Read the Label: An AI Ingredients List for Accessibility

An image of a man in a blue shirt and tie holding two symbols representing artificial intelligence (AI) and disability accessibility

When I pick up a box of crackers, I cannot see what is inside, so I rely on the label. In the US, federal law decides what that label says. Ingredients are listed by weight under 21 CFR 101.4, and the Nutrition Facts panel follows a fixed format under 21 CFR 101.9, so I can find the same information in the same place on almost any package in the store. The label content is rigidly regulated because in some cases, mistakes can lead to serious injury or death.

Digital products have no equivalent label for AI use. AI now sits upstream of much of what people with disabilities rely on. Captions, alt text, transcripts, PDF tags, front-end components, plain-language summaries, and the conformance reports procurement teams read are increasingly produced or tested by AI before anyone sees them. A qualified person may review that work, but the consumer has no way to tell how much AI went into it or how carefully anyone checked the result.

I think accessibility requires an AI ingredients list. Food labeling offers a tested model for designing one.

Where AI sits upstream of accessibility

AI can touch many aspects of accessibility long before the content reaches the person who relies on it:

  • Captions and transcripts: Automatic speech recognition produces the first draft of most captions, and many organizations publish that draft as is. WCAG 1.2.2 Captions (Prerecorded) applies the same way regardless of how the captions were produced.
  • Alt text: Image description models write alt text inside content management systems, social platforms, and document tools. WCAG 1.1.1 Non-text Content requires a text alternative that serves the equivalent purpose, and purpose depends on context the model often does not have and cannot infer.
  • Code and components: AI coding assistants produce markup, ARIA attributes, and focus management. Much of that code is inaccessible by default. Even when you request code be WCAG 2.2 conformant, AI doesn’t always get it right.
  • Document remediation: AI tools now tag PDFs, set reading order, and build table structure.
  • Audio description and sign language: Synthetic voices narrate audio description, and some vendors offer sign language avatars.
  • Testing and reporting: AI scanners and AI-assisted reviewers create test plans and produce audit findings, and some vendors use AI to draft Accessibility Conformance Reports based on the ITI VPAT template.
  • Support channels: Chatbots increasingly sit in front of accommodation requests and accessibility feedback forms.

Why disclosure matters more for people with disabilities

The person who depends most on an accessible output is often the person least able to check it. For example, a screen reader user cannot see that an AI tagging tool placed a sidebar in the middle of a PDF’s reading order. They hear the content out of sequence and may not know why it is hard to follow. A blind user reviewing AI-generated alt text, or a Deaf user watching automatic captions, faces the same problem: they have nothing to compare the output against.

Food allergen labeling exists for the same reason. A person with a peanut allergy cannot look at a cookie and see peanut protein. The law places the duty to disclose on the party who knows what went in, because the consequence falls on the party who cannot detect it.

Disclosure also changes what people do. If I know captions were machine-generated and not reviewed, I treat the names and numbers in them with suspicion, and I may look for a transcript. If I know a chart description was AI-generated, I may ask for a human description before I make a decision based on it. Without disclosure, the user absorbs the error and often blames themselves or their assistive technology.

What food labels teach us

People have debated food labeling for decades; over what consumers need to know and what manufacturers can reasonably provide. Several of the results apply to AI disclosure.

Ingredients in order of predominance

Ingredients appear in descending order by weight. An AI label should say how much of the content AI touched. “AI-generated alt text for 3 of 40 images” is different than “AI-generated alt-text on this site”

A mandatory “Contains” statement for known risks

Major food allergens must be declared. The accessibility equivalent is a short list of AI uses that always require disclosure because errors directly block access: captions, transcripts, alt text for informational images, audio description, sign language, document tagging, and any AI that sits between a person and an accommodation request.

“May contain” and cross-contamination warnings that say very little

Precautionary allergen statements in the United States are voluntary. Food Allergy Research and Education (FARE) points out that no law requires these statements or controls their wording, and that people managing allergies widely misunderstand them. FDA’s Edition 5 allergen labeling guidance says an advisory statement is no substitute for good manufacturing practices and allergen controls.

AI disclosure will drift the same way unless we design against it. A footer that says “AI may have been used in creating this content” tells a user nothing about which content was affected or whether anyone checked it. A disclaimer like that also invites organizations to treat the disclaimer as a replacement for quality control.

