{"course":{"slug":"ai-actually-explained","title":"AI, Actually Explained","url":"https://addaly.com/learn/ai-actually-explained","level":"beginner","track":"start"},"lesson":{"slug":"spotting-ai-in-a-product","title":"How to tell whether a model is involved","url":"https://addaly.com/learn/ai-actually-explained/spotting-ai-in-a-product","position":19,"of":92,"minutes":8,"answer":"Varying wording, a mistakes disclaimer, a feedback widget and word-by-word latency mark a generative model, but the useful questions are what it predicts, what a mistake costs, and who pays for it.","keyIdea":"Varying wording, a mistakes disclaimer, a feedback widget and word-by-word latency mark a generative model, but the useful questions are what it predicts, what a mistake costs, and who pays for it."},"body":{"format":"markdown","text":"## Two opposite problems\n\nThere is the product labelled AI that is a set of `if` statements, and there is the model quietly deciding something important with no label at all. The second is more common and matters more.\n\nNo regulator polices the word, so you have to read the behaviour. Here is how.\n\n## Signs that a trained model is involved\n\n**The output varies.** Ask the same thing twice and get two differently worded answers, and you are looking at a generative model. Written rules do not paraphrase themselves.\n\n**There is a disclaimer.** \"AI can make mistakes, check important information\" is not a courtesy. It appears where a company's lawyers know the output is generated rather than retrieved, and it is one of the most reliable markers there is.\n\n**There is a feedback control.** Thumbs up and down, \"was this helpful\", \"report this suggestion\". Nobody puts a rating widget on a database lookup.\n\n**It is uneven in a particular way.** It handles the hard case beautifully and fails on something trivial. Written rules fail at the edges; models fail wherever their training was thin, and that distribution does not match human intuition about difficulty.\n\n**It degrades with connection.** If a feature stops working offline while the rest of the app continues, it is calling a model on a server. Features that keep working offline are running a small model on the device: a keyboard, face grouping in a photo gallery, offline dictation.\n\n**Latency in the wrong place.** A one to three second pause before text starts appearing, then text arriving word by word, is a language model generating. Instant, complete responses are retrieval.\n\n## Signs that it is not\n\nExactly repeated wording every time. A fixed set of possible outputs. Behaviour that changes only when the app updates. An explanation that traces cleanly to stated rules: \"your application was refused because your income is below X\" with X published. None of these prove anything on their own, but together they point at ordinary software.\n\n## The label is starting to be required, in some places\n\nUntil recently, disclosure was entirely voluntary. That is changing unevenly.\n\nThe EU's AI Act carries transparency obligations: people must be told when they are interacting with an AI system rather than a person, and synthetic image, audio and video content must be marked in a machine-readable way. Those provisions phase in during 2026. China has required labelling of synthetically generated content since 2023, tightened in 2025. Several US states have laws covering political advertising and intimate imagery specifically.\n\nTwo cautions before you rely on this. The rules bind providers operating in those jurisdictions, so an app used elsewhere may carry no such duty. And a machine-readable mark is not a guarantee: it can be stripped by re-encoding, screenshotting or a crop. A missing label will increasingly mean less than a present one.\n\n## The three questions that matter more than the label\n\nOnce you suspect a model, the label stops being interesting. Ask these instead, in this order.\n\n**1. What is it predicting?** Not what it is for. What quantity is it actually trying to get right? A feed predicts engagement, not quality. A hiring filter predicts resemblance to previous hires, not job performance. A support bot predicts a plausible answer, not a correct one. The gap between what it predicts and what you want is where all the disappointment lives.\n\n**2. What happens when it is wrong?** Does it cost a second, a phone call, or a livelihood? Is there a human in the path, and can you reach them? This determines how much scepticism the system deserves, and it varies more than the technology does.\n\n**3. Who bears the cost of the error?** If the operator bears it, expect the system to be carefully tuned. If you bear it, expect the threshold to be set where it is cheapest for them. This is not cynicism; it is what optimising a measurable objective does when your inconvenience is not in the objective.\n\n## An exercise worth doing once\n\nFor one week, keep a note on your phone. Every time you notice something being decided or generated for you, write one line: what it was, and whether it cost you anything.