How GotAlt uses AI
GotAlt uses AI models from Anthropic to compare images with their alt text, to draft alt text and to write fix plans, and the free tools label that output AI-generated. GotAlt never decides whether an image or a text was made by AI: its AI disclosure tools only read what a file or a page declares. Our guides are drafted with AI assistance and checked against their sources, and GotAlt is responsible for everything it publishes.
Published . Last reviewed .
This page describes how GotAlt works on October 22, 2026: where AI models are used, what they are given, how their output is marked, and how the text on this site is written. It is general information, not legal advice. How we source, review and correct our pages is set out in the editorial policy, and what happens to the data you submit is in the privacy policy.
Where GotAlt uses AI models
Every AI feature on this site calls Claude Haiku 5.5, a model from Anthropic. GotAlt does not build or train its own models. A feature that is not in the table below does not call a model.
| Feature | What the model does | What it is given | What you get back |
|---|---|---|---|
| Alt text checker (free) | Compares up to six images on a page with their alt text | The image files, their alt text and the page address | For each image, a verdict, a description of what it shows, the problem and a suggested replacement |
| Deep audit (free) | Does the same for images, and reads the page's link texts and headings | Up to six images with their alt text, the page address, the page title, the link texts and addresses, and the heading outline | Findings, each with the problem and a suggested replacement |
| Alt text generator (free) | Drafts alt text for an image you upload | The image, resized in your browser to at most 1200 pixels on its longest side, and any keywords or page context you type | A draft alt text that you can edit |
| AI fix plan on free scan results | Writes a prioritized plan from the scan's list of issues | The scanned address, the score and grade, and each issue's title, severity, WCAG criterion and count. No images | A plan in plain text |
| Site Image Check and paid monitoring | Compares images with their alt text, as the checker does | The images and their alt text, without the page address | A report and a fix sheet. Each flagged image has what the model says it shows and a draft alt text to check against the image |
| Bulk Alt Text Pack | Checks each image in your product CSV against its alt text and drafts new alt text where it should change | The image, its product title and its current alt text | A CSV of verdicts and drafts to review before you import it |
The rule-based scan does not call a model. Neither do the AI disclosure tools, the AI disclosure laws finder, the AI image metadata checker, the AI image metadata writer, the AI label checker and the AI label generator, or the AI disclosure section of the Site Image Check. They are ordinary code that reads what a file or a page declares.
GotAlt also runs the image check outside the site's tools, on one page of sites it chooses, to research the market or to show a prospective customer a finding on their own page. The model is given what the alt text checker gives it: up to six images with their alt text and the page address. The privacy policy and the GotAltBot page say what is read and what is kept.
How the output of a model is labeled
In the free tools and the signed-in dashboard, every passage that a model wrote carries the words AI-generated, set in front of the text it describes:
- The alt text generator puts the label above each draft, and the draft's field is named "AI-generated alt text for" the file.
- The alt text checker groups the verdict, the description of what the image shows, the problem and the suggestion into one block that opens with the label. Your page's own alt text stays outside the block, so you can tell what your page says from what the model says.
- The deep audit, on its own page and on the home page after a scan, puts the label at the top of each finding.
- The AI fix plan puts the label above the plan.
- In the signed-in dashboard, the issues list groups what the model says an image shows, the problem it found and its suggested alt text into one block that opens with the label and a link to this page. Your page's own alt text stays outside the block.
The label is plain text, placed before what it describes, so a screen reader reads it first and a person using a magnifier or a keyboard meets it in reading order. It is not an image, it does not depend on hover, and it is not carried by color alone: dark text on white with a dark outline, about 19:1 contrast where WCAG asks for 4.5:1. A sentence above each set of results says which kind of model wrote it and links to this page. Our guide on accessible AI labels explains what to look for in a label, and the AI label generator builds one for your own site.
In paid reports and fix sheets, the model's statements are named as the model's: the fix sheet's two columns of the model's words are headed "What the model says the image shows (written by an AI model)" and "Draft alt text written by an AI model (check it against the image before you use it)", a JSON export marks a defect that holds the model's words with "generated_by": "AI model", and suggested alt text is a draft to check against the image before you use it.
The label is our own wording. It is not the European Commission's AI icon, and using it does not mean that GotAlt adheres to the Code of Practice on Transparency of AI-Generated Content or meets any law.
