The glossary

Agentic AI is AI that does not stop at answering. It uses tools, moves through steps and takes actions: retrieving records, updating systems, sending messages, making bookings and generally wandering around the workflow with permission to touch things. This is useful. It is also where “the computer made a mistake” acquires a larger blast radius. A weak paragraph can sit on your screen until somebody fixes it. An agent can email the client, change the database and book the wrong Tuesday before the kettle has boiled. The system is not conscious or independently ambitious; it acts because people configured it, authorised it and gave it access. That means useful agents need fences: limited permissions, visible logs, stopping conditions, human review at consequential steps and one named person who remains answerable. Delegation can move the action. It cannot move responsibility into the machine, however briskly the machine clicks.

AI brain fry is the cognitive exhaustion that arrives when a person spends too long monitoring, correcting and approving fluent machine output. The machine can produce another page instantly. The person checking it still has one brain, one working day and a cup of coffee that went cold twenty minutes ago.

The danger is not simply feeling tired. Once attention frays, mistakes pass through and decisions become poorer. Adding another AI tool to the pile does not add another supply of judgement.

The term belongs beside botsitting because it names the cost hidden inside the productivity claim. Speed at the front of the process can become fatigue at the back. A useful system counts both.

AI slop is not simply anything made with AI. That would be like calling every meal involving a blender soup, which is plainly nonsense. Slop is what happens when machine-made content is produced at speed and nobody does the final human bit: checking it, shaping it, noticing it has invented a statistic, or asking whether another person deserves to have it dropped into their day. It often looks finished. That is part of the problem. The headings are neat, the sentences wear clean shoes, and somewhere in paragraph four there is a claim nobody can source and a phrase so generic it could have been printed on the wall of an airport business lounge. In *Human Still Wins*, slop is the public puddle left by a chain of abdications. Botsitting became tiresome. Checking was skipped. Workslop moved downstream. The machine participated; the human disappeared. That is the line.

The AI-authorship effect describes what can happen when people discover that a product or message was made, or is believed to have been made, by AI.

The reaction is not always a simple judgement that the work is poor. It can feel moral: a sense of being deceived, substituted for or asked to accept a relationship in which no person was present on the other side. The trust break concerns authorship as much as quality.

This is why ‘made by people’ may become a premium label rather than a sentimental one. Human authorship does not guarantee good work. It does tell the recipient that somebody selected this particular thing, and can be asked what they thought they were doing.

Anticipatory grief is grief that begins before the loss itself, while the outcome is still uncertain and the mind has already moved into the room with the worst possibility.

In Chapter 12, Timea recognised the term while waiting to learn whether her mother had died. The words did not remove the fear. They gave it a name, and the name gave her somewhere to put it for an hour.

That is what a useful definition can do. It does not solve the experience. It tells a person they are not failing at it.

Bobotie is a South African baked dish of gently spiced minced meat under a savoury egg topping, with household arguments available about fruit, sweetness, texture and whose mother’s version is obviously the correct one. In *Human Still Wins*, bobotie becomes an accidental test for machine prose. The author searches for a quick recipe and finds a polished page describing the dish as “a delightful culinary journey”. No. It is six in the evening, the mince needs attention and somebody has used the last clean wooden spoon. The useful recipe contains more than ingredients. It knows which apricots, how wet the mixture should be, when the quantities become instinct and why the stained family page is still in the recipe folder. That is taste: not decoration, but lived selection. A machine can describe bobotie fluently. The sentence that has never been near the kitchen gives itself away. Usually around “culinary journey”.

Bobotie — from my grandmother’s cookbook

What goes in

• 1½ lb lean beef mince
• 1 large onion
• Some butter
• ½ cup milk
• 3 eggs
• 2 slices white bread, cubed
• ¼ cup finely chopped dried apricots
• 2 apples, peeled and grated
• ¼ cup seedless raisins (optional)
• ¼ cup blanched almonds
• 1 tablespoon sugar or apricot jam
• 1 tablespoon curry powder
• 1 tablespoon lemon juice
• 1 teaspoon salt
• ¼ teaspoon pepper
• ¼ teaspoon turmeric

