The glossary

An AI system that can take action. It books, sends, buys and posts. From here, the accountability questions in this book get sharper.

BCG’s name for the exhaustion of approving fluent machine output all day: more mistakes, more decision fatigue, and a strong urge to lie down. It peaks, unhelpfully, in the people juggling the most tools at once.

Machine-generated content produced at volume with nobody checking it, nobody standing behind it, and nobody expected to. The billboard version lost an election in this book.

The drop in trust, and sometimes moral disgust, that can occur when people discover that something was made or believed to have been made by AI.

Grief felt before a feared loss has happened, while the outcome is still uncertain and part of you is already living beside the worst one.

Trusting the machine’s answer over your own, and over the evidence in front of you. Nobody arrives at it through laziness. A tool that is right nine times in ten teaches you, quite correctly, that checking it wastes an afternoon. Then the tenth arrives dressed exactly like the other nine.

A South African baked dish of gently spiced minced meat under a savoury egg topping, with family opinions attached.

What happens when a worker, burnt out by botsitting, publishes what the machine produced without checking it. The researchers who coined the word surveyed six thousand workers; a remarkable share admitted to it.

The time-consuming work behind the scenes that makes generative AI usable. Feeding it context, checking hallucinations, debugging errors and cleaning up after it.

A role in which named AI workflows handle repeatable execution while the person directs them and keeps the judgement, taste, relationship and responsibility.

The MIT brain-scanner finding that the thinking you skip doesn’t stay skipped. It compounds, like interest, payable later.

Moving part of a mental task out of your head and into a device, note or other external system. Useful, often sensible, and extremely good at turning a phone number you once knew by heart into information held hostage by a flat battery.

The subtle handover of judgement to an AI system, usually after enough plausible answers have trained the person to stop checking properly.

The market deciding that work is interchangeable. The task may still exist, but the price and power drain away because a close substitute is easy to produce.

Machine-readable provenance records attached to digital media, showing where it came from and what happened to it. A birth certificate for the file, not a certificate of truth.

Deloitte’s term for tools scaling faster than the humans around them, so a confident slide moves through the meeting with nobody quite sure where it came from.

Roles in which AI makes the work easier for more people to perform, reducing the advantage once supplied by specialist access, training or technical execution.

The second look after taste reacts. It examines whether something is genuinely weak or merely unfamiliar, and whether the first response contains judgement, habit or prejudice.

The mistake of assuming that because a doorman opens the door, opening the door is the whole job. Automating the visible task can hide everything else the role was carrying.

Cory Doctorow’s word for the deliberate decay of a platform once it has you. The service gets worse on purpose, one cent at a time, and leaving costs more than staying.

The delay between a habitat being damaged and the species depending on it finally disappearing. The book uses it for human originality that may linger after its conditions have already been eroded.

The most capable AI systems available at a given moment. The frontier moves; the book’s argument is built on the parts of work that don’t.

AI that produces new text, images, audio, code or other material from patterns learned during training. It generates rather than merely retrieves, which is why it can be useful, original-looking and wrong at the same time.

When a model states something false with complete confidence, like a granny taking her children out for cookies in a story that contained no granny and no cookies. The machine is not lying. Lying requires knowing.

The judgement, context, relationship, taste and responsibility wrapped around machine output. AI can accelerate the work underneath it. The layer is what makes the result belong to someone.

A data centre built to grow without limit, the kind that negotiates its water with a whole town rather than a landlord. When one is proposed near your house, read the fine print. Ask Mo.

Frederic Tudor, the nineteenth-century merchant known as Boston’s Ice King, and the subject of a wonderfully tidy disruption story that falls apart when checked. A photograph and a confident speaker do not make an anecdote true.

The risk that abundant machine intelligence lets institutions create wealth and make decisions while needing fewer ordinary people, weakening the leverage people once gained from their work.

The gap between what AI can theoretically do and what people are willing or able to trust it to do in real work.

An internet wink meaning “if you know, you know”: the right people already have the context.

Large language model, the engine under the chatbot. It learned language from us. That’s the marvel, and quite a lot of the problem.

Planned, rotating electricity cuts used to prevent the power grid from collapsing when demand exceeds available supply. South Africans generally learn the schedule after the lights go out.

What happens when AI trains on AI output: each generation learns from the last one’s mistakes and flattens toward a confident average. The Flattening chapter calls it cannibalism, which is ruder and more accurate.

The comforting idea that a human clicking ‘Approve’ at the end has supplied the judgement the process needed at the beginning. Usually they’ve only supplied the click.

Per- and polyfluoroalkyl substances: a large family of synthetic chemicals used to make products resist water, grease, stains and heat. Many persist in the environment and the body, which is why they are known as ‘forever chemicals’.

Roles in which AI amplifies human expertise, leaving the person to perform more of the judgement-heavy work rather than making the role easier for anyone to do.

What you type to the machine. Occasionally left at the top of a school exam by mistake, which is how the students found out.

The cryptographic paper trail showing where a photograph or video came from. Cameras now sign images at the moment of capture. A signature can help prove the file’s origin, but it can’t tell you whether the person behind it deserves your trust.

An AI system improving the process that builds the next, more capable version of itself, including choosing data, running training and checking the result.

An axolotl-shaped tea infuser whose “designed by people” label becomes a tiny, splendid argument for human provenance.

The pattern in which resource wealth can make a state less answerable to its people because power no longer depends as directly on their productive participation.

The book’s image for AI capability spreading across more and more tasks: a level that keeps climbing, task by task, rather than one dramatic wave, while people decide what to do about it.

Ethan Mollick’s name for a company’s heaviest AI users, who keep quiet about how much the machine does in case somebody does the arithmetic.

The feeling that another person has actually arrived in an exchange, even through a screen: what happens can reach them, and they can alter what happens next.

The human capacity to care enough that being wrong costs something real: the privilege of choosing, and the less glamorous privilege of owning what was chosen.

Understanding learned by doing a task alongside someone who already knows it. It is the part that never quite makes it into the manual.

A friendly name for computing that happens in physical data centres. The interface may feel weightless; the servers, electricity, cooling and water are not.

A small, often dangerous spirit in Southern African folklore, associated with night-time mischief, fear and harm. One familiar defence is to raise the bed on bricks, which is a great deal of furniture rearrangement for something nobody claims to believe in.

The material used to teach an AI model its patterns. What is abundant becomes easy to reproduce; what is missing, rare or repeatedly copied can be flattened or lost.

The fear of the number thirteen, from the Greek, and the reason lifts skip a floor that is still there. Its Friday variant, paraskevidekatriaphobia, is longer than the fear it describes. This book has no Chapter 13 either. That one is at https://humanstillwins.com/chapter-13/, where I have put the part of us that no machine can be talked into.

The book’s five-part review before consequential AI-generated work leaves your desk: True, Relevant, Understandable, Safe, Take ownership.

An economy in which information and polished output are abundant, so value shifts towards people and institutions whose judgement, provenance and promises can be believed.

AI-made work that looks finished until you touch it and discover the real thinking has been left for the recipient.