Blog
AI Can Help Write Alt Text. Publishers Must Still Own It.
AI can draft alt text in seconds. To draft it well, it needs your book. Before any publisher presses upload, there are questions about data, copyright and responsibility that deserve a clear answer.
In short
- AI can draft useful alt text for simple images, but good descriptions depend on the book's context, and that means giving AI your content.
- "A closed system" can mean very different things. Check where your files go, who keeps copies, and what your author and image contracts allow.
- The law on AI and copyright is unsettled. No current law requires publishers to label AI-written alt text, but the EU's rules turn on human review and editorial responsibility.
- My recommendation: keep a qualified person responsible for every description, use human-led writing for complex and high-risk content, and tell readers how descriptions were made.
At the Frankfurt Book Fair in October, I kept hearing the same two questions in conversations about accessible publishing.
The first was about data. Most people now accept that AI can help write alternative text, but only if it can see what surrounds each image. So what happens to a manuscript once it is uploaded? Who keeps it, who learns from it, and does the publisher give anything away in the process?
The second was about credit. If AI writes the alt text, should readers be told? Should the AI be acknowledged, the way we credit an illustrator or a translator?
Many of the vendors I spoke to had the same answer to the first question: their AI is "a closed system", and your content never leaves it. Some of those claims may well be true. But without detail on how the system works and what the contract says, I could not tell one claim from another, and neither could many of the publishers I spoke to.
I should be open about where I stand. iAccessDigital writes alt text by hand, and we do not upload client titles to AI services, so I have a view and an interest. In this article I look at three questions: can AI write useful alt text, what happens when you give it your book, and who is responsible for what the reader finally receives.
Can AI write useful alt text?
Sometimes, yes. Current models describe simple images well, and when they are given enough surrounding text they can often work out tone and purpose too. I don't think that is seriously disputed any more.
But useful alt text does more than identify an image. WCAG success criterion 1.1.1 asks for a text alternative that "serves the equivalent purpose". In practice, that means giving the reader the information the image was included to convey, at the right level of detail for the book and its audience. And that depends on the book.
Take a photograph of a crowded railway platform in 1939:
- Image only: "A crowd of people on a railway platform."
- In a history of the Second World War: "Children with luggage and name labels wait on a crowded London platform during the September 1939 evacuation."
- In a novel, where the chapter has just described the narrator searching for her brother: "The crowded platform where the narrator last sees her brother, among hundreds of children waiting for the train."
The first is accurate and almost useless. The other two depend on information that is not in the picture. An AI can get there too, but only if it is given the relevant passage, the subject and the audience. So the real question is what the system has access to, what it can reliably infer from that, and what a qualified person must check before the description reaches a reader.
The answer depends on the type of image:
| Image type | Where AI can help | What a person must do | Recommended approach |
|---|---|---|---|
| Decorative image | Flag it as possibly decorative | Confirm it carries no meaning | AI flag, then an editorial check |
| Simple photograph | Produce a first draft | Check accuracy, relevance, tone and length | AI draft, then an editorial check |
| Illustration in fiction or a picture book | Describe visual details | Make sure it serves the story and does not give away what comes next | AI draft and an editor who knows the book; human-led for picture books |
| Chart, graph or infographic | Pick out the visible labels and values | Decide what the data shows and write the point | Human-led |
| Map | Identify labels and features | Explain why the map is there | Human-led |
| STEM diagram or maths | Help with structure | Provide subject knowledge and check the meaning | Human-led, by someone who knows the subject |
For some complex images, alt text alone is not enough. WCAG's own guidance says charts and diagrams may need both a short and a long text alternative. A chart may need a data table; a detailed diagram may need a long description, or a tactile or other alternative format. AI may help flag where that is needed, but deciding what serves the reader best is an editorial decision.