What to do

1. Saute the onion in hot butter until golden. Add meat and stir with a large fork until the meat loses its redness. Remove from heat.
2. In a large bowl, mix together ¼ cup milk, one egg and the bread cubes, mashing the bread with a fork. Add the apricots, apples, raisins, almonds, sugar or jam, curry powder, lemon juice, salt and pepper and mix until well blended.
3. Add to the meat mixture and mix lightly with a fork. Turn into a greased oven-proof dish and spread evenly.
4. Bake covered for 20 minutes at 180°C. Take out of the oven and add the topping.
5. For the topping, beat the remaining 2 eggs with the remaining ¼ cup milk and the turmeric until just blended. Put the bobotie back in the oven and bake for a further 15 minutes, or until the topping has set and is no longer runny.
6. When you add the topping, make sure the mixture just covers the bobotie, otherwise it takes ages for the egg mixture to cook. Watch it carefully so the topping does not burn.
7. Serve with rice and chutney.

Transcribed exactly as typed, including the parts that assume you already know what you are doing.

Botshitting is sending AI-generated work you have not properly checked, do not fully understand and could not defend if somebody in the meeting leaned forward and said, “Right. Explain this bit.” It is the moment the draft puts on a blazer and is introduced as finished work. The term sits beside botsitting in the Work AI Index 2026 for a reason. People who spend hours supervising, correcting and rerunning AI eventually get tired. The first plausible answer begins to look very attractive, especially at 16:47 on a Thursday. Understandable. Still not somebody else’s problem. The missing thinking does not vanish when the file is sent. It moves downstream to the colleague, client, teacher or public official who now has to find the errors and rebuild the logic. In *Human Still Wins*, botshitting is not merely poor prompting. It is a failure of ownership. Do not send work you cannot explain, verify or stand behind. Even if the bullets are beautifully aligned.

Botsitting is the hidden second job that arrives after the machine was meant to save you time. You provide the missing context, rewrite the prompt, compare three strangely similar answers, check the citations, correct the confident nonsense and press “try again” until your coffee has gone cold and the original task has begun to look quite restful. The Work AI Index 2026 named and measured this labour after surveying six thousand digital workers. The useful point is not that checking AI is unfair. Important work should be checked. The problem is pretending the checking does not exist when calculating the miraculous productivity gain. Botsitting is the human supervision that makes machine output usable. It needs time, skill and recognition. When organisations budget for instant AI speed but not for the person cleaning up behind it, exhaustion follows. And exhausted people eventually stop checking. That is when botsitting slides into botshitting, which is not a promotion.

Captain of your AI workforce is the agency’s name for a role redesigned around direction rather than manual repetition.

The person maps the predictable parts of the job and gives those parts to named AI workflows. The machine drafts, schedules, structures or sorts. The captain reviews, corrects and sends, because the captain still knows whether the shape fits the situation.

This is not a naval promotion for owning several chatbots. It is a practical division of labour. Execution can happen at machine scale while judgement, taste, relationship and responsibility remain clearly human. The governing idea is simple: you cannot be commoditised by something that reports to you, provided you are still doing the part worth reporting to.

Cognitive debt is the thinking ability you fail to build when the machine repeatedly does the difficult mental work for you. The saved time feels immediate. The missing capacity appears later.

A person who asks AI to test a conclusion still brings a conclusion. A person who reaches for it before forming the question skips the part where judgement is made. Repeat that often enough and the skipped thinking compounds.

The debt is not caused by using AI. It grows when use replaces the struggle through which a person learns to frame, compare, doubt and decide. The interest is paid the first time the machine is wrong and nobody at the desk has a reference point.

Cognitive offloading is what happens when you move part of a mental task out of your head and into the world. You write a shopping list, set a reminder, save a contact, follow GPS or ask AI to hold the thread while you do something else. Humans have always done this. Paper is an external memory. So is the back of your hand, although it performs badly in the bath.

Offloading is not automatically surrender. A calendar is not a moral failure; there is no medal for remembering your dentist appointment unaided. The trouble begins when the tool does not merely store the answer but removes the practice that built your ability to reach it. Nobody announces the trade. You get turn-by-turn directions, instant recall and the name of the actor who was in that thing, and in exchange the route through your own neighbourhood quietly disappears.

