{"course":{"slug":"visual-storytelling","title":"Making It Say Something","url":"https://addaly.com/learn/visual-storytelling","level":"beginner","track":"create"},"lesson":{"slug":"ai-in-the-pipeline-honestly","title":"What the tool can be trusted with, and what it cannot","url":"https://addaly.com/learn/visual-storytelling/ai-in-the-pipeline-honestly","position":62,"of":64,"minutes":10,"answer":"Sort AI tasks by what a mistake costs and whether you would notice.","keyIdea":"Sort AI tasks by what a mistake costs and whether you would notice, because these tools fail fluently and remove the uncertainty signal you normally rely on: disclosure only survives if it is visible in the frame, the copyright position differs by country and is unresolved, and the name on the piece owns every frame."},"body":{"format":"markdown","text":"## Sort the jobs by what a mistake costs\n\nGenerative and machine-learning tools are now in every part of this pipeline. The useful way to think about them is not whether they are good, but **what happens when they are wrong, and whether you would notice.**\n\n**Low cost, error visible**: use freely:\n\n- Transcription of an interview, which you are going to read anyway.\n- Background removal, masking and rotoscoping.\n- Upscaling for a low-resolution source.\n- Rough cuts and shot logging.\n- Twenty headline variants to react against.\n- A first draft of alt text, which you then correct.\n\n**High cost, error invisible**: do not delegate:\n\n- Anything evidential. A generated image is evidence of nothing, and putting one where a reader expects a photograph is a false claim regardless of intent.\n- Facts, figures, names and dates. Fluent text is produced with the same confidence whether or not it is correct, so errors do not look like errors.\n- Translation of anything that matters, where the characteristic failure is a confident reversal of meaning.\n- Captions on sensitive material, where automatic recognition invents a plausible sentence from overlapping speech.\n\nThe pattern: these tools fail **fluently**. A human assistant who did not know something would sound uncertain. This does not, so the usual signal you rely on to know where to check is absent, and you have to check by category instead.\n\n## Disclosure, and what actually survives\n\nLabelling requirements for synthetic media exist in several jurisdictions, with different scope and different timetables (India's rules for synthetic media, China's labelling requirements, the European Union's transparency provisions) and they are being amended as they are implemented. This is not legal advice and the answer for your case depends on where you are and where you publish.\n\nThe durable design fact, independent of any of it: **metadata does not survive a screenshot.** Provenance standards such as C2PA attach a signed record of how an image was made, and that record is stripped the moment somebody photographs their screen, re-uploads, or saves through a tool that does not carry it. If the disclosure matters, it has to be visible in the frame.\n\n## The copyright position is unresolved\n\nTwo separate disagreements are live, and neither has a settled answer.\n\n**Whether training on copyrighted work without permission is lawful.** This is being litigated in several countries with different statutory starting points. The United States argues it through fair use, a flexible test applied case by case. The United Kingdom, India and other fair-dealing systems have enumerated purposes rather than a general test. Japan and Singapore have specific statutory provisions for text and data mining. The European Union has an exception with an opt-out mechanism. These are not different readings of one rule; they are different rules, and cases decided under one say little about another.\n\n**Whether a generated image can be owned.** Several jurisdictions have taken the position that a work requires human authorship, which leaves purely generated output outside copyright and therefore unprotectable: a commercial problem if a client expects exclusivity. Others have reached different conclusions, and the treatment of human-directed and human-edited work is unclear nearly everywhere.\n\nAnyone who tells you this is settled is describing one country at one moment, or selling something. The practical stance for a working designer: **tell the client what is unresolved rather than making a promise you cannot keep**, and if exclusivity matters to the deliverable, say plainly that a generated asset may not be able to carry it.\n\n## The labour question, stated without resolving it\n\nThese tools remove work that used to be entry-level (stock illustration, basic retouching, simple animation, first-draft copy) and entry-level work is how people learned. Whether what replaces it is better, worse or merely different is contested, and confident predictions in either direction are not supported by anything.\n\nWhat is not contested: it is happening, and the people it affects first are the people with the least security. That is worth saying in a course whose audience includes many of them, rather than being smoothed over.