{"course":{"slug":"ai-actually-explained","title":"AI, Actually Explained","url":"https://addaly.com/learn/ai-actually-explained","level":"beginner","track":"start"},"lesson":{"slug":"how-to-read-the-next-headline","title":"How to read the next headline","url":"https://addaly.com/learn/ai-actually-explained/how-to-read-the-next-headline","position":92,"of":92,"minutes":9,"answer":"Every claim about AI can be sorted with six questions (demo or product, benchmark or task, could or does, who benefits, is there a model in it at all.","keyIdea":"Every claim about AI can be sorted with six questions (demo or product, benchmark or task, could or does, who benefits, is there a model in it at all, and what was the control) and the same six apply to this course."},"body":{"format":"markdown","text":"## The skill this course was for\n\nYou have spent nine modules learning what the thing is, where it fails, what it costs, and what is unsettled. The point of all of it is a single capability: reading the next claim and knowing what it is worth. Six questions do most of the work. Each comes from a lesson you have already had.\n\n## One: demonstration, or product?\n\nIn 2018 a large company demonstrated, on stage, an assistant that phoned a hair salon and booked an appointment in a natural voice. The product that eventually shipped did a fraction of that, in a few cities, with a person listening in. In late 2023 the same company released a video of a model responding to live drawing and objects in real time; it was later confirmed to have been edited from still images and typed prompts. In 2024 a start-up demonstrated an autonomous software engineer completing paid tasks; independent replication found it completed a small minority of them.\n\nNone of these was a lie in the strict sense. Each was a demonstration presented so that it would be read as a deployment. The two winters were made of this. Ask: can I use it, today, on my task, and what does it do when I do?\n\n## Two: benchmark, or task?\n\nThe previous lesson's saturation record is the mechanism. A score on a public test is a score on a public test: it may have been in the training data, it measures what the test measures, and it says nothing about your document. Ask: on what, measured how, and had the model seen it?\n\n## Three: could, or does?\n\nThe exposure studies could have automated eighty per cent of workers' tasks and did not. Every study in the work module found a gap between what a tool could do in a trial and what it did in a job. Ask: is this a description of a capability or of an outcome, and is there a stopwatch anywhere in it?\n\n## Four: who benefits from the claim?\n\nA funding round, a product launch, a share price, a regulatory argument, a book. A lab's claim that its model is dangerous and a lab's claim that its model is safe can both be marketing, and the safety lesson showed you how. This does not make the claim false. It tells you which way to check. Ask: what happens to the speaker if I believe this?\n\n## Five: is there a model in it at all?\n\nThe second module taught you to ask whether a product's behaviour was written or learned. Since 2023 the word has been attached to everything. A securities regulator charged two investment firms in 2024 with claiming AI capabilities they did not have. A software company valued at over a billion dollars collapsed in 2025 after it emerged that much of what it had sold as automated was done by people. A retailer's checkout-free shops, presented as computer vision, were reported in 2024 to depend on a large team in India reviewing the footage. Ask: is this a trained model, a rule, or a person, and how would I know?\n\n## Six: what was the control?\n\nThe support agents gained fourteen per cent *relative to agents without the tool*. The developers were nineteen per cent slower *relative to their own work without it*. A number with no comparison is a number about nothing. Ask: compared with what, and who was in the other group?\n\n## Three more, briefly\n\nWhich model, and when: because the model changed under you. Which language: because it is weaker in yours. And what unit (per query, per day, per token, per task) because the cost and energy lessons showed how one true number is reported as a different one.\n\n## Turn it on this course\n\nThe same six apply here. This course has quoted studies you can find, numbers you can check, and mechanisms you can test on a free model in an afternoon. Where it was uncertain it said so, and where the dates were moving it said that. It has also been written by a model, under a person's direction, and every fact in it is exactly as reliable as the checking that person did. Read it the way it taught you to read everything else.