That is why forgetting phone numbers matters beyond nostalgia for landlines and those little alphabet tabs in an address book. The number was once repeated, retrieved, rehearsed. The memory had reps. Save everything once, summon it on demand, and the effort vanishes first. The capacity may follow.

AI widens the category. We can now offload not only memory but framing, comparison, drafting and doubt. This can be marvellous. It can also leave the human at the end of the process approving an answer they no longer know how to test.

The useful question is not, “Did I use a tool?” It is, “Which part did I hand over, and do I still want that part of me to work?”

Cognitive surrender is what happens when the decision moves out of your hands so gradually that you do not notice the handover.

The model gives the correct answer and the invented one in the same calm voice. After approving a hundred adequate responses, the person at the desk reads the next one with less resistance. Soon the tool is not assisting judgement. It is supplying it, and the human approval step has become a small ceremony performed at the end.

This is not the same as using a calculator for a sum. The calculator is not pretending to weigh context or make a recommendation. Cognitive surrender begins when fluency is mistaken for authority. The defence is not permanent suspicion of the machine. It is keeping enough knowledge and attention to know what deserves suspicion.

Commoditisation is what happens when the market decides your work is interchangeable.

The task does not necessarily disappear. It becomes easier to source, cheaper to buy and harder to distinguish from the next competent version. More work may be demanded for the same pay. A specialist function may be folded into somebody else’s role. Nobody announces that the value has moved. The rate simply starts behaving as though it has.

AI accelerates this wherever output can be replicated quickly and at scale. The defence is not insisting that your draft took longer. It is building value around what cannot be swapped without consequence: judgement, trust, context, a standard and ownership of the outcome.

Content Credentials are machine-readable provenance records attached to digital media. They can record where an image came from, which device captured it and what edits followed.

The book calls this a birth certificate for the file. That is useful because looking for strange fingers and melted lettering is not a serious long-term verification system. As generated material improves, appearance tells us less.

A credential does not prove that the photograph is honest, kind or correctly interpreted. It proves something narrower and valuable: a traceable history. The difficult part is keeping that history attached when platforms strip metadata on upload. Provenance can exist before the plumbing reliably carries it.

Cultural debt is the slow organisational damage that accumulates when AI tools spread faster than the rules, habits and human judgement needed to use them well.

A recommendation appears in a slide. Nobody is quite sure where it came from, but the meeting is running late and it looks plausible, so it moves. Nothing explodes. The purchase order is ninety-eight per cent right, the report is fine again, and a few thousand defensible approvals gradually lower the standard.

Technical adoption can happen in a quarter. A culture capable of questioning, correcting and taking ownership takes longer. Pretending otherwise does not remove the debt. It only hides the statement.

Democratised roles are jobs in which AI makes the work easier for more people to perform.

A task that once required specialist software, technical training or a separate hire can be done by an ordinary user with a prompt and a decent afternoon. Access expands. The old barrier around the skill shrinks.

That can be genuinely useful. It can also lower the price of the work and reduce the number of people paid to do it as a distinct role. Democratisation describes access, not a guarantee that the transition will be fair to the people whose living depended on the old barrier.

Discernment is the second look.

Taste reacts first. Something fits or it does not. Discernment asks the response to explain itself. Is this weak, or merely unfamiliar? Is it elegant, or merely expensive? Did the work fail, or did it violate a rule you inherited without examining?

The distinction keeps taste from becoming prejudice in a better jacket. It also keeps discernment from producing work that passes every check and moves nobody. You need the instinct and the examination. One catches what a checklist misses. The other makes sure your history is not speaking with more authority than it earned.

The doorman fallacy is the mistake of assuming that because a doorman opens the door, opening the door is the job.

A doorman also recognises residents, screens strangers, receives packages, notices when something is wrong and maintains the order of a building. Automate the door and you have removed the most visible task. You have not necessarily replaced the role.

The same mistake runs through arguments about AI and work. Jobs look like bundles of tasks on paper, so a machine completing the obvious task can appear to have swallowed the job. The missing work is often context, judgement and relationship, the things that were load-bearing precisely because nobody bothered to list them.

