*Safe Enough* follows artificial intelligence out of the laboratory and into systems where a mistake can injure or kill: the aerospace certification lab, the hospital procurement committee, the autonomous vehicle programme that ran out of edge cases. A model is usually one part of a larger system, so the authority to declare the whole thing acceptable has to sit somewhere. The book is about where it sits, who earned it, and what is changing underneath. **Status** — A reported nonfiction project in development. [Follow the research](https://theaimonitor.substack.com/subscribe) [The Five Questions](/five-questions) ## What the book covers *Safe Enough* is reported nonfiction for curious general readers. It follows real deployment decisions in aviation, emergency medicine, cancer screening, autonomous transport, heavy industry, and defence, and asks the people involved how the decision was actually made. It is not a book about whether AI is good or bad. It is about evidence, judgement, and responsibility: what a test result establishes, who turns that result into permission, and what happens when the system changes after the decision is made. ## The Five Questions  Five questions run through every chapter. None of them is new, but AI makes each one harder to answer. 1. What would change your mind? 2. Who checks the machine after it changes? 3. What happens when they all fail at once? 4. Who answers when it goes wrong? 5. Who is your strongest opponent? [Read the Five Questions and download the field card](/five-questions) ## Why I am writing it I have spent 28 years in software for systems where failure can cause physical harm: aerospace, automotive, nuclear, and medical devices. Much of that work is about producing the evidence on which other people decide. The book asks how that process holds up when the software learns, adapts, and is updated in the field. I publish from the research as I go, in [The AI Monitor](/newsletter) and on [LinkedIn](https://www.linkedin.com/in/adammackay/). The same method (interviews with named specialists, who check their quotes before publication) produced *[You can't patch a safety case](https://www.hipeac.net/news/magazine/7175/hipeacinfo-79/#/)*, a reported feature in HiPEACinfo 79 (October 2026) on keeping certified systems secure when the cryptography they depend on has to change. ## Contribute to the research If you have been part of a decision to deploy AI where a mistake could cause physical harm, whether you approved it, refused it, or produced the evidence someone else relied on, I would value hearing how it was decided. Engineers, clinicians, regulators, operators, researchers, and critics are all welcome. [Get in touch](/contact)