“Free from” claims must be consistent

The same FDA guidance says a product claiming to be free of an allergen should not carry a “may contain” statement for that allergen. If an organization claims its captions are human-verified, a blanket AI disclaimer covering those same captions contradicts the claim, and users have no way to know which statement to believe.

A standard format in a predictable place

The Nutrition Facts panel works partly because it looks the same on every package. The FCC applied the same idea to internet service with its broadband consumer labels. An AI accessibility label should use the same headings in the same order and live in a predictable location, such as the accessibility statement, so a screen reader user can find it the same way on any site. The label must also be accessible itself: real text with proper headings, never an image.

Lot numbers and recalls

Food packaging carries lot codes so you can trace a contaminated batch and pull it. AI models are updated often, and a given version may describe charts poorly or mis-transcribe certain accents. An organization that recorded which model version produced each piece of content, and when, can find and fix what was affected, and notify impacted end users. This is no different than a grocery store notifying me that I purchased a bag of salad from a lot that was discovered to have been contaminated by e. coli.

What current law requires, and where it stops

Several recent rules touch AI disclosure. However, none of them was written to address accessibility specifically.

EU AI Act, Article 50

The transparency obligations in Article 50 of the EU AI Act became applicable on August 2, 2026. The Digital Omnibus gave generative systems already on the market until December 2, 2026, to meet the machine-readable marking requirement in Article 50(2). Article 50 aims to ensure people know when they are interacting with AI or viewing synthetic content, protecting against deception. It does not ask whether the output is accurate or usable. Also, the marking requirement does not apply to AI that performs an assistive function for standard editing, which describes much accessibility tooling.

California AI Transparency Act

California’s SB 942, as amended by AB 853, became operative on August 2, 2026. It requires large generative AI providers to offer a free detection tool and to embed disclosures in image, video, and audio content. It targets the companies that build the models, not the city or retailer that publishes the output, and it doesn’t cover text, where alt text and transcripts live.

AI software bills of materials

In May 2026, CISA and its G7 partners released Software Bill of Materials for AI: Minimum Elements. An SBOM is commonly described as an ingredients list for software, and the AI version adds models, datasets, infrastructure, and performance indicators. It is the closest existing model to what I am proposing, but it is voluntary guidance written for security and supply chain risk and does not address whether an AI component affects the accessibility of the output.

Accessibility Conformance Reports

An ACR tells a buyer whether a product supports each WCAG success criterion, but it does not explain how the vendor evaluated the product or whether AI drafted the remarks. It also does not indicate whether the product itself uses AI at runtime to generate accessibility features such as captions or descriptions. Procurement teams often read these reports as though a person manually tested every row with assistive technology, yet with AI now common in both testing and report writing, that is often not how they are produced.

ADA Title II

The DOJ web rule at 28 CFR part 35, subpart H adopts WCAG 2.1 Level AA. An interim final rule published April 20, 2026 moved the compliance dates to April 26, 2027 for public entities with a population of 50,000 or more and to April 26, 2028 for smaller entities and special districts (see the DOJ web rule page). The rule says nothing about how content is produced. A city that posts inaccurate AI captions on a council meeting video owns that failure the same way it would own bad human captions. Disclosure would not create a safe harbor, but it would tell residents when to ask for an alternative and give the city a record of what needs fixing.

What an AI accessibility ingredients label should contain?

Here is a starting set of fields. I have tried to keep the list short enough that organizations will actually fill it out.

  • Content or feature: What the label covers, such as video captions, image descriptions, PDF tags, or a component library.
  • AI role: Whether AI generated the content, edited human work, translated it, or tested it.
  • Percentage of content: Roughly how much of that content type AI touched, for example, all captions or only captions on archived videos.
  • Human review: None, spot check, or full review, plus the reviewer’s qualifications. A review of alt text by someone who uses a screen reader every day is different from a content editor confirming that the field is filled in.
  • Tool and date: The product or service used and the date of generation, so affected content can be traced later.
  • Model: The specific AI model and version, plus the base model for any fine-tuned or custom model, or a statement that the vendor did not disclose it.
  • Testing method: Whether conformance claims rest on automated scanning, AI-assisted review, manual testing with assistive technology, or testing with disabled users.
  • Report Problems: A specific way an end user can report an issue with the material.
  • Known limits: Weaknesses the organization already knows about, for example, that the caption engine struggles with proper names or that chart descriptions omit data values.