Enshittification is Cory Doctorow’s term for the slow ruining of a platform after it has made itself difficult to leave. First it is wonderful to users. Then suppliers are squeezed. Then everybody gets more adverts, worse search, stranger rules and a button that used to be visible is now hiding behind six menus and a subscription. No single update announces, “Hello, we have made this worse on purpose.” The decline arrives dressed as optimisation. Convenience keeps people there while the bargain quietly changes underneath them. The book places enshittification beside model collapse because both systems can deteriorate through their own incentives while still looking successful from the outside. One hollows out the platform; the other can flatten the model. Human judgement begins with the slightly irritating person in the room saying, “Hang on. Wasn’t this supposed to help us?”

Extinction debt is an ecological term for the delay between a habitat being destroyed and the species that depended on it finally disappearing. The damage has happened. The full loss has not arrived yet.

In *Human Still Wins*, the idea is borrowed for culture. Original human work may remain present long enough for us to notice when the copies become flatter, even while the conditions that produced the original signal are being weakened. The old richness can survive in the system for a while.

That delay is dangerous because it looks like safety. We can still recognise the real thing, so we assume the real thing will keep replenishing itself. A debt is not evidence that nothing was lost. It is loss waiting for its date.

A frontier model is one of the most capable general-purpose AI models available right now. Right now is doing heavy work in that sentence. The frontier moves every few months, occasionally every few days, and yesterday’s marvel can become the option you scroll past by Thursday. These models attract attention because they can handle broader tasks, longer context, more complicated reasoning and tool use. They raise the floor of what can be automated and extend how far a machine can travel through a job before asking for help. *Human Still Wins* avoids hanging its argument on one product name because product names age like yoghurt in a hot car. The durable question is what the new capability changes – and what it does not. A smarter model may complete more of the task. It does not become trustworthy, tasteful, courageous or responsible simply because the benchmark number went up. Capability moves. Accountability stays with us.

Generative AI is the branch of AI that makes things: text, images, audio, video, code and the first draft of the email you did not want to write. It learns patterns from large bodies of material and uses those patterns to produce a new response to a prompt.

New does not mean lived, witnessed or true. The system can generate a convincing answer without having an experience behind it, and it can produce a false detail with the same composure as a correct one. That is not a small technical footnote. It is the reason the output still needs a person.

In *Human Still Wins*, generative AI is less a mechanical colleague than an acceleration engine. It removes blank pages, moves information between forms and makes competent output cheap. The human job begins where the generation ends: deciding whether the thing is right, useful, safe and worth sending.

An AI hallucination is a made-up fact, source, quote or detail delivered in exactly the same polished tone as the true bits. It does not blush. It does not look shifty. It simply hands you a fictional road and says the hospital is at the end of it. The model is not lying in the human sense. Lying requires knowing what is true and choosing otherwise. An LLM is generating the most plausible next pattern, and sometimes plausibility arrives wearing a name badge marked FACT. This can be funny when the machine introduces an imaginary granny into a family story. It is less adorable in medicine, law, finance or a school assignment that now cites a journal invented five seconds ago. The remedy is not to ask for more confidence. Give it sources. Require evidence. Check the important claims. Keep a person answerable for the result. In this book, every AI answer remains a draft, even when it has excellent punctuation.

The human layer is everything wrapped around the output that the output cannot provide for itself: the context, the relationship, the standard, the decision and the person who will answer when it lands badly.

It appears in small places. The follow-up question in an interview. The designer who notices the wrong line on a bottle. The colleague who knows a technically correct email will inflame the client. The person who reads the draft and asks whether it deserves to enter somebody else’s day.

AI can make the layer easier to neglect because the work underneath it arrives looking finished. It is not finished. A polished answer still needs an owner. When competent production becomes cheap, the human layer is where value concentrates.

Hyperscale is computing built to grow across enormous numbers of servers, data centres, cables, cooling systems and power supplies so that millions of people can ask for something at once and receive it before they have finished sighing. The interface makes this feel weightless. You type a prompt on the sofa; somewhere else an industrial building gets on with the frankly less poetic business of electricity, heat and water. *Human Still Wins* follows a proposed hyperscale facility in Farmington, Minnesota, where residents questioned a contract permitting water demand greater than the city’s ordinary daily use. That does not mean every chatbot query requires a candlelit apology. It means the magic has plumbing. Hyperscale gives us speed, reach and astonishing capability. It also puts very physical costs into particular communities. Good judgement is being able to admire the machine and still ask who is paying for the cold water.

