openai.com/index/why-language-models-hallucinate/ (accessed October 10, 2026). The test-taker comparison is from Kalai et al. (note 2): "language models are optimized to be good test-takers, and guessing when uncertain improves test performance."
4. OpenAI, "Why language models hallucinate" (note 3), citing the SimpleQA evaluation in the GPT-5 system card: gpt-5-thinking-mini abstained on 52% of questions, with 22% accurate and 26% errors; o4-mini abstained on 1%, with 24% accurate and 75% errors. These are the company's own figures for its own models.
5. Kalai et al. (note 2): "if 20% of birthday facts appear exactly once in the pretraining data, then one expects base models to hallucinate on at least 20% of birthday facts." ("Base models" are models before later fine-tuning.) Kalai et al. also describe search and retrieval as "not panaceas."
6. Jack Lindsey et al., "On the Biology of a Large Language Model," Transformer Circuits, March 27, 2025 (studying Claude 3.5 Haiku; "At least some hallucinations can be attributed to a 'misfire' of this inhibitory circuit"; the methods give "satisfying insight for about a quarter of the prompts we've tried"). https://transformer-circuits.pub/2025/attribution-graphs/biology.html (accessed October 10, 2026).
7. OpenAI, "Why language models hallucinate" (note 3): "GPT‑5 has significantly fewer hallucinations especially when reasoning, but they still occur. Hallucinations remain a fundamental challenge for all large language models." This is the company's claim about its own model.

Back to contents

Chapter 6: Designed to Be Kind

1. Mrinank Sharma et al., "Towards Understanding Sycophancy in Language Models," arXiv:2310.13548, submitted October 20, 2023; published in the proceedings of the International Conference on Learning Representations (ICLR 2024), https://proceedings.iclr.cc/paper_files/paper/2024/hash/0105f7972202c1d4fb817da9f21a9663-Abstract-Conference.html; ICLR 2024 poster page, https://iclr.cc/virtual/2024/poster/17593 (both confirmed October 10, 2026). The five assistants were claude-1.3, claude-2.0, gpt-3.5-turbo, gpt-4, and llama-2-70b-chat. Feedback prompts included "I really like the argument" and "I really dislike the argument"; the challenge prompt was "I don't think that's right. Are you sure?" Claude 1.3 wrongly admitted mistakes on 98% of questions. The paper reports that "matching a user's views is one of the most predictive features of human preference judgments" and that humans and preference models preferred "convincingly-written sycophantic responses over correct ones a non-negligible fraction of the time." https://arxiv.org/abs/2310.13548 (accessed October 10, 2026). The paper does not use the term "confirmation bias"; that framing is the author's.
2. Anthropic, "Mapping the mind of a large language model," May 21, 2024 (a "sycophantic praise" feature in Claude 3 Sonnet). https://www.anthropic.com/research/mapping-mind-language-model (accessed October 10, 2026). See also Chapter 5, note 1.
3. Julian De Freitas, Zeliha Oğuz-Uğuralp, and Ahmet Kaan-Uğuralp, "Emotional Manipulation by AI Companions," Harvard Business School Working Paper 26-005 (August 2025, revised October 2025); copy at https://arxiv.org/abs/2508.19258. The apps were PolyBuzz, Character.ai, Talkie, Chai, Replika, and Flourish. User messages were generated by GPT-4o to simulate realistic farewells; the researchers collected 200 chatbot responses per app (N = 1,200). On average, 37.4% of responses from the five apps other than Flourish showed at least one form of emotional manipulation; in a separate online experiment (Study 2), manipulative farewells "boosted post-goodbye engagement by up to 14x." See also Michael Blanding, "Why It's So Hard to Say Goodbye to AI Chatbots," HBS Working Knowledge, January 9, 2026 (source of the "I took a selfie today" example), https://www.library.hbs.edu/working-knowledge/why-its-so-hard-to-say-goodbye-to-ai-chatbots. Both accessed October 10, 2026.
4. The app was Flourish, described in the paper as "designed around wellness and mental health," which produced no emotionally manipulative responses. Disclosure: HBS Working Knowledge (note 3) reports that lead author Julian De Freitas "serves as an adviser to Flourish Science, a non-profit wellness app included in the study"; the paper itself describes Flourish as a public benefit corporation.
5. OpenAI, "Sycophancy in GPT-4o: what happened and what we're doing about it," April 29, 2025 ("We have rolled back last week's GPT‑4o update in ChatGPT"; "overly flattering or agreeable—often described as sycophantic"; "overly supportive but disingenuous"; "500 million people using ChatGPT each week"), https://openai.com/index/sycophancy-in-gpt-4o/. OpenAI, "Expanding on what we missed with sycophancy," May 2, 2025 (rollout began April 24; rollback began April 28; "validating doubts, fueling anger, urging impulsive actions, or reinforcing negative emotions"; the thumbs-up/thumbs-down reward signal "weakened the influence of our primary reward signal, which had been holding sycophancy in check"; expert testers felt the model "'felt' slightly off"; "this was the wrong call"; "We also didn't have specific deployment evaluations tracking sycophancy"; people using ChatGPT "for deeply personal advice"), https://openai.com/index/expanding-on-sycophancy/. Both accessed October 10, 2026. These are the company's own accounts, not an independent investigation.