The opposite also happens. Sometimes the surrounding text already says everything the image shows. If the image adds nothing beyond that text, the W3C's alt decision tree says the alt text can be left empty (alt=""), so a screen reader skips it rather than repeating the same information. If the image is informative and its description sits in the text nearby, the alt text should still identify the image and say where the description is, for example "Chart of quarterly sales; described in the paragraph below" (WCAG technique G74). Either way, the alt attribute must be present. Deciding which case applies means reading the page and judging whether the text really does the image's job. A tool that writes a description for every image by default cannot make that call, and when nobody makes it, screen reader users end up hearing the same information twice.
One rule sits above the table. In education, health, legal and public-information titles, I would have every description human-led and checked by a second reviewer, whatever the image. A wrong description in a textbook or a medicines leaflet can mislead a student or a patient.
Speed, cost and quality
In practice, the pressure is time and money. Production schedules are tight, alt text usually comes last, and AI is much cheaper per image than paying a specialist to write descriptions. So speed and cost tend to win over quality. AI can cut both dramatically. Whether its descriptions capture the context and nuance a visually impaired student needs to learn from an image is far less certain.
The fair comparison is not the cost of generating a draft. It is the cost of a description that is accurate, fits the book and is safe to publish: AI drafting plus proper review, set against human-led writing. Once review is properly resourced, and corrections after reader complaints are counted, the saving may be much smaller than vendors suggest. If AI saves time and money on drafting, some of that should go into review, not disappear from the schedule.
It may already be in your workflow
AI alt text can also arrive without anyone deciding to use it. Adobe added AI-generated alt text to InDesign 21.2 in January 2026. Some users report that it can generate descriptions automatically when images are placed, and in February users on Adobe's own forum reported plain backgrounds and logos being described as nude figures. Adobe said in March that its product team was reviewing the behaviour. If your production team uses InDesign, check the settings and make sure you know where your descriptions are coming from.
What happens when you give AI your book?
This is where "a closed system" comes in. The phrase has no agreed technical meaning, and it can describe very different set-ups:
| What the vendor may mean | Does your content leave? | How much it protects you |
|---|---|---|
| The model runs on the vendor's or the publisher's own servers | No | Genuinely in-house. Less common and more expensive, and often uses smaller models |
| A major AI model used through a cloud platform under enterprise terms | Yes, to the cloud provider | Can be strong. AWS, for example, states that model providers on its Bedrock service cannot see customer prompts. Training, retention and location depend on the contract |
| The vendor's own product built on a third-party AI service | Yes, to the vendor and the AI company | Only as strong as both companies' terms |
So when a vendor says "closed system", ask which of these they mean, and ask them to show you where the contract says:
- whether your content is used to train or improve any model
- how long it is kept, and by whom
- which other companies (subprocessors) handle it
- where it is stored, which matters under GDPR if a book contains personal data
- what happens to copies held by the AI provider and the cloud host when you ask for your content to be deleted
Deletion deserves particular attention. When a vendor builds on a third-party AI service and a cloud platform, your content may be copied into each of those systems for security, logging or operational reasons. Deleting it from the vendor's platform does not automatically delete those copies, because each provider follows its own retention terms. Where a book contains personal data, GDPR requires a processor to bind its own subcontractors to the same obligations (Article 28). For everything else, it comes down to the contracts. Ask whether a deletion request reaches every layer, and how long copies can remain.
Even a genuinely closed system only answers the question of where your data goes. It does not tell you whether the alt text is right.
Contracts and licences
The more immediate risks often sit in your own agreements rather than in the AI. Many author contracts were signed before generative AI existed, and some may not cover sharing an author's text with a third-party AI service. Look for clauses that restrict third-party processing, derivative works or machine learning, and remember that newer contracts may address AI directly. Some picture agency licences restrict how images may be used with AI. And alt text is often written before publication, while a title is under embargo. Check all of this before anything is uploaded.
One point is often misunderstood. In the EU, rightsholders can reserve their rights against text and data mining (DSM Directive, Article 4(3)), and the AI Act requires providers of general-purpose AI models to have a policy to respect those reservations (Article 53(1)(c)). That protects you against content being mined from the open web. It does not govern what happens to files you upload yourself. That is a matter for the vendor's contract.
Who owns AI-written alt text?