Boston’s Ice King was Frederic Tudor, a nineteenth-century merchant who built a trade in natural ice. In the conference story that reached this book, he was wiped out almost overnight by manufactured ice. It was a beautiful anecdote. It was also wrong.

The natural-ice trade continued for decades, and the refrigerator arrived much later. The story survived because it had the right shape: old industry, new machine, clean extinction. A photograph made it feel sourced. A confident speaker made it feel checked.

Here, the Ice King is a warning about stories that fit too well. Judgement sometimes looks like spending a dull half-hour ruining the anecdote you most wanted to use.

The intelligence curse is the risk that abundant machine intelligence changes who a system needs.

The argument borrows from the resource curse. When wealth arrives from oil or diamonds rather than from citizens’ labour, governments can become less answerable to those citizens. If intelligence becomes something firms can buy from machines at scale, workers may remain capable while their work buys them less influence.

This sits above the question of whether one person can stay employable. Expertise can still earn a place in the room. It does not guarantee a vote over what the room is for. The curse is not that intelligence disappears. It is that human intelligence may stop being the source from which power needs to draw.

The invisible line is the distance between what AI can do in theory and what people actually hand over to it at work.

Part of the gap is technical. The tool may not be good enough, the systems may not connect or the data may not be allowed to leave the building. The rest is human: context nobody wrote down, consequences somebody must own, and the checking that consumes the hour the machine supposedly saved.

The line moves as the tools improve. It does not vanish simply because a benchmark rises. Work crosses it when somebody judges that the benefit is worth the risk and puts their name near the result. The valuable skill is not refusing to delegate. It is knowing what can be given away and which part would be dangerous to lose.

IYKYK means “if you know, you know”. It is the internet’s tiny raised eyebrow: a reference offered without the full PowerPoint because the right people already have the context. In the manuscript it follows a joke about Richard Starkey not being Ringo Starr. Some readers will catch it immediately. Others will continue with their lives, which is also allowed. The phrase belongs here because shared context is not decorative fluff around communication. It is communication. The odd family line, the local joke, the exact song somebody played in 1997 while packing a suitcase – these details tell us that a particular person is speaking to particular people, not spraying pleasant beige language over everybody. A model can expand IYKYK into its four words. The human work is knowing when explanation would kill the thing. Sometimes the little door is for those who recognise it. Everybody else may use the main entrance.

LLM stands for large language model, the pattern engine behind many chatbots and generative writing tools. It has been trained on enormous quantities of language so it can predict and produce plausible sequences of text at startling speed. There is not a tiny person inside the laptop remembering facts and composing a view. The model works from patterns, prompts and context. This is why it can summarise a report beautifully, translate a paragraph, draft an email and then invent a book that has never existed with the calm confidence of somebody recommending it at dinner. The same mechanism produces the magic and the mess. LLMs are extraordinarily useful at rearranging language. They can also hallucinate, flatten distinctive voices and make unfinished thinking look polished. Calling the thing an LLM rather than simply “AI” helps keep the claims tidy. It is a powerful language model. It is not a person, a conscience or the new office sage.

Load shedding is the planned switching-off of electricity in different areas when the power system cannot meet demand. The cuts rotate so the entire grid does not fail at once.

In South Africa the term escaped energy policy and entered ordinary life. It means charging the laptop before the next stage, checking whether dinner can be cooked at six, and discovering that the battery percentage has become a moral judgement.

The book uses it in a small scene: the power goes, the laptop runs down and darkness decides when the work stops. International readers may hear a technical phrase. South Africans hear the click of everything in the house giving up together.

