Risks, ethics and adoption

The F chapter lists fifteen key risks (p. 62-63). Grouped, they come down to five worries:

• Unfair output. Models learn bias from their training data.
• Fake and misleading content. Deepfakes, fake reviews, false personas.
• Data and security. Privacy, leaks of training data, attacks that fool the model.
• Blind trust. People stop checking because they assume the model is right.
• Wider effects. Jobs, less value placed on human creativity, energy use, rules that lag behind, and models that get worse as the world changes.

The book’s line is clear: be careful, as with any new technology, but inaction is not an option (p. 62).

The R chapter gives five strategies (p. 200-203):

1. Transparency. Say how and why you use data, and why the model decides what it decides.
2. An ethical AI maturity model. Assess your practice on fairness, privacy, accountability and bias, then improve step by step.
3. Data privacy and security. Protect data at rest, in transit and while it is processed.
4. Explainable AI. The model shows what drove its answer, so people can check it.
5. Legislation. Keep legal, risk and compliance people inside the AI program.

The F chapter adds a longer checklist of fifteen steps, from data quality and bias checks to fail-safes, external audits and AI-specific insurance (p. 63-65).

The book describes the EU proposal as it stood when it was written: a risk-based approach. Unacceptable risk is banned. High risk is allowed, with requirements. Limited risk gets light transparency duties (p. 202-203). For the US it describes the blueprint for an AI bill of rights, which is guidance, not law (p. 202).

Its conclusion: rules move slower than the technology, so leaders must set their own guardrails while they wait (p. 202-203).

Human-in-the-loop means people review, correct and approve what the AI produces before it goes out (p. 92). The F chapter makes it a rule: keep humans in critical decisions, with the power to override the AI (p. 64). The T chapter says the same for code: AI-written code still needs testing and human review (p. 223).

The book does not pretend jobs stay the same. The N chapter describes a mix of replacement, displacement and collaboration, with repetitive work, translation, simple replies and summaries most exposed (p. 160). Its advice is to build the human skills machines do not copy well, like empathy, creativity and ethical judgment, and to keep learning (p. 161).

The P chapter gives the short version: generative AI will not replace people, but people who use it well will replace those who do not (p. 182). The Q chapter adds that every company should be teaching its senior executives and board members about generative AI (p. 191).

The H chapter shows six steps for leaders (p. 91): build the business case, prove the concept, invest in an AI team, set up governance and a responsible AI policy and train people, invest in data security, then build, test and learn.

The D chapter adds eleven moves for leaders (p. 47-48). A few stand out: state the company’s stance on generative AI clearly, educate the board and executives first, build a prompt library everyone can use, create a central team to govern approved models, and make good use of AI part of everyone’s objectives.

Best practice guide #1 lists the roles (p. 36-40). To build: an executive sponsor, a project manager, a chief AI officer or architect, data scientists, business analysts, machine learning engineers, software developers, and legal and compliance. After go-live: support engineers, a trainer for users, and data scientists and engineers who keep the model up to date. A RACI chart says who is responsible, accountable, consulted and informed.

Guide #4 widens the circle. Citizen innovators are staff outside IT who build with low-code and generative tools. The book is for them, with guardrails: training, mentoring from IT, approved tools, and no shadow IT (p. 135-147).

Guide #5 ends with three pieces of advice (p. 174-180). Do proper assessment and discovery before you build. Plan for support and maintenance: support reacts to problems, maintenance prevents them. And stay tech agnostic: do not lock yourself into one model, framework or vendor, because the field moves too fast.

Pascal Bornet’s foreword makes the same point from the top (p. 8-9): a clear strategy tied to business goals, and ethics treated as part of success, not an extra.