AI and Machine Learning - Replace "overlays" as a section

[aardrian]

Section 5 is a top-level section devoted to a product category called "accessibility overlays."

I don't think it belongs in this document.

Suggested Action

I propose stripping the overlays section completely and replacing it with a section that addresses browser extensions, native apps, and perhaps other avenues where users can choose to use the features this document outlines:

• These can learn from the user's preferences.
• They can offer better privacy controls (see Ethical Web Principles on privacy).
• They can provide more control for customizations (see Ethical Web Principles on control).
• They let users decide when to throw compute cycles and battery power at a challenge (see Ethical Web Principles on sustainability).
• Unlike overlays, they can arguably support an entire device or set of devices (see Ethical Web Principles on multi-platform).

Overall, replacing overlays with extensions / apps better aligns with the ethics cited in #13 and sustainability cited in #15.

It has the added benefit of keeping this note from being associated with a product offering that users and the industry widely agree is problematic.

Background

While overlay vendors have claimed to use "AI," they have mostly only used the term "AI" in the marketing sense. Most of their remediation efforts are simple scripted features that a reasonably skilled developer could implement in JavaScript or CSS as a bookmarklet.

Looking over the use cases outlined in this document:

• Some overlays have used computer vision and LLMs to generate alternative text, but that is not unique to overlays.

• Most, if not all, of them exclude captioning from their services. Automated captioning existed before generative AI, so they could conceivably offer it today but have chosen not to.

• I am not aware of any performing automated language detection nor converting content to "plain language," let alone publicizing those as planned features.

• Instances I have reviewed of overlays adjusting color contrast appear to be brute-force methods (backplates for text, CSS overrides, etc) and don't seem to rely on computer vision nor have they generally appeared to trend that way.

• Similarly, cases I have seen of addressing headings appears to be algorithmic versus any higher-level machine learning.

• Same for visual spacing.

• I am not sure I have seen link purpose tackled, though that could be as simple as pre-fetching the destination and appending it to the link text.

• The few cases I have seen of generative sign language are (I have been told by fluent ASL speakers) not very compelling. This isn't to say it cannot improve dramatically, but I am unaware of overlay vendors expressing much interest.

All of those features could make it into overlays as "AI" in the future, of course. They could just as easily make it into browser extensions and custom native apps. However, I see neither browser extensions nor native apps as a discrete item in this document.

I believe browser extensions and native apps would be far better vehicles to replace Section 5 in the document. Partly because of the reputation of overlay products and vendors generally:

• One of the biggest vendors settled with the FTC "for Deceptive Claims that its AI Product Could Make Websites Compliant with Accessibility Guidelines"

• That same vendor is also the subject of a class action for its claims (PDF)

• One of its peers is also the subject of a class action

• They tend to be litigious when the quality of their output is challenged (1, 2, 3), which is a significant problem when users try to improve the product.

• The foray of one vendor into LLM code generation was so problematic that it became a running joke among practitioners.

• Disabled users built tools (1, 2) to block overlays on their sites.

While the overlays section hints at this by saying overlays are "viewed as rather controversial by" disabled users, I think there is far more evidence saying they are actual barriers. The uptick in lawsuits against sites with overlays seems to support that.

I hope that's enough background detail, but I am happy to offer more.

[rscano]

Fully agree. Not only lawsuit but also European Commission disagree about overlay usage.

The legal accessibility requirements in the EU are underpinned by technical criteria specified in the harmonised European standard EN 301 549 v3.2.1. Accessibility overlays are tools that can be added to websites with the aim of improving accessibility. Overlays, or any other tools which do not ensure the website itself meets the detailed criteria of the standard, are not an appropriate solution. It is best to fix accessibility issues at their source.

Another good practice is to provide feedback on accessibility issues to those developing or maintaining digital tools, to ensure that those responsible learn about digital accessibility and avoid future problems. To ensure that websites are genuinely accessible, it is important to always involve persons with disabilities in testing.

The European Disability Forum, users with disabilities and many accessibility experts have expressed concern about the use of overlays. For more information visit:

Joint statement by EDF & IAAP
Article by The A11Y project
Overlay Fact Sheet signed by nearly 900 accessibility experts.

[UsableNet-WCAG]

Section 5 is a top-level section devoted to a product category called "accessibility overlays."

I don't think it belongs in this document.