Model collapse is what can happen when generative AI is repeatedly trained on AI-generated material and begins to lose the rare, peculiar edges of the original human data. Common patterns get louder. Unusual details fade. Eventually the great buffet of language starts serving fourteen trays of the same beige pudding. The book’s ruder name is AI cannibalism: the machine eating its own output and calling the smaller menu knowledge. This is not an unavoidable curse in every training design. Keeping and accumulating real data changes the outcome, which is precisely why human-originated material matters. The strange sentence, the minority pattern, the awkward regional detail and the thing only one person thought to record are not clutter. They are what averages erase first. Model collapse turns the book’s argument about AI slop into a longer warning. Bad content does not only waste today’s reader’s time. At enough scale, it can become tomorrow’s training material. And then the photocopy gets photocopied again.

The myth of verification is the comforting belief that placing a human at the end of an AI process guarantees human judgement.

It does not. A person clicking Approve may be tired, rushed or unable to recognise what is wrong. The important judgement should have started earlier: in the question, the assumptions, the missing context and the decision about what deserves suspicion.

A human in the loop is useful only when the human is actually thinking. Otherwise the loop is theatre with a button.

PFAS is the collective name for thousands of synthetic chemicals used because they resist water, grease, stains and heat. The useful nickname is ‘forever chemicals’, which tells you the important part before the acronym has finished: many of them persist for a very long time in the environment, and some can accumulate in people and animals.

They have been used in products including firefighting foams, coatings, textiles and grease-resistant packaging. Certain PFAS exposures have been associated with immune, developmental, liver and cancer risks. The family is large, the compounds are not identical, and the science should not be flattened into one frightening sentence.

The term belongs near the book’s environmental questions because invisible convenience still has a source, a pathway and a place where the residue ends up. ‘Away’ remains one of our least convincing locations.

Professionalised roles are jobs in which AI acts as a force multiplier for human expertise.

The repeatable work becomes faster, but the role does not dissolve into a button. More of the person’s time moves towards judgement, client handling, interpretation and the decisions that require context. The tool raises what one capable person can do, and the expertise still determines whether the result is useful.

This sounds like the kinder track, and often it is. It also raises the entry price. Junior workers are asked for judgement that used to be built through years of routine work. The role becomes more valuable and harder to enter at the same time.

A prompt is the instruction, question, example or lump of context you give an AI so it knows what you are asking it to do. It can be six words, six pages or a feverish paragraph written after the third failed attempt. Good prompts help. They clarify the task, audience, constraints and standard. They are the difference between “write something about trust” and a proper brief that tells the machine what trust means here, who is reading and which bits it must not invent. But prompting is not the grand human superpower some people have tried to sell it as. A magnificent prompt cannot decide whether the original problem is stupid, whether the source is trustworthy, whether the answer fits the moment or who will take responsibility when it goes sideways. The prompt starts the exchange. Judgement frames it, checks it and decides what happens next. Also, delete the prompt before submitting the homework. Honestly.

Provenance is the digital paper trail: where a piece of content came from, who or what created it, and what happened to it on the way here. Technical standards can attach tamper-evident records to an image or other media, showing that a particular device captured it or that an edit was made. This is useful because our old method – squinting at the pixels and announcing “that hand looks funny” – is not a sustainable verification system. But provenance is not a little golden badge marked TRUE. A perfectly traceable photograph can still be misleading, cruelly framed or attached to a rotten argument. The trace answers the first question: where did this come from? Human judgement must answer the rest. As generated material becomes harder to spot by appearance alone, provenance becomes part of ordinary due diligence. Trust still belongs to people and institutions willing to put their names near the history and answer for what it shows.

Recursive self-improvement is the possibility that an AI system could help build a more capable version of itself, then use that increased capability to improve the next one.

In the version discussed in the book, the system would choose data, run training and check the result without a human executing each step. The recursion is the loop: the improved tool becomes part of the machinery that produces a still better tool.

This is not the same as an ordinary software update, and it is not proof that such a loop will arrive on schedule. It is a precise version of the safety question: what happens when making the next machine is no longer work only humans can do? The honest answer in the chapter is that nobody knows.

Relaxolotl is an axolotl-shaped tea infuser made by Genuine Fred, which means a small amphibian hangs over the side of your mug while the tea brews. Civilisation has done worse things with its engineering talent. In the book, the important bit is the packaging line: “designed in Rhode Island, by people.” The company used it before generative AI turned human authorship into something that might require labelling. The object is gloriously unserious. Real people sat in a room, considered the ancient question “Should the axolotl hold the tea?” and decided yes. That tiny decision carries humour, taste and a source somebody can name. The label does not promise greatness. Humans make rubbish too. It promises provenance: somebody was there, chose this particular absurdity and can stand behind it. The machine may reproduce the shape. It did not have the meeting.