\n\nMost people's list comes back at somewhere between fifteen and forty entries, and nearly all of them are unlabelled. Feed order. Autocorrect. Which emails were prioritised. A card declined. A price that differed from your friend's. A route. A photograph enhanced. A CV filtered before a human saw it.\n\nThe value of the exercise is not the count. It is that you stop thinking of AI as something you go to a website to use, and start seeing it as an infrastructure you are already inside, which is the accurate picture, and the one that makes the rest of this course useful rather than abstract."},"check":{"question":"An insurance app's help feature answers questions in slightly different wording each time, pauses for two seconds, streams its reply word by word, and carries a note saying responses may be inaccurate. What does this combination indicate, and what follows?","options":[{"text":"A search index over the company's policy documents, so the answers are quotations and can be relied on.","correct":false,"tempting":"Reads the streaming and variation as retrieval; a search index returns identical text for identical queries."},{"text":"A rules engine with randomised phrasing added for a friendlier tone, so the underlying answers are authoritative.","correct":false,"tempting":"Explains the variation with a cosmetic feature, which would not require the disclaimer or the generation delay."},{"text":"A generative model producing plausible answers, so a claim about your cover needs confirming against the policy document.","correct":true},{"text":"A human support agent typing, since the pause and word-by-word delivery match someone composing a reply.","correct":false,"tempting":"Mistakes token streaming for typing; a human would not produce a machine-generated inaccuracy disclaimer."}]},"prev":{"slug":"factories-fields-and-clinics","title":"The uses that are not on a screen","url":"https://addaly.com/learn/ai-actually-explained/factories-fields-and-clinics"},"next":{"slug":"what-a-chatbot-does","title":"What a chatbot is doing when it answers","url":"https://addaly.com/learn/ai-actually-explained/what-a-chatbot-does"},"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 tell whether a model is involved', in AI, Actually Explained. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-actually-explained/spotting-ai-in-a-product (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 tell whether a model is involved","courseTitle":"AI, Actually Explained","module":{"index":2,"slug":"where-it-already-is","title":"Where it already is, without a label"},"answer":null,"keyIdea":"Varying wording, a mistakes disclaimer, a feedback widget and word-by-word latency mark a generative model, but the useful questions are what it predicts, what a mistake costs, and who pays for it.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"two-opposite-problems","runs":[{"text":"Two opposite problems"}]},{"t":"p","runs":[{"text":"There is the product labelled AI that is a set of "},{"text":"if","code":true},{"text":" statements, and there is the model quietly deciding something important with no label at all. The second is more common and matters more."}]},{"t":"p","runs":[{"text":"No regulator polices the word, so you have to read the behaviour. Here is how."}]},{"t":"h2","id":"signs-that-a-trained-model-is-involved","runs":[{"text":"Signs that a trained model is involved"}]},{"t":"p","runs":[{"text":"The output varies.","b":true},{"text":" Ask the same thing twice and get two differently worded answers, and you are looking at a generative model. Written rules do not paraphrase themselves."}]},{"t":"p","runs":[{"text":"There is a disclaimer.","b":true},{"text":" “AI can make mistakes, check important information” is not a courtesy. It appears where a company’s lawyers know the output is generated rather than retrieved, and it is one of the most reliable markers there is."}]},{"t":"p","runs":[{"text":"There is a feedback control.","b":true},{"text":" Thumbs up and down, “was this helpful”, “report this suggestion”. Nobody puts a rating widget on a database lookup."}]},{"t":"p","runs":[{"text":"It is uneven in a particular way.","b":true},{"text":" It handles the hard case beautifully and fails on something trivial. Written rules fail at the edges; models fail wherever their training was thin, and that distribution does not match human intuition about difficulty."}]},{"t":"p","runs":[{"text":"It degrades with connection.","b":true},{"text":" If a feature stops working offline while the rest of the app continues, it is calling a model on a server. Features that keep working offline are running a small model on the device: a keyboard, face grouping in a photo gallery, offline dictation."}]},{"t":"p","runs":[{"text":"Latency in the wrong place.","b":true},{"text":" A one to three second pause before text starts appearing, then text arriving word by word, is a language model generating. Instant, complete responses are retrieval."