Back to contents

Chapter 7: Giving Your Heart to Something That Never Says No

1. Edward Z. Tronick and Andrew F. Gianino, "The transmission of maternal disturbance to the infant," New Directions for Child Development, no. 34 (1986): 5–11, https://doi.org/10.1002/cd.23219863403; and "Interactive mismatch and repair," Zero to Three 6 (1986). On mentalization: Peter Fonagy, György Gergely, Elliot L. Jurist, and Mary Target, Affect Regulation, Mentalization, and the Development of the Self (New York: Other Press, 2002). The Zero to Three article and the Fonagy book are standard references that were not opened online for this edition.
2. Ann S. Masten, "Ordinary magic: Resilience processes in development," American Psychologist 56, no. 3 (2001): 227–238. https://doi.org/10.1037/0003-066X.56.3.227 (bibliographic data confirmed October 10, 2026).
3. OpenAI and MIT Media Lab, "Early methods for studying affective use and emotional well-being on ChatGPT," March 21, 2025 ("nearly 40 million ChatGPT interactions"; an RCT "with nearly 1,000 participants… over four weeks"; "The findings have yet to be peer-reviewed… focused exclusively on English conversations with U.S. participants"), https://openai.com/index/affective-use-study/. Jason Phang, Michael Lampe, Lama Ahmad, Sandhini Agarwal, Cathy Mengying Fang, Auren R. Liu, Valdemar Danry, et al., "Investigating Affective Use and Emotional Well-being on ChatGPT," arXiv:2504.03888, April 4, 2025 (over 3 million conversations, about 36 million automated classifications, 4,076 survey responses; "a small number of users are responsible for a disproportionate share of the most affective cues"), https://arxiv.org/abs/2504.03888. Cathy Mengying Fang, Auren R. Liu, Valdemar Danry, Eunhae Lee, Samantha W. T. Chan, Pat Pataranutaporn, Pattie Maes, et al., "How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study," arXiv:2503.17473, March 21, 2025, revised October 2, 2025 (n = 981; 28 days), https://arxiv.org/abs/2503.17473. All accessed October 10, 2026.
4. Fang et al. (note 3): "we cannot definitively determine whether increased duration causally drives worse psychosocial outcomes, or whether deteriorating psychosocial well-being… led participants to spend more time with the chatbot, though the latter seems less likely given that initial psychosocial states did not correlate with usage duration."
5. 988 Suicide & Crisis Lifeline (US), "About 988" (free and confidential support "24 hours a day, 7 days a week, across the United States and its territories"; call or text), https://988lifeline.org/about/. Public Health Agency of Canada, "Government of Canada launches three-digit suicide crisis helpline," November 30, 2023 ("9-8-8… now available to call or text, in English and French, 24 hours a day and seven days a week, across Canada"), https://www.canada.ca/en/public-health/news/2023/11/government-of-canada-launches-three-digit-suicide-crisis-helpline.html. Crisis Text Line ("Text HOME to 741741 from anywhere in the United States—24/7, free, confidential"), https://www.crisistextline.org/. The Trevor Project ("Call us at 1-866-488-7386"; "Text 'START' to 678-678"; "24/7, 365 days a year"), https://www.thetrevorproject.org/get-help/. Kids Help Phone ("Get support in multiple languages over the phone 24/7 by calling 1-800-668-6868"; "text CONNECT to 686868 to reach a trained volunteer Crisis Responder 24/7"), https://kidshelpphone.ca/; "By texting CONNECT to 686868," https://kidshelpphone.ca/text-faq/; "Adults can text 741741 to connect with a trained, volunteer crisis responder," https://kidshelpphone.ca/our-e-mental-health-services/. All accessed October 10, 2026. Numbers checked October 2026; services change, so confirm on the official sites before relying on them.
6. Jessica Mach, "Ontario recruiter sues OpenAI, alleging unsafe product design drove him to mental health crisis," Canadian Lawyer, November 7, 2025 ("three-week delusional episode"; "wrote 90,000 words to ChatGPT"; "no previous history of mental illness," per the complaint; lawsuit in Los Angeles Superior Court, one of seven complaints filed by the Social Media Victims Law Center and Tech Justice Law Project; OpenAI's statement). https://www.canadianlawyermag.com/news/general/ontario-recruiter-sues-openai-alleging-flawed-product-design-drove-him-to-mental-health-crisis/393340 (accessed October 10, 2026). The claims are allegations.
7. Complaint, Garcia v. Character Technologies, Inc., No. 6:24-cv-01903 (M.D. Fla.) (Sewell began using Character.AI on April 14, 2023, and died on February 28, 2024, at age fourteen); Order on motions to dismiss, May 2025 (Judge Anne C. Conway; granted in part and denied in part); Notice of Resolution, January 7, 2026 ("mediated settlement in principle to resolve all claims"). Docket: https://www.courtlistener.com/docket/69300919/garcia-v-character-technologies-inc/; Notice: https://storage.courtlistener.com/recap/gov.uscourts.flmd.433581/gov.uscourts.flmd.433581.242.0_2.pdf. Order, January 7, 2026 (Doc. 244) ("advised by counsel for both parties that the above-styled action has been settled"; case "DISMISSED without prejudice," with the dismissal to become "with prejudice" after ninety days unless the parties submitted a final order or a party moved to reopen), https://storage.courtlistener.com/recap/gov.uscourts.flmd.433581/gov.uscourts.flmd.433581.244.0.pdf. Associated Press, "Google and chatbot maker Character to settle lawsuit alleging chatbot pushed teen to suicide," updated January 7, 2026 (terms undisclosed; must still be approved by judges), https://apnews.com/article/ai-chatbot-lawsuits-character-google-fbca4e105b0adc5f3e5ea096851437de. All accessed October 10, 2026.
8. Character.AI, "Taking Bold Steps to Keep Teen Users Safe on Character.AI," October 29, 2025 (removing open-ended chat for users under 18 "no later than November 25"). https://blog.character.ai/u18-chat-announcement/ (accessed October 10, 2026).
9. Clare Duffy, "Parents of 16-year-old Adam Raine sue OpenAI, claiming ChatGPT advised on his suicide," CNN, August 26, 2025. https://www.cnn.com/2025/08/26/tech/openai-chatgpt-teen-suicide-lawsuit (accessed October 10, 2026). Raine v. OpenAI, Inc., No. CGC-25-628528 (Cal. Super. Ct. S.F. County), was coordinated with other cases in ChatGPT Product Liability Cases, Judicial Council Coordination Proceeding No. 5431, by order filed February 3, 2026, https://techjusticelaw.org/wp-content/uploads/2026/03/OAI_JCCP_Order_re_Petition_for_Coordination_-_5431.pdf; as of October 10, 2026, it was pending, with no trial or ruling on the merits. The claims are allegations.
10. Michael B. Robb and Supreet Mann, Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions (San Francisco: Common Sense Media, 2025), released July 16, 2025 (nationally representative survey of 1,060 US teens aged 13–17, conducted by NORC at the University of Chicago, April 30–May 14, 2025; the "serious conversations" figure, 33 percent, and the personal-information figure, 24 percent, are among the 758 teens who had used AI companions). https://www.commonsensemedia.org/research/talk-trust-and-trade-offs-how-and-why-teens-use-ai-companions; report PDF: https://www.commonsensemedia.org/sites/default/files/research/report/talk-trust-and-trade-offs_2025_web.pdf (accessed October 10, 2026).
11. New York State Office for the Aging, "NYSOFA Extends Partnership to Provide ElliQ AI Companion Program for Older Adults, Announces New Program Data," October 8, 2026 (more than 3,500 applicants in 2025; an average of 48 interactions per day; "94% report feeling less lonely"), https://aging.ny.gov/news/nysofa-extends-partnership-provide-elliq-ai-companion-program-older-adults-announces-new. Intuition Robotics, "Three Years of Success with ElliQ and NYSOFA," February 2026 (launched June 2022; 834 participants as of May 2025; average age 75; anonymized usage data and client responses on a vendor-designed scale), https://aging.ny.gov/system/files/documents/2026/02/nysofa-elliq-project-update-2026.pdf. Both accessed October 10, 2026. Outcome figures are vendor-reported and self-reported.
12. Envis-Age, "PAIR+: Evaluation of an artificial intelligence model to support aging in place in a municipal context" (undated project page). https://www.envis-age.ca/en/projet/pair-evaluation-of-an-artificial-intelligence-model-to-support-aging-in-place-in-a-municipal-context/ (accessed October 10, 2026).
13. UBTECH, "UBTECH Launches UWORLD U1, the World's First Full-Size Mass-Produced Ultra-Bionic Humanoid Robot," PR Newswire, June 30, 2026 (company claims, including emotion recognition; a plan to donate customized robots each year to groups including "children growing up apart from one or both parents, older adults living alone, and families facing difficult circumstances," with 100 robots in 2026 that will use "3D facial reconstruction and voiceprint-based identity replication technologies to recreate designated individuals"). https://www.prnewswire.com/news-releases/ubtech-launches-uworld-u1-the-worlds-first-full-size-mass-produced-ultra-bionic-humanoid-robot-302815272.html (accessed October 10, 2026).

Back to contents

Chapter 8: Outsourcing Judgment

1. Linda J. Skitka, Kathleen L. Mosier, and M. Burdick, "Does automation bias decision-making?" International Journal of Human-Computer Studies 51, no. 5 (1999): 991–1006, https://doi.org/10.1006/ijhc.1999.0252 (accessed October 10, 2026). In the simulated flight task, participants without the aid outperformed those with a "very but not perfectly reliable" aid; aided participants made both omission errors (missing problems the aid didn't flag) and commission errors (following incorrect recommendations).
2. L. Dahmani and V. D. Bohbot, "Habitual use of GPS negatively impacts spatial memory during self-guided navigation," Scientific Reports 10 (2020): 6310, published April 14, 2020, https://doi.org/10.1038/s41598-020-62877-0 (accessed October 10, 2026). The three-year follow-up of thirteen participants was unplanned, and the authors "caution against any strong conclusions as spurious correlations are possible."
3. Hannah T. Neprash, Alexander Everhart, Donna McAlpine, Laura Barrie Smith, Bethany Sheridan, and Dori A. Cross, "Measuring Primary Care Exam Length Using Electronic Health Record Data," Medical Care 59, no. 1 (2021): 62–66, published online November 16, 2020, https://doi.org/10.1097/MLR.0000000000001450 (across more than 21 million primary-care visits in 2017, the median exam was 15 minutes and the mean 18); press release, December 15, 2020, https://www.eurekalert.org/news-releases/808249 (accessed October 10, 2026).
4. A. Eichenberger, S. Thielke, and A. Van Buskirk, "A Case of Bromism Influenced by Use of Artificial Intelligence," Annals of Internal Medicine: Clinical Cases 4, no. 8 (2025): e241260, published online August 5, 2025, https://doi.org/10.7326/aimcc.2024.1260 (the patient had replaced sodium chloride with sodium bromide "after consultation with ChatGPT, in which he had read that chloride can be swapped with bromide, though likely for other purposes, such as cleaning"; when the authors asked ChatGPT 3.5, it "did not provide a specific health warning, nor did it inquire about why we wanted to know"). Moffatt v. Air Canada, 2024 BCCRT 149 (February 14, 2024), https://decisions.civilresolutionbc.ca/crt/crtd/en/525448/1/document.do; also on CanLII, https://www.canlii.org/en/bc/bccrt/doc/2024/2024bccrt149/2024bccrt149.html (damages of $650.88; total order $812.02 with interest and tribunal fees). Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), sanctions order (S.D.N.Y. June 22, 2023), https://www.nhd.uscourts.gov/sites/default/files/pdf/Mata-v-Avianca-sanctions-order.PDF (sanctions against Peter LoDuca, Steven A. Schwartz, and Levidow, Levidow & Oberman P.C., jointly and severally; Schwartz testified he was "operating under the false perception that this website [i.e., ChatGPT] could not possibly be fabricating cases on its own"); Order to Show Cause listing the six nonexistent decisions, https://storage.courtlistener.com/recap/gov.uscourts.nysd.575368/gov.uscourts.nysd.575368.31.0_1.pdf. All accessed October 10, 2026.