\n\n## The rule that does not change\n\n**The person whose name is on the piece is responsible for every frame of it, whatever made the frame.**\n\nNot the tool, not the model, not the supplier. If the caption is wrong, you published a wrong caption. If the hands have six fingers, you shipped six fingers. If the translated line says the opposite of what the client meant, that is your line now.\n\nThis single rule resolves most day-to-day questions without needing to know the law, because it puts the checking back where it belongs: on you, before it goes out, at full size, in the language it will be read in."},"check":{"question":"Why is C2PA provenance metadata an incomplete solution for disclosing that an image was generated?","options":[{"text":"It is attached to the file, so it is lost the moment someone screenshots or re-uploads the image.","correct":true},{"text":"Most platforms do not yet read the standard, so the signature is present but never displayed to viewers.","correct":false,"tempting":"Describes a temporary adoption gap that wider support would close, whereas the underlying weakness persists no matter how many platforms adopt it."},{"text":"The signature can be forged by anyone with the generating tool, so a viewer cannot rely on its absence meaning anything.","correct":false,"tempting":"Misstates how signed provenance works; the cryptographic record resists forgery, and the practical failure lies elsewhere."},{"text":"Provenance records describe the editing history rather than the origin, so they cannot establish whether an image was generated or captured.","correct":false,"tempting":"Misdescribes the standard's contents; the record is designed to cover origin as well as subsequent edits."}]},"prev":{"slug":"representing-people-you-are-not","title":"Somebody is in this piece, and it is not you","url":"https://addaly.com/learn/visual-storytelling/representing-people-you-are-not"},"next":{"slug":"learning-by-deconstruction","title":"Take apart the thing that worked on you","url":"https://addaly.com/learn/visual-storytelling/learning-by-deconstruction"},"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) 'What the tool can be trusted with, and what it cannot', in Making It Say Something. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/visual-storytelling/ai-in-the-pipeline-honestly (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":"What the tool can be trusted with, and what it cannot","courseTitle":"Making It Say Something","module":{"index":8,"slug":"judgement-and-the-body-of-work","title":"Judgement, and the work that accumulates"},"answer":null,"keyIdea":"Sort AI tasks by what a mistake costs and whether you would notice, because these tools fail fluently and remove the uncertainty signal you normally rely on: disclosure only survives if it is visible in the frame, the copyright position differs by country and is unresolved, and the name on the piece owns every frame.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"sort-the-jobs-by-what-a-mistake-costs","runs":[{"text":"Sort the jobs by what a mistake costs"}]},{"t":"p","runs":[{"text":"Generative and machine-learning tools are now in every part of this pipeline. The useful way to think about them is not whether they are good, but "},{"text":"what happens when they are wrong, and whether you would notice.","b":true}]},{"t":"p","runs":[{"text":"Low cost, error visible","b":true},{"text":": use freely:"}]},{"t":"ul","items":[[{"text":"Transcription of an interview, which you are going to read anyway."}],[{"text":"Background removal, masking and rotoscoping."}],[{"text":"Upscaling for a low-resolution source."}],[{"text":"Rough cuts and shot logging."}],[{"text":"Twenty headline variants to react against."}],[{"text":"A first draft of alt text, which you then correct."}]]},{"t":"p","runs":[{"text":"High cost, error invisible","b":true},{"text":": do not delegate:"}]},{"t":"ul","items":[[{"text":"Anything evidential. A generated image is evidence of nothing, and putting one where a reader expects a photograph is a false claim regardless of intent."}],[{"text":"Facts, figures, names and dates. Fluent text is produced with the same confidence whether or not it is correct, so errors do not look like errors."}],[{"text":"Translation of anything that matters, where the characteristic failure is a confident reversal of meaning."}],[{"text":"Captions on sensitive material, where automatic recognition invents a plausible sentence from overlapping speech."}]]},{"t":"p","runs":[{"text":"The pattern: these tools fail "},{"text":"fluently","b":true},{"text":". A human assistant who did not know something would sound uncertain. This does not, so the usual signal you rely on to know where to check is absent, and you have to check by category instead."}]},{"t":"h2","id":"disclosure-and-what-actually-survives","runs":[{"text":"Disclosure, and what actually survives"}]},{"t":"p","runs":[{"text":"Labelling requirements for synthetic media exist in several jurisdictions, with different scope and different timetables (India’s rules for synthetic media, China’s labelling requirements, the European Union’s transparency provisions) and they are being amended as they are implemented. This is not legal advice and the answer for your case depends on where you are and where you publish."