Preface

Why this book, and why now

Consider an ordinary afternoon in a health system in 2026. The consultation note is drafted by an ambient system that listened to the visit and returned a structured summary for the clinician to correct and sign (Tierney et al., 2024). The differential the team reasons against was sketched, in part, by a model that now answers medical-licensing questions at a level that would pass the examination (Nori et al., 2023; Singhal et al., 2025), and that, in a controlled comparison, produced answers to patient questions that blinded clinicians rated as more responsive and more empathetic than the physicians’ own (Ayers et al., 2023). Down the hall, the systematic review that will anchor next quarter’s guideline update was screened, at first pass, by a language model working through several thousand abstracts overnight. In the public-health department across the city, an analyst is drafting a rapid risk assessment for an emerging outbreak with the same class of tool. None of this was routine when today’s entering students sat their undergraduate examinations.

We do not rehearse these developments to impress the reader, and the reader should be wary of anyone who recounts them in order to. The case this book makes is narrower and, we think, more durable. A working knowledge of generative AI is becoming a condition of competent practice in the health sciences, in the specific sense that a professional who can use these systems, and who can say precisely when they should not be trusted, will hold a measurable advantage over one who cannot. The advantage is not the tool. The advantage is the judgment that governs the tool.

That distinction matters because the same models that pass the licensing examination also reproduce race-based clinical claims that the field discredited decades ago (Omiye et al., 2023), and the most fluent paragraph a model produces is, often enough, the one that is most confidently wrong. Unfortunately, the failures are rarely loud. They arrive as a plausible citation to a paper that does not exist, a summary that quietly inverts a study’s finding, a subgroup the model never saw and therefore never served. The professional’s task, then, is neither uncritical adoption nor principled refusal, both of which are abdications, but disciplined use: knowing what to delegate, what to verify, and how much verification the stakes demand. That habit of mind, not fluency with any particular product, is what this book sets out to build.

But why learn this now, when the specific systems are changing month to month? Precisely because they are. The models named in this preface will be superseded, in all likelihood before the second printing; the frontier that belonged to GPT-4 in 2023 is held, as we write, by models two full generations further on. What does not change at that pace is the shape of the problem: capability and the difficulty of verifying it rise together, and the durable skill is learning to work at that moving edge rather than memorizing the tool of the season. A reader who internalises the method will remain competent across releases that a reader who memorized one interface will not.

For the public-health researcher the argument carries an additional weight, because the unit of consequence is a population rather than a patient. When the World Health Organization judges it necessary to issue governance specifically for large multi-modal models in health (World Health Organization, 2024), and when disease surveillance, evidence synthesis for policy, and health communication are already being reshaped by these tools, the technology has passed from novelty into infrastructure. The field has been here before in miniature: Google Flu Trends was, for a season, a celebrated instance of population-scale prediction, until it drifted and overshot and taught a generation of epidemiologists why a model that is not maintained and verified is worse than no model at all (Lazer et al., 2014). The lesson did not expire; it scaled. This book is written so that its readers, whether they will sign a clinical note or advise a health department, meet the current generation of these tools with that lesson already in hand.

The book in the sequence

This volume is the fourth in a graduate sequence on quantitative methods for clinical and public-health research. The first three volumes cover the methods and infrastructure of research practice: introductory and advanced statistical computing, plus a practicum on the workflow tools that surround them. This fourth volume covers the orthogonal axis that has reshaped applied analytic work between 2023 and 2026: the use of generative AI as a working collaborator.

The intended audience is the graduate student or working researcher in biostatistics, clinical research, or public health who runs the analyses and signs the reports. Throughout the book, the researcher refers to whoever sits in that seat regardless of disciplinary label. The book assumes familiarity with statistical methods at the level of a first-year graduate sequence and fluency in R at the level of R for Data Science. It does not assume prior experience with large language models, agents, or multimodal AI; those are introduced from the ground up, oriented to the research use cases the audience will encounter.

What this book covers

The 14-chapter structure is organized in five parts:

  1. Foundations: a brief genealogy of generative AI; the contemporary landscape and capability classes; reasoning models, context, and the verification problem; and programming with AI as a collaborator in R and Python.
  2. Grounding models in biomedical evidence and data: retrieval-augmented generation over biomedical corpora; synthetic data with privacy guarantees; multimodal medical AI.
  3. Agentic workflows and tool use: agents, tool use, and the Model Context Protocol; deep research and evidence synthesis pipelines.
  4. Evaluation, safety, and governance: evaluation beyond the benchmark; safety, bias, and red-teaming; regulation, privacy, and the IRB.
  5. Customisation, deployment, and practice: fine-tuning, distillation, and AI-augmented teams; deploying AI in clinical and public-health practice.

The chapter list was constructed by surveying current graduate-level applied-AI curricula across major US biostatistics, public-health, and adjacent programs (the survey is documented in docs/syllabi-survey.md) and cross-checking the result against a 24-month digest of Ethan Mollick’s One Useful Thing Substack (docs/mollick-digest.md), a touchstone for contemporary applied genAI thinking. The result is a curriculum that reflects mainstream 2026 practice while staying anchored to clinical and public-health research use cases.

What this book does not cover

The book deliberately omits topics that are taught elsewhere in the sequence or in dedicated courses:

  • The mechanics of training large language models from scratch, including pre-training and reinforcement learning from human feedback. We treat models as capability tiers to be selected and verified, not artifacts to be trained.
  • Classical machine learning, which is treated in Advanced Statistical Computing in the Age of AI, Chapter 10.
  • General software engineering for statisticians, which is treated in Advanced Statistical Computing in the Age of AI, Chapter 11.

Pointers to relevant chapters appear where they arise.

Three load-bearing concepts

Three concepts thread through every chapter and earn the ‘applied’ framing.

The jagged frontier. Generative AI capability is uneven: superhuman in some dimensions, subhuman in adjacent ones. The frontier is also opaque, capability and verification difficulty grow together. The researcher’s contribution is to identify which jagged edges still require human judgment and which do not, and to design verification proportionate to the stakes.

The cybernetic teammate. Modern AI assistance is more fruitfully framed as collaboration with a teammate than as operation of a tool. The 2024–2026 arc has shifted the human role from prompter to manager: the binding skill is precise specification of deliverables, fast evaluation, and judicious delegation. This framing organizes the book from Chapter 1 forward.

Verification as a first-class concern. Every chapter treats verification not as a final step but as a design property. The section template in every chapter front-loads The researcher’s contribution, the judgments no AI can make on the reader’s behalf, and ends with Collaborating with an LLM, where specific prompts are paired with what to watch for and how to verify. The dialectic structures the chapter rather than decorating it.

How to read this book

Each content chapter follows the same structure: Learning objectives, Orientation, The researcher’s contribution, content sections (with collapsible Check-your-understanding callouts at natural pauses), Worked example, Collaborating with an LLM, Principle in use, Exercises, Further reading.

Chapters can be read in order or out of order. Topics with a chain of dependencies (e.g., Chapter 8 builds on Chapter 3; Chapter 14 builds on Chapter 13) are noted in the relevant Orientation sections.

Acknowledgments

The chapter list reflects a survey of graduate-level genAI curricula across roughly twenty US programs and several industry-adjacent open courses. Ethan Mollick’s One Useful Thing Substack provided the contemporary frame against which the chapter list was tested. The authors of R for Data Science, Advanced R, R Packages, Bayesian Data Analysis, and Statistical Rethinking established conventions this book inherits from its sister volumes.