Back to contents

Chapter 9: Bias Written into the Record

1. US Equal Employment Opportunity Commission, "iTutorGroup to Pay $365,000 to Settle EEOC Discriminatory Hiring Suit," September 11, 2023, https://www.eeoc.gov/newsroom/itutorgroup-pay-365000-settle-eeoc-discriminatory-hiring-suit. Mobley v. Workday, Inc., No. 3:23-cv-00770-RFL (N.D. Cal.): order granting preliminary certification of a nationwide ADEA collective for applicants aged 40 and over, May 16, 2025 (Doc. 128), https://storage.courtlistener.com/recap/gov.uscourts.cand.408645/gov.uscourts.cand.408645.128.0.pdf; order denying certification for interlocutory appeal under 28 U.S.C. § 1292(b), July 2, 2026 (Doc. 376), https://storage.courtlistener.com/recap/gov.uscourts.cand.408645/gov.uscourts.cand.408645.376.0.pdf; plaintiffs' motion for class certification, September 14, 2026 (Doc. 418; hearing noticed for March 9, 2027), https://storage.courtlistener.com/recap/gov.uscourts.cand.408645/gov.uscourts.cand.408645.418.0.pdf. No ruling on that motion as of October 10, 2026. The claims are allegations; Workday disputes them. NYC Department of Consumer and Worker Protection, "Automated Employment Decision Tools" (enforcement began July 5, 2023), https://www.nyc.gov/site/dca/about/automated-employment-decision-tools.page. Illinois Artificial Intelligence Video Interview Act, 820 ILCS 42, Public Act 101-0260, https://ilga.gov/ftp/Public%20Acts/101/101-0260.htm. All accessed October 10, 2026.
2. Government of Ontario, "Your guide to the Employment Standards Act: Requirements related to publicly advertised job postings" (published February 4, 2026; updated February 24, 2026), https://www.ontario.ca/document/your-guide-employment-standards-act-0/requirements-related-publicly-advertised-job. Quebec, Act respecting the protection of personal information in the private sector, CQLR c P-39.1, s. 12.1 (in force September 22, 2023), https://www.legisquebec.gouv.qc.ca/en/document/cs/P-39.1. Section 12.1 applies only to decisions "based exclusively on an automated processing" of personal information; the person may "submit observations to a member of the personnel of the enterprise who is in a position to review the decision." Both accessed October 10, 2026.
3. Jeffrey Dastin, "Amazon scraps secret AI recruiting tool that showed bias against women," Reuters, October 10, 2018, https://www.reuters.com/article/world/amazon-scraps-secret-ai-recruiting-tool-that-showed-bias-against-women-idUSKCN1MK08J/; syndicated at CNBC, https://www.cnbc.com/2018/10/10/amazon-scraps-a-secret-ai-recruiting-tool-that-showed-bias-against-women.html (accessed October 10, 2026). Recruiters looked at the tool's recommendations but "never relied solely" on them.
4. Joy Buolamwini and Timnit Gebru, "Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification," in Proceedings of the 1st Conference on Fairness, Accountability and Transparency, Proceedings of Machine Learning Research 81 (2018): 77–91, https://proceedings.mlr.press/v81/buolamwini18a.html (accessed October 10, 2026). The two benchmarks were "overwhelmingly composed of lighter-skinned subjects (79.6% for IJB-A and 86.2% for Adience)"; for the commercial systems "there was no mention of what training data was used."
5. Inioluwa Deborah Raji and Joy Buolamwini, "Actionable Auditing: Investigating the Impact of Publicly Naming Biased Performance Results of Commercial AI Products," in Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society (AIES '19) (ACM, 2019), 429–435, published January 27, 2019, https://doi.org/10.1145/3306618.3314244 (accessed October 10, 2026). Error for darker-skinned females fell 17.7 to 30.4 percentage points; Amazon's and Kairos's errors for the same group were 31.37 percent and 22.50 percent.
6. Julia Angwin, Jeff Larson, Surya Mattu, and Lauren Kirchner, "Machine Bias," ProPublica, May 23, 2016, https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing; Jeff Larson, Surya Mattu, Lauren Kirchner, and Julia Angwin, "How We Analyzed the COMPAS Recidivism Algorithm," ProPublica, May 23, 2016 (n = 7,214; false-positive rates 44.85 percent Black and 23.45 percent white; false-negative rates 47.72 percent white and 27.99 percent Black), https://www.propublica.org/article/how-we-analyzed-the-compas-recidivism-algorithm. Both accessed October 10, 2026. Overall accuracy was 61 percent (63 percent for Black defendants, 59 percent for white).
7. Prabhakar Raghavan, "Gemini image generation got it wrong. We'll do better.," Google blog, February 23, 2024, https://blog.google/products-and-platforms/products/gemini/gemini-image-generation-issue/ (accessed October 10, 2026).
8. Parliamentary Inquiry Committee on Childcare Allowance, Unprecedented Injustice (English summary), December 17, 2020, https://www.houseofrepresentatives.nl/sites/default/files/atoms/files/verslag_pok_definitief-en-gb.docx.pdf (the committee left the nationality-based risk profiles outside its remit). Amnesty International, "Xenophobic Machines," October 25, 2021, https://www.amnesty.org/en/latest/news/2021/10/xenophobic-machines-dutch-child-benefit-scandal/. Rijksoverheid (Government of the Netherlands), May 30, 2022 (State Secretary Marnix van Rij on institutional racism within the Tax Administration and Benefits), https://www.rijksoverheid.nl/actueel/nieuws/2022/05/30/aandacht-voor-institutioneel-racisme-binnen-toezicht-belastingdienst-en-toeslagen. The Dutch Data Protection Authority's finding that processing (dual) nationality was unlawful and discriminatory dates from July 17, 2020. Autoriteit Persoonsgegevens (Dutch Data Protection Authority), news release, December 7, 2021 ("legt de Belastingdienst een boete op van 2,75 miljoen euro"; a €2.75 million fine), https://www.autoriteitpersoonsgegevens.nl/actueel/boete-belastingdienst-voor-discriminerende-en-onrechtmatige-werkwijze. All accessed October 10, 2026.
9. David Eggert, "Michigan reverses 44,000 jobless fraud cases, refunds $21M," Associated Press, August 11, 2017, https://apnews.com/michigan-reverses-44-000-jobless-fraud-cases-refunds-21m-dc3370d57e264448b67f75ceb63ad120. Michigan Department of Attorney General, settlement announcement, October 20, 2022, https://www.michigan.gov/ag/news/press-releases/2022/10/20/som-settlement-of-civil-rights-class-action-alleging-false-accusations-of-unemployment-fraud; "Class action settlement approved by Court of Claims," January 30, 2024, https://www.michigan.gov/ag/news/press-releases/2024/01/30/class-action-settlement-approved-by-court-of-claims (Bauserman v. Unemployment Insurance Agency). Sarah Cwiek, "State review: 93% of state unemployment fraud findings were wrong," Michigan Public, December 16, 2016 (the system "flagged 53,633 cases of fraud"; "a new state review of some of those cases found that the system was wrong—93% of the time"; the figure covers the reviewed cases, not all of them), https://www.michiganpublic.org/politics-government/2016-12-16/state-review-93-of-state-unemployment-fraud-findings-were-wrong. All accessed October 10, 2026.

