Opening Edition
Notes and references
The company, its people, supplier demonstration, internal records and numerical results are fictional. Case figures illustrate decision-making and are not reported research findings. The technical claims below are supported by external sources. The remaining chapters and their references will be added in later editions.
Numbered notes
1. Generated language can be fluent yet contain hallucinated or unsupported content. Ji et al., abstract and survey overview. The fictional wrong-rule example is the author’s case, not a result reported in that paper. Reference R19.
2. Generative models produce synthetic content; language models generate successive tokens and can produce false or inconsistent output. Generative AI Profile, Section 2.2 and its introductory definition. Forecasting demand versus composing a reply is the author’s practical company example. Reference R03.
3. Decoding and sampling choices affect generated text and its diversity. Holtzman et al., abstract. The chapter’s saved-run and version-recording advice is the author’s method. Reference R20.
4. Retrieval-augmented generation combines retrieved external material with generation. Lewis et al., abstract. The company’s source-date requirements are the author’s design judgement, not a claim that the paper validates this service. Reference R21.
5. AI assistance can improve some tasks and worsen others within a knowledge workflow. Dell’Acqua et al., journal abstract. The study does not establish this fictional product’s accuracy or the company’s expected benefit. Reference R14.
6. Classifier confidence estimates can be poorly calibrated. Guo et al., abstract. The definition of tested decision score and the 0.8 illustration are explanatory method, not a guarantee or a result from the case. Reference R15.
7. Language-model systems can combine reasoning and action steps and interface with external tools or environments. Yao et al., abstract. The proposed stopping and recovery boundaries are the author’s operating recommendations. Reference R22.
Reference records
R19
Ziwei Ji et al. Survey of Hallucination in Natural Language Generation. ACM Computing Surveys, DOI 10.1145/3571730; author version arXiv:2202.03629v7, 2024.
Type: Paper
Live source: https://arxiv.org/abs/2202.03629
Claim supported: Chapter 3, note 1: fluency does not establish factual or source faithfulness.
Verified on: 24 September 2026. Supporting claim checked against the live source; all remain subject to final publication verification.
Archive: http://web.archive.org/web/20260909044219/https://arxiv.org/abs/2202.03629
Archive status: existing snapshot, not a new capture. Recapture needed; snapshot contents not independently rechecked in this round.
Status: Verify before publish. Recheck within 30 days of publication.
R03
NIST. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1, July 2024. DOI 10.6028/NIST.AI.600-1.
Type: Government guidance
Live source: https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf
Claim supported: Chapter 3, note 2: generated content and token prediction; confident erroneous output (Section 2.2).
Verified on: 24 September 2026. Supporting claim checked against the live source; all remain subject to final publication verification.
Archive: http://web.archive.org/web/20260921040857/https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf
Archive status: existing snapshot, not a new capture. Recapture needed; snapshot contents not independently rechecked in this round.
Status: Verify before publish. Recheck within 30 days of publication.
R20
Ari Holtzman et al. The Curious Case of Neural Text Degeneration. ICLR 2020; arXiv:1904.09751v2.
Type: Paper
Live source: https://arxiv.org/abs/1904.09751
Claim supported: Chapter 3, note 3: decoding and sampling choices affect generated text and diversity.
Verified on: 24 September 2026. Supporting claim checked against the live source; all remain subject to final publication verification.
Archive: http://web.archive.org/web/20260917122236/https://arxiv.org/abs/1904.09751
Archive status: existing snapshot, not a new capture. Recapture needed; snapshot contents not independently rechecked in this round.
Status: Verify before publish. Recheck within 30 days of publication.
R21
Patrick Lewis et al. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 2020; author version arXiv:2005.11401v4, 2021.
Type: Paper
Live source: https://arxiv.org/abs/2005.11401
Claim supported: Chapter 3, note 4 and FIG-3.1: external retrieval combined with generation.
Verified on: 24 September 2026. Supporting claim checked against the live source; all remain subject to final publication verification.
Archive: http://web.archive.org/web/20260915175627/https://arxiv.org/abs/2005.11401
Archive status: existing snapshot, not a new capture. Recapture needed; snapshot contents not independently rechecked in this round.
Status: Verify before publish. Recheck within 30 days of publication.
R14
Fabrizio Dell’Acqua et al. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science, published online 11 March 2026. DOI 10.1287/orsc.2025.21838.
Type: Paper
Live source: https://pubsonline.informs.org/doi/abs/10.1287/orsc.2025.21838
Claim supported: Chapter 3, note 5: uneven effects across tasks. Replaces the starter working-paper version for this claim.
Verified on: 24 September 2026. Supporting claim checked against the live source; all remain subject to final publication verification.
Archive: capture attempt failed; existing snapshot could not be established. Recapture needed before publication.
Status: Verify before publish. Recheck within 30 days of publication.
R15
Chuan Guo, Geoff Pleiss, Yu Sun and Kilian Q. Weinberger. On Calibration of Modern Neural Networks. ICML 2017, PMLR 70, pp. 1321–1330.
Type: Paper
Live source: https://proceedings.mlr.press/v70/guo17a.html
Claim supported: Chapter 3, note 6 and FIG-3.2: numerical classifier confidence can be poorly calibrated; not a study of verbal certainty.
Verified on: 24 September 2026. Supporting claim checked against the live source; all remain subject to final publication verification.
Archive: http://web.archive.org/web/20260920133227/https://proceedings.mlr.press/v70/guo17a.html
Archive status: existing snapshot, not a new capture. Recapture needed; snapshot contents not independently rechecked in this round.
Status: Verify before publish. Recheck within 30 days of publication.
R22
Shunyu Yao et al. ReAct: Synergizing Reasoning and Acting in Language Models. ICLR 2023; arXiv:2210.03629v3.
Type: Paper
Live source: https://arxiv.org/abs/2210.03629
Claim supported: Chapter 3, note 7 and FIG-3.3: model-generated reasoning and actions interacting with tools or environments.
Verified on: 24 September 2026. Supporting claim checked against the live source; all remain subject to final publication verification.
Archive: http://web.archive.org/web/20260921035253/https://arxiv.org/abs/2210.03629
Archive status: existing snapshot, not a new capture. Recapture needed; snapshot contents not independently rechecked in this round.
Status: Verify before publish. Recheck within 30 days of publication.