}]},{"t":"h2","id":"signs-that-it-is-not","runs":[{"text":"Signs that it is not"}]},{"t":"p","runs":[{"text":"Exactly repeated wording every time. A fixed set of possible outputs. Behaviour that changes only when the app updates. An explanation that traces cleanly to stated rules: “your application was refused because your income is below X” with X published. None of these prove anything on their own, but together they point at ordinary software."}]},{"t":"h2","id":"the-label-is-starting-to-be-required-in-some-places","runs":[{"text":"The label is starting to be required, in some places"}]},{"t":"p","runs":[{"text":"Until recently, disclosure was entirely voluntary. That is changing unevenly."}]},{"t":"p","runs":[{"text":"The EU’s AI Act carries transparency obligations: people must be told when they are interacting with an AI system rather than a person, and synthetic image, audio and video content must be marked in a machine-readable way. Those provisions phase in during 2026. China has required labelling of synthetically generated content since 2023, tightened in 2025. Several US states have laws covering political advertising and intimate imagery specifically."}]},{"t":"p","runs":[{"text":"Two cautions before you rely on this. The rules bind providers operating in those jurisdictions, so an app used elsewhere may carry no such duty. And a machine-readable mark is not a guarantee: it can be stripped by re-encoding, screenshotting or a crop. A missing label will increasingly mean less than a present one."}]},{"t":"h2","id":"the-three-questions-that-matter-more-than-the-label","runs":[{"text":"The three questions that matter more than the label"}]},{"t":"p","runs":[{"text":"Once you suspect a model, the label stops being interesting. Ask these instead, in this order."}]},{"t":"p","runs":[{"text":"1. What is it predicting?","b":true},{"text":" Not what it is for. What quantity is it actually trying to get right? A feed predicts engagement, not quality. A hiring filter predicts resemblance to previous hires, not job performance. A support bot predicts a plausible answer, not a correct one. The gap between what it predicts and what you want is where all the disappointment lives."}]},{"t":"p","runs":[{"text":"2. What happens when it is wrong?","b":true},{"text":" Does it cost a second, a phone call, or a livelihood? Is there a human in the path, and can you reach them? This determines how much scepticism the system deserves, and it varies more than the technology does."}]},{"t":"p","runs":[{"text":"3. Who bears the cost of the error?","b":true},{"text":" If the operator bears it, expect the system to be carefully tuned. If you bear it, expect the threshold to be set where it is cheapest for them. This is not cynicism; it is what optimising a measurable objective does when your inconvenience is not in the objective."}]},{"t":"h2","id":"an-exercise-worth-doing-once","runs":[{"text":"An exercise worth doing once"}]},{"t":"p","runs":[{"text":"For one week, keep a note on your phone. Every time you notice something being decided or generated for you, write one line: what it was, and whether it cost you anything."}]},{"t":"p","runs":[{"text":"Most people’s list comes back at somewhere between fifteen and forty entries, and nearly all of them are unlabelled. Feed order. Autocorrect. Which emails were prioritised. A card declined. A price that differed from your friend’s. A route. A photograph enhanced. A CV filtered before a human saw it."}]},{"t":"p","runs":[{"text":"The value of the exercise is not the count. It is that you stop thinking of AI as something you go to a website to use, and start seeing it as an infrastructure you are already inside, which is the accurate picture, and the one that makes the rest of this course useful rather than abstract."}]}],"check":{"question":[{"text":"An insurance app’s help feature answers questions in slightly different wording each time, pauses for two seconds, streams its reply word by word, and carries a note saying responses may be inaccurate. What does this combination indicate, and what follows?"}],"options":[{"text":[{"text":"A search index over the company’s policy documents, so the answers are quotations and can be relied on."}],"correct":false,"tempting":[{"text":"Reads the streaming and variation as retrieval; a search index returns identical text for identical queries."}]},{"text":[{"text":"A rules engine with randomised phrasing added for a friendlier tone, so the underlying answers are authoritative."}],"correct":false,"tempting":[{"text":"Explains the variation with a cosmetic feature, which would not require the disclaimer or the generation delay."}]},{"text":[{"text":"A generative model producing plausible answers, so a claim about your cover needs confirming against the policy document."}],"correct":true,"tempting":null},{"text":[{"text":"A human support agent typing, since the pause and word-by-word delivery match someone composing a reply."}],"correct":false,"tempting":[{"text":"Mistakes token streaming for typing; a human would not produce a machine-generated inaccuracy disclaimer."}]}],"order":[1,0,3,2],"why":null},"prev":{"slug":"factories-fields-and-clinics","title":"The uses that are not on a screen","minutes":9},"next":{"slug":"what-a-chatbot-does","title":"What a chatbot is doing when it answers","minutes":8}}}