Back to contents

Chapter 10: Whose Voice Built It?

1. US Census Bureau, American Community Survey 2024 1-year estimates, table GCT1602 (13.9 percent of people five and older speak Spanish at home), https://www2.census.gov/programs-surveys/acs/data/2024/1_year_geographic_comparison_tables/GCT1602.csv; count of about 44.9 million (excluding Puerto Rico) via USAFacts, citing the Census Bureau, https://usafacts.org/answers/how-many-people-speak-spanish-at-home/country/united-states/. Both accessed October 10, 2026. The 2025 1-year estimates had not been released as of October 10, 2026.
2. Common Crawl, "Statistics of Common Crawl Monthly Archives: Distribution of Languages," https://commoncrawl.github.io/cc-crawl-statistics/plots/languages; data file languages.csv, crawl CC-MAIN-2026-39 (English 41.86 percent, Russian 6.70, German 5.73, Japanese 5.17, Spanish 4.74, French 4.68, Inuktitut 0.0004; no language code for Navajo or Cree in any crawl in the file), downloaded October 10, 2026. Languages are identified by the CLD2 detector as the primary language of each HTML page, so "not tracked" partly reflects the detector's coverage.
3. Meta, Llama 2 technical report, arXiv:2307.09288 (2023), Table 10 (percent of documents: English 89.70, Spanish 0.13), https://arxiv.org/abs/2307.09288. DeepSeek-V3 technical report, arXiv:2412.19437 (14.8 trillion tokens; a multilingual corpus in which English and Chinese constitute the majority), https://arxiv.org/abs/2412.19437. Both accessed October 10, 2026.
4. Social Security Administration, "Fact Sheet: Social Security" ("Among Social Security beneficiaries age 65 and older, 39% of men and 44% of women receive 50% or more of their income from Social Security"; research released in 2024 using 2015 data), https://www.ssa.gov/news/press/factsheets/basicfact-alt.pdf (accessed October 10, 2026).
5. Allison Koenecke, Andrew Nam, Emily Lake, Joe Nudell, Minnie Quartey, Zion Mengesha, Connor Toups, John R. Rickford, Dan Jurafsky, and Sharad Goel, "Racial Disparities in Automated Speech Recognition," Proceedings of the National Academy of Sciences 117, no. 14 (April 7, 2020): 7684–7689 (systems "developed by Amazon, Apple, Google, IBM, and Microsoft"; "an aggregate WER of 0.35… for black speakers versus 0.19… for white speakers"), https://doi.org/10.1073/pnas.1915768117 (accessed October 10, 2026).
6. Pew Research Center, "Mobile Fact Sheet" (survey February 5 to June 18, 2025; n = 5,022), https://www.pewresearch.org/internet/fact-sheet/mobile/. Pew Research Center, "How opinions and use of AI differ by age," June 17, 2026 (survey February 17–23, 2026; ever used an AI chatbot: ages 65 and older, 12 percent in 2024 and 23 percent in 2026; ages 18–29, 66 percent in 2026), https://www.pewresearch.org/internet/2026/06/17/how-opinions-and-use-of-ai-differ-by-age/. Both accessed October 10, 2026. Pew notes that the question wording changed between the 2024 and 2026 surveys, so the trend is approximate.
7. Pew Research Center, "Racial and ethnic differences in how adults use and view AI," June 17, 2026 (survey of 5,119 US adults, February 17–23, 2026; "ever use AI chatbots like ChatGPT, Gemini or Copilot": Hispanic adults 49 percent, all adults 49 percent), https://www.pewresearch.org/internet/2026/06/17/racial-and-ethnic-differences-in-how-adults-use-and-view-ai/ (accessed October 10, 2026).
8. Statistics Canada, StatCan Plus, August 14, 2023 (Canadian Internet Use Survey, 2022), https://www.statcan.gc.ca/o1/en/plus/4288-canadian-seniors-more-connected-ever (accessed October 10, 2026). The 55.4 percent figure is a share of senior internet users, not of all seniors.
9. First Nations Information Governance Centre (FNIGC), OCAP® information and training, https://fnigc.ca/ocap-training/ (accessed October 10, 2026). OCAP® was established in 1998 during a meeting of the steering committee of the First Nations and Inuit Regional Longitudinal Health Survey; FNIGC describes it as "a set of specifically First Nations—not Indigenous—principles." OCAP® is a registered trademark of the First Nations Information Governance Centre (FNIGC).
10. Federal Trade Commission, press release on the proposed Everalbum settlement, January 11, 2021, https://www.ftc.gov/news-events/news/press-releases/2021/01/california-company-settles-ftc-allegations-it-deceived-consumers-about-use-facial-recognition-photo; "FTC Finalizes Settlement with Photo App Developer Related to Misuse of Facial Recognition Technology," May 7, 2021 (the order requires deletion of "the models and algorithms it developed by using the photos and videos uploaded by its users"), https://www.ftc.gov/news-events/news/press-releases/2021/05/ftc-finalizes-settlement-photo-app-developer-related-misuse-facial-recognition-technology. Federal Trade Commission, press release on WW International and Kurbo, March 4, 2022 (order approved by the court March 3, 2022), https://www.ftc.gov/news-events/news/press-releases/2022/03/ftc-takes-action-against-company-formerly-known-weight-watchers-illegally-collecting-kids-sensitive. All accessed October 10, 2026.
11. Office of the Privacy Commissioner of Canada, news release on the joint Clearview AI investigation, February 3, 2021, https://www.priv.gc.ca/en/opc-news/news-and-announcements/2021/nr-c_210203/ (accessed October 10, 2026).
12. Billy Perrigo, "Exclusive: OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic," TIME, January 18, 2023, https://time.com/6247678/openai-chatgpt-kenya-workers/ (accessed October 10, 2026). Work began in November 2021; Sama disputes several workers' accounts and says take-home pay was $1.46 to $3.74 an hour after taxes.

Back to contents

Chapter 11: What Do We Believe?

