AI Disclosure

Where we use AI

Three features call a language model, all of them OpenAI’s gpt-4o-mini. Each is listed below with exactly what it sends.

• Executive Summary. The plain-language paragraph at the top of a report, turning the scan’s statistics into a few sentences a non-specialist can act on. It receives aggregate, non-identifying statistics only: the score, inferred industry, risk level, issue counts by severity, the most-affected component categories, and the titles of the top issues. It does not receive your page’s HTML, screenshots, or contact details.
• Industry classification. When a website is scanned, we infer which industry it belongs to so the report can be compared against the right benchmark. This sends text read from the public pages we scanned, along with the hostname. It runs only on public web pages you asked us to scan, never on an uploaded document. If the model is unavailable or answers off-list, a deterministic keyword classifier is used instead.
• Image descriptions in PDF repair. Drafting alternative text for the pictures in a document you uploaded. This one sends the images themselves — the picture data extracted from your PDF — together with the document title, one image per request.

The image feature is opt-in and nothing is sent until you ask for it. Asking us to repair a document is not consent to send its contents anywhere, so the drafts sit behind a button that says what it will do before you press it. If you never press it, no part of your document reaches an AI provider at any point in the repair. Everything else the repair does — building the tag structure, embedding fonts, setting the language and title, fixing the tab order — runs entirely on our own servers with no model involved.

Where we do not use AI

This is the important part for trusting your results:

• The score is not AI. Issue detection runs on the deterministic axe-core engine plus our own custom rules and our keyboard checks, and the accessibility score is computed by a fixed, rule-based formula (see Methodology). The same site always produces the same score — no model randomness is involved.
• Issue detection is not AI. Every flagged issue comes from a defined rule, not from a model’s opinion.
• Remediation guidance is curated, not generated per request. The developer fix notes shown for each issue are written and maintained by us, mapped to the specific rule that fired — they are not produced by a language model on the fly.

What data is sent to the AI provider

It differs per feature, so it is worth being exact. The Executive Summary receives aggregate statistics only. Industry classification receives text from the public pages scanned. Image descriptions receive picture data out of your uploaded document.

That last one is the one to read twice. An image inside a document can contain personal information — a signature, a photograph, a scanned form — and if you ask us to describe your pictures, that image data is sent to OpenAI. We do not inspect what is in them first and we cannot judge on your behalf whether a given picture is sensitive. If a document contains images you would not send to a third party, do not use the description feature on it; the rest of the repair still works, and the pictures are simply listed for you to describe yourself.

Across all three features we do not send your contact details, your payment details, or any account information. We never send an uploaded document as a whole file to an AI provider — only the extracted images, and only on request.

If no AI provider is configured, the summary falls back to a deterministic, template-based sentence generated entirely on our own servers — the Service still works without any AI call.

Data handling and training

Data sent to our AI provider is processed under that provider’s API terms. Per OpenAI’s current API policy, content submitted via the API is not used to train their models. We do not use your data to train any models of our own.

Limitations and human oversight

AI-generated text can occasionally be imprecise or omit nuance. The Executive Summary is an explanatory aid layered on top of the deterministic findings — it never changes the score or the list of issues. Always rely on the underlying findings, and verify anything material before acting on it.

Image descriptions are drafts, and are presented as drafts. They are never written into your document automatically, because a description that is confidently wrong is worse than no description at all — an automated checker then reports the picture as handled and nobody looks at it again. Read each one against the picture it describes before you use it.

Questions

Questions about our use of AI? Email support@accesslumens.com.