It depends where you are. In the US, the Copyright Office's January 2025 report concludes that prompts alone do not make someone the author of AI output, although people can own their own creative changes to it. The courts agree that copyright needs a human author (Thaler v. Perlmutter, D.C. Circuit, March 2025; the Supreme Court declined to hear an appeal in March 2026). An editor who chooses which details matter and reshapes a description for the reader is making that kind of contribution. Fixing a typo probably is not.
The UK is different. Section 9(3) of the Copyright, Designs and Patents Act 1988 protects "computer-generated works", but the government's March 2026 report on copyright and AI proposes removing that protection. No court has yet tested how the provision applies to modern AI output, and it may be repealed. Either way, copyright ownership is a weak reason to choose, or to disclose, any particular workflow.
What the US courts have said about training on books
Much of this debate turns on "fair use", a US legal defence that allows some use of copyrighted work without permission. Courts decide it case by case, weighing four factors set out in the Copyright Act: the purpose and character of the use, including whether it is commercial; the nature of the copyrighted work; how much of the work is used; and the effect on the market for the original. The UK and EU have no general fair use defence, only narrower, specific exceptions.
Two US rulings from June 2025 are often quoted, and they point in different directions:
- In Bartz v. Anthropic, Judge Alsup held that training on books was "exceedingly transformative" and a fair use, and so was scanning print books the company had bought. But pirated copies used to build a central library "were not justified by a fair use". The case then settled for $1.5 billion, about $3,000 per book, with final approval in July 2026. Because it settled, the fair use ruling will not be tested on appeal.
- In Kadrey v. Meta, Judge Chhabria ruled for Meta on fair use, but stressed that his ruling "does not stand for the proposition that Meta's use of copyrighted materials to train its language models is lawful", only that "these plaintiffs made the wrong arguments". He added that "in many circumstances it will be illegal to copy copyright-protected works to train generative AI models without permission."
Both are first-instance decisions. They do not bind other courts, and the two judges reasoned differently. The US Copyright Office's pre-publication report on AI training (May 2025) also concluded that training is not automatically fair use (summary from the Authors Guild). The honest position is that the law is unsettled. What is clear is that books have value as training material: HarperCollins, for example, offered authors $2,500 per title in 2024 to license some backlist books to an AI company. A publisher that uploads its list under loose terms may be giving away something others are paying for.
I am not a lawyer, and nothing here is legal advice. If you are considering uploading titles to an AI service, check your contracts and take advice on your own situation.
Who is responsible for what the reader receives?
This is the question I care about most. Readers who rely on alt text cannot see the image to check it. They are trusting the publisher completely, and that trust only works if someone at the publisher stands behind every description.
What the law says, and what it does not
The EU AI Act has a transparency rule for AI-generated text, Article 50(4), which has applied since 2 August 2026. It is narrower than it is sometimes presented. It covers text "published with the purpose of informing the public on matters of public interest", and the duty to disclose falls on whoever deploys the AI to publish it. It does not apply where the content "has undergone a process of human review or editorial control" and a person or organisation "holds editorial responsibility for the publication".
It is not a general labelling rule for alt text. Whether alt text falls within it at all is unclear. The European Commission's guidance on Article 50, published in July 2026, does not mention alt text. It does say what counts as human review: a person with relevant expertise deliberately examining the substance of the content. Superficial or purely formal checks, such as spell-checking, do not count. What the rule does show is how the law thinks about the issue: editorial responsibility, backed by real review, is what matters.
The US has no equivalent for text. California's AI Transparency Act took effect on 2 August 2026 after AB 853 delayed it, and was amended by SB 1000 on 30 September 2026. It now applies to any developer of a publicly available generative AI system, not only large ones, but it still covers images, video and audio, not text. No law I know of currently requires a publisher to label AI-written alt text.
What I recommend
None of the following is required by law. It is what I believe responsible practice looks like:
- A trained editor reads the surrounding text, not just the image.
- They check accuracy, purpose, tone, length and terminology.
- They check for bias and stereotypes. AI descriptions can add assumptions about race, gender, disability, age or body size that neither the image nor the text supports.
