{"course":{"slug":"ai-actually-explained","title":"AI, Actually Explained","url":"https://addaly.com/learn/ai-actually-explained","level":"beginner","track":"start"},"lesson":{"slug":"already-around-you","title":"Where it already is, without the label","url":"https://addaly.com/learn/ai-actually-explained/already-around-you","position":10,"of":92,"minutes":6,"answer":"Most AI in your life is unlabelled, and what matters is who pays when it is wrong.","keyIdea":"Most AI in your life is unlabelled, and what matters is who pays when it is wrong."},"body":{"format":"markdown","text":"## None of this was called AI when it shipped\n\nYou have been using trained models for years. They were not marketed as AI because at the time that word was reserved for robots in films.\n\n**Your bank declining a card.** A model trained on hundreds of millions of past transactions, each labelled fraud or not fraud. It scores your transaction in a few milliseconds. Notice how it fails: it declines your perfectly real payment on your first day in a new country. The bank tolerates that because a wrong decline costs a phone call, and a missed fraud costs money. The error rate was chosen, not accidental.\n\n**Searching your photos for beach.** Nobody tagged your photos. A model trained on millions of labelled images learned what beaches tend to look like, and it runs over your library. It will also confidently show you a photo of a car park by the sea.\n\n**Your keyboard suggesting the next word.** This is the same basic idea as a chatbot, scaled down by a factor of millions and running on your phone. If you understand your keyboard, you are most of the way to understanding lesson five.\n\n**Maps changing your route.** Part written rules, part trained model. Finding a path through the road network is ordinary code. Predicting that this road will take 34 minutes at 6pm on a Friday comes from historical and live traffic data.\n\n**Turning a voice note into text.** Trained on enormous amounts of recorded speech paired with transcripts. Try it in a language with fewer speakers and watch the quality drop, for exactly the reason in the last lesson.\n\n**What your feed shows you next.** A model predicting which items you are most likely to engage with, ranked. Worth pausing on: engagement is a target that people chose. A different target would produce a different feed. The model has no view about what is good for you, because nobody made that the thing it was nudged toward.\n\n## The question that separates these\n\nEvery one of those systems is wrong regularly. Some of that is fine and some is not, and accuracy alone does not tell you which is which.\n\nAsk instead: **when it is wrong, who pays, and can they do anything about it?**\n\nA bad song recommendation costs you thirty seconds and you skip it. A wrong route costs you twenty minutes. Both are recoverable by the person affected, immediately.\n\nNow the same technology decides whether your loan is approved, whether your visa application is flagged, whether your exam script is marked as cheating, whether your name goes on a watchlist. The error rate might be lower than the music app's. It does not matter. The cost lands on one person, that person often cannot see that a model was involved, and there may be nowhere to appeal.\n\nThat is the line worth learning. Not whether a system is AI, and not even how accurate it is, but whether the person carrying the cost of the mistake can see it and undo it."},"check":{"question":"A video app's recommender and a court's bail-risk system are both trained models, and both are wrong roughly one time in twenty. What best explains why one is broadly acceptable and the other is not?","options":[{"text":"The cost of the error lands differently. A wrong recommendation costs the viewer seconds and they can undo it themselves; a wrong bail score costs a person their liberty and they may not even know a model was involved.","correct":true},{"text":"The recommender is not really AI, just a ranking system, so the comparison does not hold.","correct":false,"tempting":"People reserve the word AI for chatbots and robots, which quietly exempts the familiar systems from any scrutiny at all."},{"text":"One in twenty is an acceptable rate anywhere, since human judges are wrong more often than that.","correct":false,"tempting":"The comparison to human error is a real and often correct argument, which makes it easy to stop there and skip the questions of who is harmed and whether they can appeal."},{"text":"The court needs a better model. At 99.9 percent accuracy the objection would go away.","correct":false,"tempting":"It reframes a question about power and appeal as a purely technical problem, which is comfortable because technical problems have known fixes."}]},"prev":{"slug":"the-words-people-use-wrongly","title":"Five words that hide more than they say","url":"https://addaly.com/learn/ai-actually-explained/the-words-people-use-wrongly"},"next":{"slug":"the-feed-that-learns-you","title":"The feed is a prediction machine","url":"https://addaly.com/learn/ai-actually-explained/the-feed-that-learns-you"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'Where it already is, without the label', in AI, Actually Explained. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-actually-explained/already-around-you (Accessed: 31 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"Where it already is, without the label","courseTitle":"AI, Actually Explained","module":{"index":2,"slug":"where-it-already-is","title":"Where it already is, without a label"},"answer":null,"keyIdea":"Most AI in your life is unlabelled, and what matters is who pays when it is wrong.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"none-of-this-was-called-ai-when-it-shipped","runs":[{"text":"None of this was called AI when it shipped"}]},{"t":"p","runs":[{"text":"You have been using trained models for years. They were not marketed as AI because at the time that word was reserved for robots in films."}]},{"t":"p","runs":[{"text":"Your bank declining a card.","b":true},{"text":" A model trained on hundreds of millions of past transactions, each labelled fraud or not fraud. It scores your transaction in a few milliseconds. Notice how it fails: it declines your perfectly real payment on your first day in a new country. The bank tolerates that because a wrong decline costs a phone call, and a missed fraud costs money. The error rate was chosen, not accidental."}]},{"t":"p","runs":[{"text":"Searching your photos for beach.","b":true},{"text":" Nobody tagged your photos. A model trained on millions of labelled images learned what beaches tend to look like, and it runs over your library. It will also confidently show you a photo of a car park by the sea."}]},{"t":"p","runs":[{"text":"Your keyboard suggesting the next word.","b":true},{"text":" This is the same basic idea as a chatbot, scaled down by a factor of millions and running on your phone. If you understand your keyboard, you are most of the way to understanding lesson five."}]},{"t":"p","runs":[{"text":"Maps changing your route.","b":true},{"text":" Part written rules, part trained model. Finding a path through the road network is ordinary code. Predicting that this road will take 34 minutes at 6pm on a Friday comes from historical and live traffic data."}]},{"t":"p","runs":[{"text":"Turning a voice note into text.","b":true},{"text":" Trained on enormous amounts of recorded speech paired with transcripts. Try it in a language with fewer speakers and watch the quality drop, for exactly the reason in the last lesson."}]},{"t":"p","runs":[{"text":"What your feed shows you next.","b":true},{"text":" A model predicting which items you are most likely to engage with, ranked. Worth pausing on: engagement is a target that people chose. A different target would produce a different feed. The model has no view about what is good for you, because nobody made that the thing it was nudged toward."}]},{"t":"h2","id":"the-question-that-separates-these","runs":[{"text":"The question that separates these"}]},{"t":"p","runs":[{"text":"Every one of those systems is wrong regularly. Some of that is fine and some is not, and accuracy alone does not tell you which is which."}]},{"t":"p","runs":[{"text":"Ask instead: "},{"text":"when it is wrong, who pays, and can they do anything about it?","b":true}]},{"t":"p","runs":[{"text":"A bad song recommendation costs you thirty seconds and you skip it. A wrong route costs you twenty minutes. Both are recoverable by the person affected, immediately."}]},{"t":"p","runs":[{"text":"Now the same technology decides whether your loan is approved, whether your visa application is flagged, whether your exam script is marked as cheating, whether your name goes on a watchlist. The error rate might be lower than the music app’s. It does not matter. The cost lands on one person, that person often cannot see that a model was involved, and there may be nowhere to appeal."}]},{"t":"p","runs":[{"text":"That is the line worth learning. Not whether a system is AI, and not even how accurate it is, but whether the person carrying the cost of the mistake can see it and undo it."}]}],"check":{"question":[{"text":"A video app’s recommender and a court’s bail-risk system are both trained models, and both are wrong roughly one time in twenty. What best explains why one is broadly acceptable and the other is not?"}],"options":[{"text":[{"text":"The cost of the error lands differently. A wrong recommendation costs the viewer seconds and they can undo it themselves; a wrong bail score costs a person their liberty and they may not even know a model was involved."}],"correct":true,"tempting":null},{"text":[{"text":"The recommender is not really AI, just a ranking system, so the comparison does not hold."}],"correct":false,"tempting":[{"text":"People reserve the word AI for chatbots and robots, which quietly exempts the familiar systems from any scrutiny at all."}]},{"text":[{"text":"One in twenty is an acceptable rate anywhere, since human judges are wrong more often than that."}],"correct":false,"tempting":[{"text":"The comparison to human error is a real and often correct argument, which makes it easy to stop there and skip the questions of who is harmed and whether they can appeal."}]},{"text":[{"text":"The court needs a better model. At 99.9 percent accuracy the objection would go away."}],"correct":false,"tempting":[{"text":"It reframes a question about power and appeal as a purely technical problem, which is comfortable because technical problems have known fixes."}]}],"order":[2,3,0,1],"why":null},"prev":{"slug":"the-words-people-use-wrongly","title":"Five words that hide more than they say","minutes":8},"next":{"slug":"the-feed-that-learns-you","title":"The feed is a prediction machine","minutes":9}}}