What GotAlt does not do with AI
It does not decide whether an image was made by AI
Anthropic's documentation says that Claude cannot determine whether an image is AI-generated and that you should not rely on it to detect fake or synthetic images (page read October 22, 2026). The vision check is asked one question about an image: whether its alt text is true of what the image shows. It is not asked who or what made the image, and no GotAlt result says an image is, or is not, AI-generated.
Our AI disclosure tools do something different. They read what a file or a page declares: the digital source type recorded in an image's metadata, the presence of content credentials, the wording of a label on a live page. Finding no marker proves nothing about how an image was made, and the tools say so wherever they report that nothing was found.
It does not talk to visitors through a chatbot
gotalt.com has no chatbot or AI assistant that converses with visitors. The AI features above run when you start one, and some of them ask for an email address first, as the privacy policy explains.
It does not use a model to make decisions about people
The models judge images and text on websites. The privacy policy sets out what the AI tools send to Anthropic and how long Anthropic says it keeps it.
How our guides are drafted and checked
The guides, and the law and standards pages beside them, are drafted with AI assistance: a model writes a first draft from a plan and from the primary sources we name. A second pass then opens each source the draft cites and compares the claim with the source's text; a claim the source does not support is corrected or removed. Scripts check the structure of each page, its links, the wording rules in our editorial policy and its dates. Each page shows when it was published and when its citations were last re-checked.
The second pass is carried out by AI models and by scripts. We do not describe it as human editorial review. Every guide, law page, standards page and platform page carries this line directly under its dates, with a link to this page: "Drafted with AI assistance and checked against the sources linked on this page."
If you find an error, email hello@gotalt.com with the page address and, if you can, a link to the source. We reply within two working days (Monday to Friday) and correct the page against the source.
Who is responsible
GotAlt publishes every page on this site and is responsible for it, whether a person or a model drafted it. A label or a note on this site describes how a result was produced. It is not a statement by GotAlt that your site meets or fails a law, and no tool, ours included, can confirm that.
What we don't claim
Models make mistakes. In our own test of 124 labeled alt texts (62 correct and 62 with a planted error), each checked three times, the production setting left about 98% of the correct alt text alone and flagged about 98% of the planted errors (182 and 183 of 186 verdicts each), so some wrong alt text goes unflagged and some correct alt text gets questioned; AI alt text accuracy has the details. Nothing on this page is legal advice, and the labels and notes it describes do not show that GotAlt or your site meets Article 50 of the EU AI Act or any other rule. The AI disclosure laws hub lists the rules with their sources.
Questions about how GotAlt uses AI
Does GotAlt use AI?
Yes. The alt text checker, the deep audit, the alt text generator, the AI fix plan on scan results, the Site Image Check with paid monitoring, and the Bulk Alt Text Pack each call an AI model. The rule-based scan and the AI disclosure tools do not.
Which AI model does GotAlt use?
On October 22, 2026, every GotAlt feature that uses AI runs on Claude Haiku 5.5, a model from Anthropic. GotAlt does not build or train its own models. The model can change, so this page names the one in use on the date it was last reviewed.
Can GotAlt tell me whether an image was made by AI?
No. Anthropic's documentation says Claude cannot determine whether an image is AI-generated, and GotAlt does not offer AI detection. Our AI disclosure tools read what a file or page declares, such as a digital source type in metadata, and finding nothing proves nothing about how an image was made.
Are GotAlt's guides written by AI?
They are drafted with AI assistance and checked against the sources they link, by AI models and by scripts. Each guide, law page, standards page and platform page says so under its dates and links to this page. GotAlt publishes the guides and is responsible for them, and we do not describe the checking as human editorial review.
Why do the tool results say AI-generated?
So you can tell what a model wrote from what your page says. The label is plain text in front of the model's words, readable by a screen reader, with contrast above the WCAG minimum. It is our own wording, not a legal mark, and it does not mean a person has checked the output.
Who is responsible for what GotAlt publishes?
GotAlt, the company that runs this site, is responsible for every page and every result, whether a person or a model drafted it. To report an error or ask a question, email hello@gotalt.com. We reply within two working days (Monday to Friday).
See what the vision check finds on your site
Run a free scan to see which problems automated checks find, or look at a real report first.
Scan my site for free See a real report first