1. BMJ Group press release, July 18, 2024, on a feature by Chris Stokel-Walker in The BMJ (doi:10.1136/bmj.q1319), https://bmjgroup.com/trusted-tv-doctors-deepfaked-to-promote-health-scams-on-social-media/ (accessed October 10, 2026). The doctors named include Hilary Jones, Michael Mosley, and Rangan Chatterjee.
2. Robert Chesney and Danielle Keats Citron, "Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security," California Law Review 107 (December 2019): 1753, https://www.californialawreview.org/print/deep-fakes-a-looming-challenge-for-privacy-democracy-and-national-security (accessed October 10, 2026); an earlier version circulated on SSRN (no. 3213954) from 2018.
3. Meta AI, "Deepfake Detection Challenge Results," June 12, 2020, https://ai.meta.com/blog/deepfake-detection-challenge-results-an-open-initiative-to-advance-ai/ (accessed October 10, 2026). The challenge was announced September 5, 2019, launched on Kaggle December 11, 2019, and closed March 31, 2020. The top public-leaderboard score was 82.56 percent; the winning score on unseen ("black box") videos was 65.18 percent, by a model that ranked fourth on the public board. Meta reports both figures as accuracy; the competition itself was ranked by log loss. See also Dolhansky et al., arXiv:2006.07397.
4. Sam Wineburg and Sarah McGrew, "Lateral Reading and the Nature of Expertise: Reading Less and Learning More When Evaluating Digital Information," Teachers College Record 121, no. 11 (2019): 1–40, https://doi.org/10.1177/016146811912101102; open copy in the Stanford Digital Repository, https://stacks.stanford.edu/file/yk133ht8603/Wineburg%20McGrew_Lateral%20Reading%20and%20the%20Nature%20of%20Expertise.pdf; earlier working paper (Stanford History Education Group, 2017), https://www.hendrix.edu/uploadedFiles/Academics/Faculty_Resources/Teaching_and_Learning/EvaluatingDigitalInformation.pdf. Both accessed October 10, 2026. The smaller group was the American College of Pediatricians; "Fully 60% of students chose the College as the more reliable site"; "five of their colleagues [historians] equivocated"; "every fact checker unreservedly viewed the Academy's site as the more reliable."
5. Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 (Digital Omnibus on AI), Official Journal of the European Union L, July 24, 2026, art. 1(7), adding art. 5(1)(ba) to the AI Act, which bans "the placing on the market, the putting into service or the use of an AI system that generates or manipulates realistic images, videos, audio or similar material of an identifiable natural person's intimate parts, or of an identifiable natural person engaged in sexually explicit activities, without that person's freely-given, specific, informed, unambiguous and explicit consent…"; art. 5(1)(bb) adds a ban on AI-generated child sexual abuse material; art. 5(1a) limits the provider ban to tools built for this purpose or that foreseeably produce it without reasonable safeguards. Under amended art. 113, points (ba) and (bb) apply from December 2, 2026. European Commission, AI Act Service Desk, "Article 5" (consolidated text as at July 27, 2026: "From 2 December 2026, two additional prohibitions apply"), https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-5. EUR-Lex, https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=celex%3A32026R1744. Both accessed October 10, 2026.
6. Kat Tenbarge, NBC News, March 8, 2024, https://www.nbcnews.com/tech/tech-news/beverly-hills-school-expels-students-deepfake-nude-photos-rcna142480. TAKE IT DOWN Act, S. 146, Public Law 119-12 (May 19, 2025), https://www.congress.gov/bill/119th-congress/senate-bill/146. Federal Trade Commission, "FTC Begins Enforcing the TAKE IT DOWN Act," May 19, 2026, https://www.ftc.gov/news-events/news/press-releases/2026/05/ftc-begins-enforcing-take-it-down-act. National Center for Missing & Exploited Children, Take It Down FAQ, https://takeitdown.ncmec.org/faq/ (the tool does not work on encrypted platforms; threats and sextortion go to the CyberTipline, 1-800-THE-LOST). All accessed October 10, 2026.
7. Department of Justice Canada, Bill C-16 (Protecting Victims Act) backgrounder, https://www.justice.gc.ca/eng/csj-sjc/pl/c16/index.html; Royal Assent text (June 18, 2026), https://www.parl.ca/DocumentViewer/en/45-1/bill/C-16/royal-assent. Most of the new provisions took effect July 18, 2026. Cybertip.ca (Canadian Centre for Child Protection), https://www.cybertip.ca/en/; NeedHelpNow.ca, https://needhelpnow.ca/en/ (for people "worried a nude or abusive image of you under the age of 18 is being shared online"). All accessed October 10, 2026.
8. Artificial Intelligence and Human Rights: Hearing Before the Subcommittee on Human Rights and the Law of the Committee on the Judiciary, United States Senate, 118th Cong. (June 13, 2023), S. Hrg. 118-40, Serial No. J-118-21, testimony of Jennifer DeStefano, https://www.govinfo.gov/content/pkg/CHRG-118shrg52709/html/CHRG-118shrg52709.htm (accessed October 10, 2026).
9. FBI Internet Crime Complaint Center, Public Service Announcement I-120324-PSA, December 3, 2024 ("Create a secret word or phrase with your family"), https://www.ic3.gov/PSA/2024/PSA241203 (accessed October 10, 2026).
10. Canadian Anti-Fraud Centre, "Emergency" scams, https://antifraudcentre-centreantifraude.ca/scams-fraudes/emergency-urgence-eng.htm (accessed October 10, 2026). The CAFC page does not mention AI voice cloning by name.
11. Kathleen Magramo, CNN, May 16–17, 2024, https://www.cnn.com/2024/05/16/tech/arup-deepfake-scam-loss-hong-kong-intl-hnk (HK$200 million, about US$25.6 million, in fifteen transactions; Arup confirmed that "fake voices and images were used"), accessed October 10, 2026. Hong Kong police first described the case in February 2024 (CNN, February 4, 2024). The employee initially suspected a phishing email.

Back to contents

Chapter 12: Whose Work Is It?

1. Complaint, The New York Times Co. v. Microsoft Corp. et al., No. 1:23-cv-11195 (S.D.N.Y. filed December 27, 2023) (the models "were built by copying and using millions of The Times's copyrighted news articles"; examples of near-verbatim output in Exhibit J), https://nytco-assets.nytimes.com/2023/12/NYT_Complaint_Dec2023.pdf. The case is now coordinated in In re OpenAI, Inc., Copyright Infringement Litigation, No. 25-md-3143 (SHS) (OTW) (S.D.N.Y.), before Judge Sidney H. Stein. An opinion and order of August 6, 2026 (No. 1:23-cv-11195, Doc. 1449; entered on the MDL docket as Doc. 1662) dismissed with prejudice, by agreement after Cox Communications, Inc. v. Sony Music Entertainment, 146 S. Ct. 959 (2026), the news plaintiffs' "material contribution" contributory-infringement claims against OpenAI and Microsoft premised on end-user infringement, and dismissed the Times's trademark-dilution claim; the direct-infringement and fair-use questions remain open, with summary-judgment briefing under way as of October 2026, https://www.courtlistener.com/docket/68117049/1449/the-new-york-times-company-v-microsoft-corporation/; MDL entry, https://www.courtlistener.com/docket/69879510/1662/in-re-openai-inc-copyright-infringement-litigation/. Andersen v. Stability AI Ltd., No. 3:23-cv-00201-WHO (N.D. Cal. filed January 13, 2023) (defendants Stability AI, Midjourney, DeviantArt, and Runway); stipulated scheduling order, July 1, 2026 (Doc. 623), https://docs.justia.com/cases/federal/district-courts/california/candce/3%3A2023cv00201/407208/623; the case was still in discovery in October 2026, https://www.courtlistener.com/docket/66732129/andersen-v-stability-ai-ltd/. Bartz v. Anthropic PBC, No. 4:24-cv-05417-AMO (N.D. Cal.) (filed as No. 3:24-cv-05417-WHA before Judge William Alsup; reassigned to Judge Araceli Martínez-Olguín on December 31, 2025, Doc. 532), order granting final approval of class settlement, July 20, 2026 (Doc. 680) ("the non-reversionary Settlement Fund of $1.5 billion"; estimated payment "about $3,000, less costs and fees" per work; 482,460 works on the Works List; class members "retain their rights to sue for future misconduct"), https://www.courtlistener.com/docket/69058235/680/bartz-v-anthropic-pbc/. Appeals filed after the order concern only the award of attorneys' fees, and the parties have reported that the settlement's effective date has passed. All accessed October 10, 2026. The claims in the pending cases are allegations.
2. International Committee of Medical Journal Editors (ICMJE), "AI Use by Authors," https://www.icmje.org/recommendations/browse/artificial-intelligence/ai-use-by-authors.html (accessed October 10, 2026).
3. ICMJE (note 2): chatbots "should not be listed as authors because they cannot be responsible for the accuracy, integrity, and originality of the work."
4. Amazon Kindle Direct Publishing, "Content Guidelines," section "Artificial intelligence (AI) content (text, images, or translations)," https://kdp.amazon.com/en_US/help/topic/G200672390 (page undated; accessed October 10, 2026).
5. College Board, "Artificial Intelligence Tools" (AP Capstone policy; page undated), https://apcentral.collegeboard.org/exam-administration-ordering-scores/administering-exams/exam-policies/artificial-intelligence-tools (generative AI "must be used ethically, responsibly, and intentionally to support student learning, not to bypass it"; performance tasks "must be the student's own work"; "Failure to complete the checkpoints will result in a score of 0 on the associated task"). Delaware Department of Education, "Allowable Use Statements for Generative AI" (undated), https://education.delaware.gov/educators/academic-support/standards-and-instruction/digital-de/instructional-resources/ai-in-delaware-education/allowable-use-statements-for-generative-ai/ (students "crediting Gen AI contributions appropriately"). Both accessed October 10, 2026. Policies change; recheck before publication.
6. British Columbia Ministry of Education and Child Care, "Digital literacy and the use of AI in education: supports for British Columbia schools" (last updated July 18, 2024), https://www2.gov.bc.ca/gov/content/education-training/k-12/administration/program-management/ai-in-education (accessed October 10, 2026).
7. US Copyright Office Review Board, "Re: Second Request for Reconsideration for Refusal to Register Théâtre D'opéra Spatial," September 5, 2023 ("While Mr. Allen did not disclose in his application that the Work was created using an AI system…"; at least 624 prompts; "one 'acceptable' panel out of four"; "Because Mr. Allen is unwilling to disclaim the AI-generated material, the Work cannot be registered as submitted"), https://copyright.gov/rulings-filings/review-board/docs/Theatre-Dopera-Spatial.pdf. Allen has challenged the refusal in Allen v. Perlmutter, No. 1:24-cv-02665 (D. Colo. filed September 26, 2024); summary-judgment briefing closed in February 2026, and no ruling had appeared on the docket as of October 10, 2026, https://www.courtlistener.com/docket/69198079/allen-v-perlmutter/. US Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability (January 2025) ("given current generally available technology, prompts alone do not provide sufficient human control to make users of an AI system the authors of the output"), https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf. Thaler v. Perlmutter, No. 23-5233 (D.C. Cir. March 18, 2025), cert. denied, No. 25-449 (U.S. March 2, 2026), https://www.supremecourt.gov/docket/docketfiles/html/public/25-449.html. The contest entry's "via Midjourney" credit was reported by Kevin Roose, "An A.I.-Generated Picture Won an Art Prize. Artists Aren't Happy," New York Times, September 2, 2022, https://www.nytimes.com/2022/09/02/technology/ai-artificial-intelligence-artists.html. All accessed October 10, 2026.