Naming the model

A company name alone isn’t enough, and neither is a hosting platform like Hugging Face, which isn’t a model. The label needs the specific model and version. For a fine-tuned or custom model, it should name both the fine-tuned model and the base model it was built on. The base model indicates which known weaknesses may have been inherited, and the fine-tuned model’s name shows what changed.

There are practical limits to requiring the model be named:

  • Deployers often do not know: A city using a captioning service usually sees only the vendor’s product name, and vendors can switch models without notice.
  • Model names can mislead: The same model can perform very differently depending on prompts, settings, and human review, so a model name should not be read as a quality rating.
  • Chained models: One caption file may pass through a speech recognition model, and then a separate language model for cleanup, and each one can independently affect the result.
  • Vendor confidentiality: Some vendors treat their model choices as trade secrets.

Two layers address most of these limits. The public label names the product and gives one of three answers for the model: the model name and version, “multiple models,” or “not disclosed by vendor.” A “not disclosed” answer is useful information for a buyer or user. A separate machine-readable record, following the AI SBOM (Artificial Intelligence Software Bill of Materials) structure, contains full model identifiers, base models, versions, dates, and fine-tuning details, and procurement teams can request it. Update both layers whenever the underlying model changes, just as a food label is updated when the recipe changes.

The following table shows how this could look for a recorded city council meeting.

Content Captions and transcript for the October 14 city council meeting video AI role Captions generated by automatic speech recognition Percentage of content All captions for this video Human review Full review against the audio by a trained captioner within five business days of posting Tool and date Vendor captioning service, generated October 15, 2026 Model Vendor’s fine-tuned speech model, version 4.2, built on an open-source base speech model; both model names are listed in the city’s detailed AI record Testing method Caption accuracy and timing checked manually; media player controls tested with NVDA, JAWS, and VoiceOver Report Problems If issues are discovered in this transcript, please report to accessibility@yourcity.com Known limits Speaker names and street names corrected by hand; comments from the public microphone may be incomplete

Who reads the label

The label serves several kinds of consumers, and each reads it for a different reason:

  • People with disabilities: Deciding how far to trust the content and when to ask for an alternative.
  • Caregivers, educators, and support staff: Deciding whether material is ready to hand to a student, client, or family member.
  • Procurement teams: Comparing vendors on how accessibility is produced and verified, in addition to what the ACR claims.
  • Public entities and businesses: Learning what is inside the products and content they buy before they publish it under their own name.
  • Regulators, courts, and attorneys: Getting a record of how content was produced when something goes wrong.
  • Researchers and standards bodies: Collecting data on where AI performs well and where it fails, which today is mostly anecdotal.

Objections I expect

“Everything uses AI now, so every label will say yes.”

That is the “may contain” problem, and the fix is the same: require specificity. A label that names the content type, the AI role, and the review level carries useful information even if nearly every site has one.

“Disclosure is an admission of liability.”

I see it the other way. An organization that can show which content was AI-generated and who reviewed it is better positioned than one that cannot explain how its captions were made, especially after a user has been harmed by an error.

“Users will distrust good AI output.”

Some users will distrust it, and given how many caption and alt text errors people have already experienced, that is a reasonable starting point. Including the review level on the label gives well-reviewed AI output a way to build trust over time, which cannot happen when AI involvement is hidden.

“This is too much work.”

The team that ran the tool already knows most of these fields, so the main effort is recording them when the content is produced. Organizations that track remediation in an accessibility management platform can add these fields to the same record.

What to do now

No regulator is likely to require this label soon, so the first adopters will be organizations and buyers who decide it is worth doing:

  • Procurement: Add solicitation questions asking vendors where AI is used to produce accessibility features, how their ACR was evaluated, and what human review occurs.
  • Publishers and public entities: Add an AI section with consistent fields to the accessibility statement, starting with captions, alt text, and documents.
  • Vendors: Record tool, version, date, and review status for every AI-produced accessibility artifact, and make that record available to customers.
  • Governance and maturity work: Treat AI disclosure as part of accessibility governance. The W3C Accessibility Maturity Model already asks organizations to examine procurement and the ICT development life cycle, and upstream AI disclosure belongs in both.

I would like to hear from anyone who has tried something like this

Many thanks to Tzviya Siegman, who chaired the AI in Society breakout session at the World Wide Web Consortium (W3C) Technical Plenary Advisory Council (TPAC) session, which made me think critically about this topic.