\n\n## Where to go next\n\nIf you want the mechanism at depth, the course on how language models work. If you want to use them well, the course on prompting. If you want to test them properly, the course on evaluating AI. If you want the ethics and the law in full, the course on AI and you. If you want the workplace, the courses on AI at work and AI careers. If you want to build, the course on Python for AI. All free, all here, and all written to the standard this lesson just gave you for reading them."},"check":{"question":"A company releases a video of its model responding in real time to a person's drawings, and the video is later found to have been edited from still images and typed prompts. Which of the six questions does this most directly answer, and how?","options":[{"text":"\"Who benefits?\": the video was made to lift the company's share price ahead of a launch, so its claims must be false.","correct":false,"tempting":"Uses the benefit question to conclude falsity, when it only tells you which way to check."},{"text":"\"Benchmark or task?\": the video was a benchmark result presented as a task.","correct":false,"tempting":"Misfiles it; no benchmark was involved, and the gap was between a staged demonstration and a working product."},{"text":"\"Demo or product?\": the video showed what could be staged, not what a user could do with the shipped system.","correct":true},{"text":"\"Is there a model in it?\": the video was produced by people rather than a model.","correct":false,"tempting":"Overstates the finding; a model did produce the responses, from inputs that were not what the video implied."}]},"prev":{"slug":"the-questions-still-open","title":"Questions nobody has answered yet","url":"https://addaly.com/learn/ai-actually-explained/the-questions-still-open"},"next":null,"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) 'How to read the next headline', in AI, Actually Explained. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-actually-explained/how-to-read-the-next-headline (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"How to read the next headline","courseTitle":"AI, Actually Explained","module":{"index":9,"slug":"who-decides","title":"Who controls it, and what is unsettled"},"answer":null,"keyIdea":"Every claim about AI can be sorted with six questions (demo or product, benchmark or task, could or does, who benefits, is there a model in it at all, and what was the control) and the same six apply to this course.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"the-skill-this-course-was-for","runs":[{"text":"The skill this course was for"}]},{"t":"p","runs":[{"text":"You have spent nine modules learning what the thing is, where it fails, what it costs, and what is unsettled. The point of all of it is a single capability: reading the next claim and knowing what it is worth. Six questions do most of the work. Each comes from a lesson you have already had."}]},{"t":"h2","id":"one-demonstration-or-product","runs":[{"text":"One: demonstration, or product?"}]},{"t":"p","runs":[{"text":"In 2018 a large company demonstrated, on stage, an assistant that phoned a hair salon and booked an appointment in a natural voice. The product that eventually shipped did a fraction of that, in a few cities, with a person listening in. In late 2023 the same company released a video of a model responding to live drawing and objects in real time; it was later confirmed to have been edited from still images and typed prompts. In 2024 a start-up demonstrated an autonomous software engineer completing paid tasks; independent replication found it completed a small minority of them."}]},{"t":"p","runs":[{"text":"None of these was a lie in the strict sense. Each was a demonstration presented so that it would be read as a deployment. The two winters were made of this. Ask: can I use it, today, on my task, and what does it do when I do?"}]},{"t":"h2","id":"two-benchmark-or-task","runs":[{"text":"Two: benchmark, or task?"}]},{"t":"p","runs":[{"text":"The previous lesson’s saturation record is the mechanism. A score on a public test is a score on a public test: it may have been in the training data, it measures what the test measures, and it says nothing about your document. Ask: on what, measured how, and had the model seen it?"}]},{"t":"h2","id":"three-could-or-does","runs":[{"text":"Three: could, or does?"}]},{"t":"p","runs":[{"text":"The exposure studies could have automated eighty per cent of workers’ tasks and did not. Every study in the work module found a gap between what a tool could do in a trial and what it did in a job. Ask: is this a description of a capability or of an outcome, and is there a stopwatch anywhere in it?"