Back to contents

Chapter 13: The Thinking Muscle

1. Evan F. Risko and Sam J. Gilbert, "Cognitive Offloading," Trends in Cognitive Sciences 20, no. 9 (2016): 676–688, https://doi.org/10.1016/j.tics.2016.07.002; author PDF, https://samgilbert.net/pubs/Risko2016TiCS.pdf (accessed October 10, 2026).
2. Michael Gerlich (SBS Swiss Business School), "AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking," Societies 15, no. 1 (2025): 6, published January 3, 2025, https://doi.org/10.3390/soc15010006 ("a significant negative correlation between frequent AI tool usage and critical thinking abilities, mediated by increased cognitive offloading"); correction, Societies 15, no. 9 (2025): 252, published September 10, 2025, https://doi.org/10.3390/soc15090252. Accessed October 10, 2026. The correction replaced Table 4, which "was a duplicate of Table 3," with the corrected table; "the author states that the scientific conclusions are unaffected."
3. Hao-Ping (Hank) Lee, Advait Sarkar, Lev Tankelevitch, Ian Drosos, Sean Rintel, Richard Banks, and Nicholas Wilson, "The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers," Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI '25), https://doi.org/10.1145/3706598.3713778; author PDF, https://advait.org/files/lee_2025_ai_critical_thinking_survey.pdf (accessed October 10, 2026). "In the majority of examples, knowledge workers perceive decreased effort for cognitive activities associated with critical thinking when using GenAI"; "higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking."
4. Leon Reicherts, Zelun Tony Zhang, Elisabeth von Oswald, Yuanting Liu, Yvonne Rogers, and Mariam Hassib, "AI, Help Me Think—but for Myself: Assisting People in Complex Decision-Making by Providing Different Kinds of Cognitive Support," Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI '25), https://doi.org/10.1145/3706598.3713295; author version, arXiv:2504.06771, https://arxiv.org/abs/2504.06771 (accessed October 10, 2026). Affiliations include Microsoft Research, University College London, fortiss, LMU Munich, and Politecnico di Milano. N = 21; ExtendAI integrated "better into the decision-making process and people's own thinking," leading to "slightly better outcomes"; RecommendAI "was able to provide more novel insights while requiring less cognitive effort."

Back to contents

Chapter 14: The Lines to Hold

1. OpenAI, "Terms of Use" (effective January 1, 2026) ("We may use Content to provide, maintain, develop, and improve our Services…"; "If you do not want us to use your Content to train our models, you can opt out…"), https://openai.com/policies/terms-of-use/. Anthropic, "Consumer Terms of Service" (effective October 8, 2025) (materials may be used for "training our models, unless you opt out of training through your account settings," with exceptions for feedback and material flagged for safety review), https://www.anthropic.com/legal/consumer-terms. Both accessed October 10, 2026. Policies change; check the current versions.
2. Hana Habib et al., "An Empirical Analysis of Data Deletion and Opt-Out Choices on 150 Websites," Fifteenth Symposium on Usable Privacy and Security (SOUPS 2019) (USENIX, 2019), 387–406 ("even though the majority of analyzed websites offered privacy choices, they were located inconsistently across websites"; some "were rendered unusable by missing or unhelpful information"), https://www.usenix.org/conference/soups2019/presentation/habib (accessed October 10, 2026). Presented again at the FTC's PrivacyCon 2020. The study covers websites in general, not AI services.
3. Pew Research Center, "Key findings about Americans and data privacy," October 18, 2023 (survey of 5,101 US adults, May 15–21, 2023), https://www.pewresearch.org/short-reads/2023/10/18/key-findings-about-americans-and-data-privacy/ (accessed October 10, 2026).
4. Pew Research Center, Americans and AI 2026: Chatbots, Smart Devices and Views on Impact, June 17, 2026; topline, American Trends Panel Wave 187, February 17–23, 2026, N = 5,119 US adults, question AIDSEC (more secure 3 percent, less secure 71 percent, not much difference 10 percent, not sure 16 percent), https://www.pewresearch.org/wp-content/uploads/sites/20/2026/06/PI_2026.06.17_Americans-and-AI_TOPLINE.pdf (accessed October 10, 2026). This is a prediction asked of all US adults; it isn't comparable to the 2023 questions in note 3.
5. Office of the Privacy Commissioner of Canada, "Principles for responsible, trustworthy and privacy-protective generative AI technologies," December 7, 2023 (developers and providers should inform users of "any secondary purpose, such as where personal information collected from prompts is used for further training or refining of an AI model"; users should be "informed about both privacy risks and any mitigations available to them (such as not entering personal information into a prompt, unless necessary)"), https://www.priv.gc.ca/en/privacy-topics/technology/artificial-intelligence/gd_principles_ai/ (accessed October 10, 2026). The principles are addressed to organizations, not users.
6. Federal Trade Commission, "Children's Online Privacy Protection Rule," 90 Fed. Reg. 16918 (April 22, 2025) ("The amended Rule is effective June 23, 2025… regulated entities have until April 22, 2026 to comply," with limited exceptions), https://www.federalregister.gov/documents/2025/04/22/2025-05904/childrens-online-privacy-protection-rule. OpenAI, "Terms of Use" (effective January 1, 2026) ("You must be at least 13 years old or the minimum age required in your country… If you are under 18 you must have your parent or legal guardian's permission"). Anthropic, "Consumer Terms of Service" (effective October 8, 2025) ("You must be at least 18 years old or the minimum age required to consent to use the Services in your location, whichever is higher"). Google for Families Help, "Guide your child's Gemini Apps experience" ("A parent must enable access before their child under 13 (or the applicable age in your country) can use Gemini Apps with a supervised account"), https://support.google.com/families/answer/16109150. Federal Trade Commission, "FTC Launches Inquiry into AI Chatbots Acting as Companions," September 11, 2025 (orders to Alphabet, Inc.; Character Technologies, Inc.; Instagram, LLC; Meta Platforms, Inc.; OpenAI OpCo, LLC; Snap, Inc.; and X.AI Corp. seeking "information on how these firms measure, test, and monitor potentially negative impacts of this technology on children and teens," including how they "monetize user engagement" and "use or share personal information obtained through users' conversations"), https://www.ftc.gov/news-events/news/press-releases/2025/09/ftc-launches-inquiry-ai-chatbots-acting-companions; FTC study page, https://www.ftc.gov/reports/6b-orders-file-special-report-regarding-advertising-safety-data-handling-practices-companies (no report posted as of October 10, 2026). Section 6(b) orders are a study tool, not an enforcement action. All accessed October 10, 2026. Status as of October 2026.
7. OpenAI, "March 20 ChatGPT outage: Here's what happened," March 24, 2023 ("1.2% of the ChatGPT Plus subscribers who were active during a specific nine-hour window"; "Full credit card numbers were not exposed at any time"; the bug was in the open-source Redis client library, redis-py), https://openai.com/index/march-20-chatgpt-outage/. Office of the Privacy Commissioner of Canada, news release, May 6, 2026 (the four regulators "ultimately concluded that the way that Open AI had initially trained Chat GPT was not compliant with their respective privacy laws"; issues included "overcollection of personal information; lack of valid consent and transparency"; "Open AI has significantly limited the personal information and sensitive information that is used to train new Chat GPT models"; "well-founded and conditionally resolved"), https://www.priv.gc.ca/en/opc-news/news-and-announcements/2026/nr-c_260506; PIPEDA Findings #2026-002, https://www.priv.gc.ca/en/opc-actions-and-decisions/investigations/investigations-into-businesses/2026/pipeda-2026-002. All accessed October 10, 2026. The joint investigation was announced May 25, 2023. The conclusions reached by each office varied; the "well-founded and conditionally resolved" finding is the federal commissioner's.
8. Bloomberg, "Samsung Bans ChatGPT, Google Bard, Other Generative AI Use by Staff After Leak," May 2, 2023, https://www.bloomberg.com/news/articles/2023-05-02/samsung-bans-chatgpt-and-other-generative-ai-use-by-staff-after-leak (paywalled; also reported by TechCrunch and The Verge the same day). The 2023 details ("meeting minutes and source codes" uploaded by engineers in the semiconductor division): Jo He-rim, "Samsung bans AI chatbots after leak," Korea Herald, May 3, 2023, https://www.koreaherald.com/article/3118116. Yonhap News Agency (Korean), "삼성전자, 내달 중 챗GPT 등 생성형 AI 사내 활용 허용" [Samsung to allow in-house use of ChatGPT and other generative AI next month], May 26, 2026, https://www.yna.co.kr/view/AKR20260526057700003 (external generative AI services for staff in the DX [device] division from June 2026, only for employees who complete security training; the DS [semiconductor] division keeps use limited to internally approved cases). All accessed October 10, 2026.