The resource curse is the pattern in which great natural wealth can leave a country’s people with less power rather than more.

Oil and diamonds bring money straight out of the ground. A state that can fund itself without relying as heavily on citizens’ work or taxes has less practical reason to listen to them. The resource becomes valuable; the people become optional in the arithmetic.

*Human Still Wins* uses the pattern to explain the intelligence curse. If machine intelligence begins producing more of the value, the ownership question becomes larger than who uses the best tool. It becomes who owns the source, who receives the return and who no longer needs to be consulted.

The rising tide is the book’s image for how AI reaches work. A wave suggests one violent arrival, a peak and then a retreat. A tide needs no single dramatic moment. It keeps coming, task by task, until the room is different and your shoes are wet.

That distinction changes the useful response. You do not have to predict the exact day the wave hits. You have to notice what is already becoming faster, cheaper or easier, and move before standing still becomes the decision.

The tide is not a promise that everybody floats. Some roles shrink, some ladders lose their bottom rungs and some people absorb more of the cost than others. Its one mercy is visibility. Gradual change can still be seen. Seeing it is the opportunity.

Secret cyborgs are employees who use AI extensively but keep the assistance quiet. Ethan Mollick’s term captures the most capable users hiding the very practice that makes them capable, often because admitting how much the machine contributed feels uncomfortably close to admitting that the task might be replaceable.

The secrecy leaves everybody else with the polished result and none of the method: no prompts, corrections, failed drafts or judgement calls. Organisations cannot learn from work they are not allowed to see being made.

It also helps poor habits spread. The careful users hide their checking. The careless users hide that they did not check. Soon the whole office is working with the machine and pretending it is not.

Social presence is the sense that another person is properly there with you in a mediated exchange. Their camera may be off. Their face may be twelve centimetres high and paused on an expression they would never voluntarily hold. Still, they have arrived.

An online-status dot can remain green from breakfast until supper while the human behind it contributes roughly the companionship of a printer cartridge. Social presence begins when the exchange is allowed to interfere with both people: somebody notices the hesitation under “all good”, changes the question, offers the missing context, or carries what was said into the next decision instead of letting it die politely in the transcript.

Researchers use the term for the feeling of togetherness that can form across distance. Stephanie Tietz, Evi Kneisel and Katja Werner examined 148 successful and unsuccessful knowledge exchanges in virtual teams. Social presence appeared far more often in the successful ones. The knowledge travelled better when the people felt that other people, rather than merely messages, were present.

Remote work did not make this less important. It removed the office’s habit of leaking context for free by the kettle, in the corridor and on the walk back from the meeting. A remote team must create more of those chances deliberately.

In *Human Still Wins*, presence is not proof of a body in a room. It is the moment the situation gets inside you far enough to change what you do next. A connection has been established. Social presence is when somebody actually comes through it.

The Soul Advantage is the capacity to care about something enough that being wrong about it costs you something real.

A machine can produce the language of concern, regret or conviction. It does not sit in the wreckage when the decision goes badly. A person chooses, carries the consequence and can decide to repair what happened. That is agency, with the invoice attached.

The advantage is not proof that humans are always wiser, kinder or more tasteful. We have supplied ample evidence to the contrary. It is that a human can be responsible. Trust, judgement, taste and courage grow from that fact. The privilege is not merely that we get to choose. It is that the choice remains ours after the applause has stopped.

Tacit knowledge is the understanding built by doing the work beside someone who already knows how. It is not the procedure in the manual. It is the reason the experienced person knows the procedure will fail on this particular Tuesday.

You learn it in the gap between the written rule and the live situation: what a client’s silence means, when a spreadsheet that balances still smells wrong, which problem needs escalating before the dashboard turns red.

This is why the disappearance of junior work creates a problem larger than fewer entry-level jobs. The repetitive task was also the room where judgement travelled from one generation to the next. Tacit knowledge does not transfer neatly in a document. It passes through shared work.

The cloud is somebody else’s building.