- Charts, maps, maths and technical images are checked by someone who knows the subject.
- Education, health, legal and public-information titles get a second reviewer.
- Who reviewed each description, and when, is recorded.
Telling readers how descriptions were made
This is the part of the debate I think the industry has not yet faced. It helps to separate what exists today, what has been proposed and what I would like to see.
What exists today. Accessibility metadata can say that a book has image descriptions. In the EPUB file this is done with values such as alternativeText and longDescription; in ONIX, with Codelist 196 codes 14 and 15 for short and full descriptions. As far as I can find, neither has a way to say who wrote them or how. Both do carry a free-text accessibility summary: accessibilitySummary in EPUB (DAISY guidance), and Codelist 196 code 00 in ONIX.
What has been proposed. This autumn, the W3C AI Content Disclosure Community Group approved, with no objections, a Final Community Group Report dated 9 October 2026, AI Content Disclosure for HTML. It proposes an ai-disclosure attribute that explicitly covers alt text, with three values: human-only; ai-assisted, where AI was involved and a person reviewed the text before publication; and ai-autonomous, where it was published without human review. That line between reviewed and unreviewed output is the same one the EU rule draws. The report is a community proposal, not a W3C standard, and no reading system acts on it yet.
What I recommend. Publishers can say it now, in the accessibility summary. A sentence or two is enough: "All image descriptions were written by our editors", or "Image descriptions in this title were drafted with AI assistance, then reviewed, edited and approved by [Publisher]'s editorial team. Complex charts and diagrams were checked by a subject specialist." I would also like to see a proper way to declare this in the EPUB and ONIX vocabularies, so retailers and libraries can show it alongside the rest of the accessibility information.
Two cautions. First, disclosure is not a substitute for quality. A description that is not fit to publish should not be published, whatever the label says. Second, how a description was made tells you its origin, not whether it is right. A carefully reviewed AI-assisted description can be excellent, and a human-written one can still be wrong.
Still, being open matters. Two well-known universities found that out this autumn. Dartmouth's provost resigned in October after a faculty review found that he "did not sufficiently disclose his AI use to publishers in contravention of those publications' policies", and that his public description of that use understated it (committee report; Inside Higher Ed). Stanford took down advertising banners that used an AI-altered student photo without disclosure, which the university said broke its own policy (Inside Higher Ed). In both cases, the damage came less from using AI than from not saying so.
For universities and university presses, the stakes are higher than for most publishers. Their name on a book is a promise of scholarly care: peer review, careful editing, sources that can be checked. A reader who later learns that the image descriptions in a press's titles were produced by AI, unreviewed and unmentioned in the metadata, may reasonably wonder what else was not checked. Many universities also have legal accessibility duties of their own, from ADA Title II for public universities in the US to the public sector accessibility regulations in the UK. A reputation for rigour takes decades to build. It should not be put at risk by cutting corners for the readers who depend on it most.
Conclusion
AI is not the enemy of accessible publishing. Used carefully, it can find missing descriptions, suggest first drafts for simple images and help keep a large backlist consistent. Some publishers will build good AI-assisted workflows, and I would much rather see that than books with no descriptions at all.
But the hard part of alt text has always been deciding what a reader who cannot see the image needs to know, in this book, at this point in the text. That takes someone who has read the page and who will answer for the result.
So my advice to publishers is simple. Know where your content goes before you upload it. Get the promises in writing. Keep a qualified person responsible for every description. And tell your readers how their descriptions were made.
Small presses cannot always build elaborate workflows, but three safeguards cost very little: never publish AI alt text that nobody has reviewed, keep a simple record of who approved each description, and keep embargoed and unpublished titles out of third-party AI tools.
Accessibility has always been about people, not tools. AI can help us work faster, but it cannot take responsibility for what a reader is told. That part is still ours.
I would be glad to hear from anyone who sees it differently, especially publishers who have tried AI-assisted alt text and measured the results.
This article reflects my personal opinion and provides general information only. It is not legal advice. Publishers should take advice from a qualified lawyer on their own contracts and obligations.