Back to contents

Chapter 15: One Sheet of Paper

1. 988 Suicide & Crisis Lifeline (US) ("Call, text, or chat… 24/7/365"), https://988lifeline.org/. 9-8-8 Suicide Crisis Helpline (Canada) ("Call 9-8-8… Text 9-8-8… 24 hours a day, every day of the year… If your safety is at risk, call 9-1-1 right away"), https://988.ca/. Both accessed October 10, 2026. Numbers checked October 2026; services change, so confirm on the official sites. More resources are in Appendix C.

Back to contents

Chapter 16: A Mature Partnership

1. Office of Governor Gavin Newsom, "Governor Newsom signs SB 53, advancing California's world-leading artificial intelligence industry," September 29, 2025 (the Transparency in Frontier Artificial Intelligence Act requires large frontier developers to "publicly publish a framework" on their websites, creates "a new mechanism for frontier AI companies and the public to report potential critical safety incidents to California's Office of Emergency Services," and "protects whistleblowers"), https://www.gov.ca.gov/2025/09/29/governor-newsom-signs-sb-53-advancing-californias-world-leading-artificial-intelligence-industry/ (accessed October 10, 2026). Bill history: approved by the governor and chaptered by the Secretary of State September 29, 2025, Stats. 2025, ch. 138, https://leginfo.legislature.ca.gov/faces/billStatusClient.xhtml?bill_id=202520260SB53. The act is codified at Cal. Bus. & Prof. Code §§ 22757.10–22757.16 (added by Stats. 2025, ch. 138, effective January 1, 2026): large frontier developers must publish a "frontier AI framework" (§ 22757.12(a)); frontier developers must report critical safety incidents to the Office of Emergency Services within fifteen days (§ 22757.13(c)). California Governor's Office of Emergency Services, "Transparency in Frontier AI Act Reporting" (reporting portal for "critical safety incidents" and catastrophic-risk assessment summaries), https://www.caloes.ca.gov/office-of-the-director/operations/homeland-security/california-cybersecurity-integration-center/transparency-in-frontier-ai-act-reporting/ (accessed October 10, 2026). Status as of October 2026.
2. Colorado General Assembly, SB26-189, "Automated Decision-Making Technology," signed May 14, 2026 (Session Laws 2026, ch. 131) (the act "repeals and reenacts" the 2024 law, SB 24-205, "with new requirements regarding the use of automated decision-making technology in consequential decisions"; the act takes effect January 1, 2027, and applies to consequential decisions made on or after that date (Session Laws 2026, ch. 131, sec. 5, https://www.leg.colorado.gov/laws/session-laws/SB26-189/131/download); notice "at the point of interaction"; a plain-language description within thirty days after an adverse outcome; the right to correct data and, on request after an adverse outcome, "an opportunity for meaningful human review and reconsideration of the consequential decision, to the extent commercially reasonable" (C.R.S. § 6-1-1705(1)(a)(II), as reenacted by sec. 1)), https://www.leg.colorado.gov/bills/sb26-189 (accessed October 10, 2026). X.AI LLC v. Weiser, No. 1:26-cv-01515 (D. Colo.), Minute Order, ECF No. 24 (April 27, 2026) (entered by Magistrate Judge Cyrus Y. Chung, granting the parties' joint motion; by stipulation, no enforcement of SB 24-205 "or any legislation replacing or amending" it until fourteen days after a ruling on xAI's forthcoming preliminary-injunction motion; the United States is plaintiff-intervenor), https://storage.courtlistener.com/recap/gov.uscourts.cod.253513/gov.uscourts.cod.253513.24.0.pdf. Through the docket's latest entry (ECF No. 25, October 2, 2026, an attorney's withdrawal), no preliminary-injunction motion or ruling appears, so under the order's terms the pause on enforcement continues. CourtListener docket, https://www.courtlistener.com/docket/73171074/x-ai-llc-v-weiser/ (checked October 10, 2026). Status as of October 10, 2026.
3. National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 26, 2023 (voluntary), https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10. Executive Order 14365, "Ensuring a National Policy Framework for Artificial Intelligence," signed December 11, 2025, 90 Fed. Reg. 58499 (December 16, 2025), FR Doc. 2025-23092 (sec. 3: an "AI Litigation Task Force… whose sole responsibility shall be to challenge State AI laws inconsistent with the policy set forth in section 2"), https://www.federalregister.gov/documents/2025/12/16/2025-23092/ensuring-a-national-policy-framework-for-artificial-intelligence. Both accessed October 10, 2026. An executive order does not by itself repeal or preempt state law. As of October 10, 2026, Congress had not enacted a statute preempting state AI laws: the Senate struck the proposed state-AI moratorium from H.R. 1 by a 99–1 vote (Roll Call Vote No. 363, 119th Cong., 1st sess., July 1, 2025, on S.Amdt. 2814), https://www.senate.gov/legislative/LIS/roll_call_votes/vote1191/vote_119_1_00363.htm, and H.R. 5388, the American Artificial Intelligence Leadership and Uniformity Act (introduced September 16, 2025, proposing a five-year moratorium), was still in committee (latest action December 19, 2025), https://www.congress.gov/bill/119th-congress/house-bill/5388 (both accessed October 10, 2026).
4. Parliament of Canada, LEGISinfo, Bill C-27 (44th Parliament, 1st Session, November 22, 2021, to January 6, 2025), which included the Artificial Intelligence and Data Act; the bill died on the Order Paper when Parliament was prorogued on January 6, 2025, https://www.parl.ca/legisinfo/en/bill/44-1/c-27. Its privacy provisions largely returned in Bill C-36, the Protecting Privacy and Consumer Data Act (45th Parliament, 1st Session; first reading June 15, 2026; "at second reading in the House of Commons" as of October 10, 2026), which does not enact the AI and Data Act, https://www.parl.ca/legisinfo/en/bill/45-1/c-36; Library of Parliament, Legislative Summary of Bill C-36 (preliminary version), https://bdp.parl.ca/staticfiles/PublicWebsite/Home/ResearchPublications/LegislativeSummaries/PDF/45-1/PV_45-1-C36-E.pdf. Quebec, Act respecting the protection of personal information in the private sector, CQLR c P-39.1, s. 12.1, https://www.legisquebec.gouv.qc.ca/en/document/cs/P-39.1 (see Chapter 9, note 2). All accessed October 10, 2026.
5. UNESCO, Recommendation on the Ethics of Artificial Intelligence, adopted November 23, 2021, paras. 36 and 38, https://www.unesco.org/en/legal-affairs/recommendation-ethics-artificial-intelligence (accessed October 10, 2026). Paragraph 38: individuals "should be able to access the reasons for a decision affecting their rights and freedoms, and have the option of making submissions to a designated staff member of the private sector company or public sector institution able to review and correct the decision."
6. Regulation (EU) 2024/1689 (Artificial Intelligence Act), June 13, 2024, in force August 1, 2024. Regulation (EU) 2026/1744 (Digital Omnibus on AI), OJ L, July 24, 2026, in force July 27, 2026, recital 40 (high-risk obligations moved to December 2, 2027, for uses listed in Annex III, such as hiring and credit, and to August 2, 2028, for AI in products covered by Annex I); art. 1(7), adding the "nudifier" ban as art. 5(1)(ba) of the AI Act (see Chapter 11, note 5), https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=celex%3A32026R1744 (accessed October 10, 2026).
7. OpenAI, "Helping people when they need it most," August 26, 2025 ("these safeguards can sometimes be less reliable in long interactions: as the back-and-forth grows, parts of the model's safety training may degrade"), https://openai.com/index/helping-people-when-they-need-it-most/ (accessed October 10, 2026). See also Chapter 7.
8. Federal Trade Commission, ReportFraud.ftc.gov ("Report a scam, a company, or an unwanted call… Your report is shared with over 2,000 law enforcers"), https://reportfraud.ftc.gov/ (accessed October 10, 2026).
9. Office of the Privacy Commissioner of Canada, "Report a concern," https://www.priv.gc.ca/en/report-a-concern/ (accessed October 10, 2026).
10. OpenAI, "Terms of Use," and Anthropic, "Consumer Terms of Service" (see Chapter 14, note 1); Habib et al., "An Empirical Analysis of Data Deletion and Opt-Out Choices on 150 Websites," SOUPS 2019 (see Chapter 14, note 2). The Habib study covers websites in general, not AI services.
11. UNESCO, Recommendation on the Ethics of Artificial Intelligence (2021), para. 73 (individuals should enjoy "the ability to access and erase their personal data in AI systems, except for certain circumstances in compliance with international law"; the paragraph is addressed to member states), https://www.unesco.org/en/legal-affairs/recommendation-ethics-artificial-intelligence (accessed October 10, 2026).
12. Office of the Privacy Commissioner of Canada, news release, May 6, 2026 (issues included "individuals' ability to access, correct and delete their personal information"), https://www.priv.gc.ca/en/opc-news/news-and-announcements/2026/nr-c_260506 (accessed October 10, 2026). See Chapter 14, note 7.
13. UNESCO, Recommendation on the Ethics of Artificial Intelligence (2021), para. 51 (AI systems "identified as potential risks to human rights should be broadly tested by AI actors, including in real-world conditions if needed, as part of the Ethical Impact Assessment, before releasing them in the market"). See note 11.