}]},{"t":"p","runs":[{"text":"The durable design fact, independent of any of it: "},{"text":"metadata does not survive a screenshot.","b":true},{"text":" Provenance standards such as C2PA attach a signed record of how an image was made, and that record is stripped the moment somebody photographs their screen, re-uploads, or saves through a tool that does not carry it. If the disclosure matters, it has to be visible in the frame."}]},{"t":"h2","id":"the-copyright-position-is-unresolved","runs":[{"text":"The copyright position is unresolved"}]},{"t":"p","runs":[{"text":"Two separate disagreements are live, and neither has a settled answer."}]},{"t":"p","runs":[{"text":"Whether training on copyrighted work without permission is lawful.","b":true},{"text":" This is being litigated in several countries with different statutory starting points. The United States argues it through fair use, a flexible test applied case by case. The United Kingdom, India and other fair-dealing systems have enumerated purposes rather than a general test. Japan and Singapore have specific statutory provisions for text and data mining. The European Union has an exception with an opt-out mechanism. These are not different readings of one rule; they are different rules, and cases decided under one say little about another."}]},{"t":"p","runs":[{"text":"Whether a generated image can be owned.","b":true},{"text":" Several jurisdictions have taken the position that a work requires human authorship, which leaves purely generated output outside copyright and therefore unprotectable: a commercial problem if a client expects exclusivity. Others have reached different conclusions, and the treatment of human-directed and human-edited work is unclear nearly everywhere."}]},{"t":"p","runs":[{"text":"Anyone who tells you this is settled is describing one country at one moment, or selling something. The practical stance for a working designer: "},{"text":"tell the client what is unresolved rather than making a promise you cannot keep","b":true},{"text":", and if exclusivity matters to the deliverable, say plainly that a generated asset may not be able to carry it."}]},{"t":"h2","id":"the-labour-question-stated-without-resolving-it","runs":[{"text":"The labour question, stated without resolving it"}]},{"t":"p","runs":[{"text":"These tools remove work that used to be entry-level (stock illustration, basic retouching, simple animation, first-draft copy) and entry-level work is how people learned. Whether what replaces it is better, worse or merely different is contested, and confident predictions in either direction are not supported by anything."}]},{"t":"p","runs":[{"text":"What is not contested: it is happening, and the people it affects first are the people with the least security. That is worth saying in a course whose audience includes many of them, rather than being smoothed over."}]},{"t":"h2","id":"the-rule-that-does-not-change","runs":[{"text":"The rule that does not change"}]},{"t":"p","runs":[{"text":"The person whose name is on the piece is responsible for every frame of it, whatever made the frame.","b":true}]},{"t":"p","runs":[{"text":"Not the tool, not the model, not the supplier. If the caption is wrong, you published a wrong caption. If the hands have six fingers, you shipped six fingers. If the translated line says the opposite of what the client meant, that is your line now."}]},{"t":"p","runs":[{"text":"This single rule resolves most day-to-day questions without needing to know the law, because it puts the checking back where it belongs: on you, before it goes out, at full size, in the language it will be read in."}]}],"check":{"question":[{"text":"Why is C2PA provenance metadata an incomplete solution for disclosing that an image was generated?"}],"options":[{"text":[{"text":"It is attached to the file, so it is lost the moment someone screenshots or re-uploads the image."}],"correct":true,"tempting":null},{"text":[{"text":"Most platforms do not yet read the standard, so the signature is present but never displayed to viewers."}],"correct":false,"tempting":[{"text":"Describes a temporary adoption gap that wider support would close, whereas the underlying weakness persists no matter how many platforms adopt it."}]},{"text":[{"text":"The signature can be forged by anyone with the generating tool, so a viewer cannot rely on its absence meaning anything."}],"correct":false,"tempting":[{"text":"Misstates how signed provenance works; the cryptographic record resists forgery, and the practical failure lies elsewhere."}]},{"text":[{"text":"Provenance records describe the editing history rather than the origin, so they cannot establish whether an image was generated or captured."}],"correct":false,"tempting":[{"text":"Misdescribes the standard’s contents; the record is designed to cover origin as well as subsequent edits."}]}],"order":[2,0,3,1],"why":null},"prev":{"slug":"representing-people-you-are-not","title":"Somebody is in this piece, and it is not you","minutes":9},"next":{"slug":"learning-by-deconstruction","title":"Take apart the thing that worked on you","minutes":9}}}