}]},{"t":"h2","id":"four-who-benefits-from-the-claim","runs":[{"text":"Four: who benefits from the claim?"}]},{"t":"p","runs":[{"text":"A funding round, a product launch, a share price, a regulatory argument, a book. A lab’s claim that its model is dangerous and a lab’s claim that its model is safe can both be marketing, and the safety lesson showed you how. This does not make the claim false. It tells you which way to check. Ask: what happens to the speaker if I believe this?"}]},{"t":"h2","id":"five-is-there-a-model-in-it-at-all","runs":[{"text":"Five: is there a model in it at all?"}]},{"t":"p","runs":[{"text":"The second module taught you to ask whether a product’s behaviour was written or learned. Since 2023 the word has been attached to everything. A securities regulator charged two investment firms in 2024 with claiming AI capabilities they did not have. A software company valued at over a billion dollars collapsed in 2025 after it emerged that much of what it had sold as automated was done by people. A retailer’s checkout-free shops, presented as computer vision, were reported in 2024 to depend on a large team in India reviewing the footage. Ask: is this a trained model, a rule, or a person, and how would I know?"}]},{"t":"h2","id":"six-what-was-the-control","runs":[{"text":"Six: what was the control?"}]},{"t":"p","runs":[{"text":"The support agents gained fourteen per cent "},{"text":"relative to agents without the tool","i":true},{"text":". The developers were nineteen per cent slower "},{"text":"relative to their own work without it","i":true},{"text":". A number with no comparison is a number about nothing. Ask: compared with what, and who was in the other group?"}]},{"t":"h2","id":"three-more-briefly","runs":[{"text":"Three more, briefly"}]},{"t":"p","runs":[{"text":"Which model, and when: because the model changed under you. Which language: because it is weaker in yours. And what unit (per query, per day, per token, per task) because the cost and energy lessons showed how one true number is reported as a different one."}]},{"t":"h2","id":"turn-it-on-this-course","runs":[{"text":"Turn it on this course"}]},{"t":"p","runs":[{"text":"The same six apply here. This course has quoted studies you can find, numbers you can check, and mechanisms you can test on a free model in an afternoon. Where it was uncertain it said so, and where the dates were moving it said that. It has also been written by a model, under a person’s direction, and every fact in it is exactly as reliable as the checking that person did. Read it the way it taught you to read everything else."}]},{"t":"h2","id":"where-to-go-next","runs":[{"text":"Where to go next"}]},{"t":"p","runs":[{"text":"If you want the mechanism at depth, the course on how language models work. If you want to use them well, the course on prompting. If you want to test them properly, the course on evaluating AI. If you want the ethics and the law in full, the course on AI and you. If you want the workplace, the courses on AI at work and AI careers. If you want to build, the course on Python for AI. All free, all here, and all written to the standard this lesson just gave you for reading them."}]}],"check":{"question":[{"text":"A company releases a video of its model responding in real time to a person’s drawings, and the video is later found to have been edited from still images and typed prompts. Which of the six questions does this most directly answer, and how?"}],"options":[{"text":[{"text":"“Who benefits?”: the video was made to lift the company’s share price ahead of a launch, so its claims must be false."}],"correct":false,"tempting":[{"text":"Uses the benefit question to conclude falsity, when it only tells you which way to check."}]},{"text":[{"text":"“Benchmark or task?”: the video was a benchmark result presented as a task."}],"correct":false,"tempting":[{"text":"Misfiles it; no benchmark was involved, and the gap was between a staged demonstration and a working product."}]},{"text":[{"text":"“Demo or product?”: the video showed what could be staged, not what a user could do with the shipped system."}],"correct":true,"tempting":null},{"text":[{"text":"“Is there a model in it?”: the video was produced by people rather than a model."}],"correct":false,"tempting":[{"text":"Overstates the finding; a model did produce the responses, from inputs that were not what the video implied."}]}],"order":[1,2,3,0],"why":null},"prev":{"slug":"the-questions-still-open","title":"Questions nobody has answered yet","minutes":10},"next":null}}