It is a useful name for computing delivered over the internet, and an excellent name if the aim is to make warehouses full of servers sound as though they drift above us without plumbing. They do not. The servers use electricity, produce heat and need cooling. The building sits in a town. Somebody pays the water bill.

The phrase matters in this book because digital convenience is easy to imagine as weightless. A prompt disappears into a clean little box and an answer returns. The physical system behind it remains out of sight, which is not the same as absent. The cloud has a roof, a power connection and neighbours.

The Soul Skills are trust, judgement, taste and courage: four human muscles that become more valuable as competent output becomes cheap. Trust is not charm. It is the evidence that you will be reliable, clear and willing to repair what goes wrong. Judgement finds the real decision when the brief is muddled or the information incomplete. Taste knows what to keep when the machine can give you two hundred perfectly acceptable options before breakfast. Courage makes the call and stays in the room for what follows. None of these is mystical, despite the name. They can be practised, observed and strengthened. Responsibility sits underneath all four. AI can support the work and imitate the language around it. It cannot become accountable for the relationship, the standard or the consequence. That is the point. Intelligence is becoming abundant. The person who can use it without surrendering authorship becomes more valuable, not less.

The tokoloshe is a small, often dangerous spirit in Southern African folklore, associated with night-time mischief, fear and harm. Accounts differ, as folklore tends to, but one familiar defence is to raise the bed on bricks so the creature cannot reach the sleeper.

That detail belongs in this book because superstition survives quite comfortably beside disbelief. A person can insist the tokoloshe is nonsense and still prefer the bed a little higher. The body keeps old instructions long after the mouth has rejected them.

The term is a reminder for the missing Chapter 13: inherited patterns, irrational systems and the odd things people do because somebody before them did them first.

Training data is the material from which an AI model learns its patterns: writing, images, code, recordings and other examples gathered before the model is asked to produce anything of its own.

The composition of that material leaves fingerprints. What appears often becomes easier to reproduce. What is rare, local, awkward or badly represented is easier to mishandle. The model does not know that the missing edge was important. It only knows what the data made available.

The book returns to training data because human work can flow back into the next machine, along with machine-made work nobody checked. That is where slop stops being merely annoying. At scale, today’s indifferent output can become tomorrow’s lesson.

Triskaidekaphobia is the fear or superstition surrounding the number thirteen, often associated with bad luck or misfortune in various cultures and individuals.

It is the reason buildings skip a floor, and the reason this site has a chapter that is not printed in the book. The lift panel knows how to count. It also knows people.

The TRUST Check is the five-part review for AI-generated work that is about to leave your desk.

True: could you stand behind every fact if challenged? Relevant: does it answer the question asked? Understandable: can the recipient act on it? Safe: are the risks ones you are prepared to own? Take ownership: is your name, implicitly or explicitly, on this?

The test is deliberately ordinary. It does not require a governance committee, a new dashboard or a branded lanyard. It requires the person sending the work to read it as though consequences exist. If it fails one of the five, it is not ready.

The trust economy begins when information stops being scarce.

AI can produce writing, analysis and presentation polish almost instantly. Sounding informed is no longer much of a moat. The scarce thing becomes the person whose judgement can be believed, whose sources can be traced and whose promise still means something after the document leaves the room.

Trust is not a soft extra added after the useful work. It changes who gets hired, funded, promoted and given the consequential decision. As synthetic output becomes harder to identify by appearance, provenance helps establish where something came from. The rest still comes down to a human or institution willing to stand near it and answer.

Workslop is AI-made work that looks finished until you touch it. The headings are tidy, the grammar is respectable and the actual thinking has been left for the recipient to discover, usually while muttering at the screen. The sender saves time by handing over a polished shell. The colleague, client or teacher must work out what was meant, verify the claims, repair the missing context or start again. The labour has not disappeared. It has changed desks. *Human Still Wins* describes the feeling before naming the term: a technically sound article that is somehow hollow because nobody cared enough to author it. That reaction is not preciousness about every sentence being handmade. It is annoyance at an unfair hand-off. Workslop damages productivity, yes. More quietly, it damages trust. After enough of it, you stop wondering whether the work is correct and start wondering whether the sender has even read it.