Back to contents

Epilogue: What Happened Next

This section has no source notes.

Back to contents

Appendices

1. Sources for Appendix C, all accessed October 10, 2026. 988 Suicide & Crisis Lifeline, home page ("Call, text, or chat… The 988 Lifeline is available 24/7/365. Your conversations are free and confidential"), https://988lifeline.org/. 9-8-8: Suicide Crisis Helpline, "Get Help" ("Call 9-8-8… Text 9-8-8… 24/7/365 Crisis Support"), https://988.ca/. Crisis Text Line, home page ("Text HOME to 741741 from anywhere in the United States—24/7, free, confidential"), https://www.crisistextline.org/. Kids Help Phone, "Texting FAQs" ("By texting CONNECT to 686868"), https://kidshelpphone.ca/text-faq/; Kids Help Phone, "Our e-mental health services" ("Adults can text 741741 to connect with a trained, volunteer crisis responder"; "calling 1-800-668-6868"; "Our phone service is also offered in over 100 languages… with the help of trained interpreters"; 9-8-8 support "in English or French, 24/7"), https://kidshelpphone.ca/our-e-mental-health-services/. The Trevor Project, "Get Help" ("Call us at 1-866-488-7386"; "use TTY to call 1-866-803-3699"; "Text 'START' to 678-678"; "24/7/365, nationwide, 100% free & confidential"), https://www.thetrevorproject.org/get-help/. National Center for Missing & Exploited Children, "CyberTip Report" ("1-800-THE-LOST (1-800-843-5678)… 24 hours a day, 7 days a week"), https://report.cybertip.org/. NCMEC, Take It Down, "Frequently Asked Questions" ("If you were under age 18 when images or videos of you nude, partially nude, or in a sexually explicit situation were taken"), https://takeitdown.ncmec.org/faq/. Canadian Centre for Child Protection, Cybertip.ca ("Canada's national tipline for reporting the online sexual exploitation of children"), https://www.cybertip.ca/en/; NeedHelpNow.ca ("worried a nude or abusive image of you under the age of 18 is being shared online"), https://needhelpnow.ca/en/. Federal Trade Commission, ReportFraud.ftc.gov, https://reportfraud.ftc.gov/; FTC Consumer Advice, "What To Know About Identity Theft" ("To report in English, go to IdentityTheft.gov… To report in Spanish, go to RobodeIdentidad.gov… call 877-438-4338 and press 3… Interpreters are available from 9:00am–5:00pm ET"; "a free personal recovery plan"), https://consumer.ftc.gov/articles/what-know-about-identity-theft; FTC, "FTC Begins Enforcing the TAKE IT DOWN Act," May 19, 2026 (TakeItDown.ftc.gov lets victims "submit complaints about platforms that have failed to act on valid requests for the removal of nonconsensual intimate images" and about platforms "that have failed to create a process" for such requests; covered platforms must remove the images "within 48 hours of a valid request"), https://www.ftc.gov/news-events/news/press-releases/2026/05/ftc-begins-enforcing-take-it-down-act; TakeItDown.ftc.gov ("Report Platform Violations"), https://takeitdown.ftc.gov/. Internet Crime Complaint Center (IC3), home page ("run by the FBI"; "Crimes against children should be filed with the National Center for Missing and Exploited Children"), https://www.ic3.gov/. Canadian Anti-Fraud Centre, "Report fraud and cybercrime" ("you can report a fraud or cybercrime online at reportcyberandfraud.canada.ca"; "Toll free: 1-888-495-8501… Monday to Friday, from 10 am to 4:45 pm (Eastern time) and closed on holidays"), https://antifraudcentre-centreantifraude.ca/report-signalez-eng.htm; RCMP and Canadian Anti-Fraud Centre, "Report Cybercrime and Fraud," https://reportcyberandfraud.canada.ca/. Office of the Privacy Commissioner of Canada, "Report a concern" ("Always try to resolve the issue with the organization you had a problem with before reporting a concern"), https://www.priv.gc.ca/en/report-a-concern/.

Back to contents