Suggested Action

I propose stripping the overlays section completely and replacing it with a section that addresses browser extensions, native apps, and perhaps other avenues where users can choose to use the features this document outlines:

• These can learn from the user's preferences.
• They can offer better privacy controls (see Ethical Web Principles on privacy).
• They can provide more control for customizations (see Ethical Web Principles on control).
• They let users decide when to throw compute cycles and battery power at a challenge (see Ethical Web Principles on sustainability).
• Unlike overlays, they can arguably support an entire device or set of devices (see Ethical Web Principles on multi-platform).

Overall, replacing overlays with extensions / apps better aligns with the ethics cited in #13 and sustainability cited in #15.

It has the added benefit of keeping this note from being associated with a product offering that users and the industry widely agree is problematic.

Background

While overlay vendors have claimed to use "AI," they have mostly only used the term "AI" in the marketing sense. Most of their remediation efforts are simple scripted features that a reasonably skilled developer could implement in JavaScript or CSS as a bookmarklet.

Looking over the use cases outlined in this document:

• Some overlays have used computer vision and LLMs to generate alternative text, but that is not unique to overlays.
• Most, if not all, of them exclude captioning from their services. Automated captioning existed before generative AI, so they could conceivably offer it today but have chosen not to.
• I am not aware of any performing automated language detection nor converting content to "plain language," let alone publicizing those as planned features.
• Instances I have reviewed of overlays adjusting color contrast appear to be brute-force methods (backplates for text, CSS overrides, etc) and don't seem to rely on computer vision nor have they generally appeared to trend that way.
• Similarly, cases I have seen of addressing headings appears to be algorithmic versus any higher-level machine learning.
• Same for visual spacing.
• I am not sure I have seen link purpose tackled, though that could be as simple as pre-fetching the destination and appending it to the link text.
• The few cases I have seen of generative sign language are (I have been told by fluent ASL speakers) not very compelling. This isn't to say it cannot improve dramatically, but I am unaware of overlay vendors expressing much interest.

All of those features could make it into overlays as "AI" in the future, of course. They could just as easily make it into browser extensions and custom native apps. However, I see neither browser extensions nor native apps as a discrete item in this document.

I believe browser extensions and native apps would be far better vehicles to replace Section 5 in the document. Partly because of the reputation of overlay products and vendors generally:

• One of the biggest vendors settled with the FTC "for Deceptive Claims that its AI Product Could Make Websites Compliant with Accessibility Guidelines"
• That same vendor is also the subject of a class action for its claims (PDF)
• One of its peers is also the subject of a class action
• They tend to be litigious when the quality of their output is challenged (1, 2, 3), which is a significant problem when users try to improve the product.
• The foray of one vendor into LLM code generation was so problematic that it became a running joke among practitioners.
• Disabled users built tools (1, 2) to block overlays on their sites.

While the overlays section hints at this by saying overlays are "viewed as rather controversial by" disabled users, I think there is far more evidence saying they are actual barriers. The uptick in lawsuits against sites with overlays seems to support that.

I hope that's enough background detail, but I am happy to offer more.

I would agree that overlays are not going anything different with AI that a browser extension could do or any other system that can bring an update to an html element. AI in all cases is being used just to recommend content and the system itself is applying changes. The same suggestion could be in a testing or dev co-pilot. No AI is actually updating the DOM or A11y tree. I would agree removing the section but not replacing with extensions or apps , as these are not AI doing anything. All these mechanisms (extensions, JS include file, etc) that can update the end user client side are not AI in action.

[nattarnoff]

No AI is actually updating the DOM or A11y tree.
I think this right here is important. Some vendors scrape the site, use AI to provide options, but what is happening on the site is JS adding or updating attributes. We've been using JS to do that for years, most of ARIA works on this principle.

[aardrian]

I would agree removing the section but not replacing with extensions or apps , as these are not AI doing anything.

You made fair points @UsableNet-WCAG and @nattarnoff.

Mostly I wanted to acknowledge the aspirational opportunities, but the current crop of tools is lacking as you note. Even though future local installs could offer users better control and privacy.

I would be ok with striking overlays completely without replacement.

[aardrian]

Looking at the most recent draft, I still think the section (now 6.1) should be stricken.

For example:

the applicability of an overlay that contributes accessibility improvements is similar to the use of AI chatbots and other prompting mechanisms, […]

This is a false equivalency. An LLM chatbot user generally has a specific outcome in mind -- playing with writing styles, generating a piece of code to solve an issue, making pictures, simplifying language, etc. One gives the user agency, the other does not.

[…] thereby suggesting this may prove to be another area where generative AI could introduce improvements.

That conclusion is based on the prior false equivalency. If the end user has no idea what to fix (perhaps they cannot perceive a control or recognize how a widget is broken), then they have to rely on a third-party vendor having a pre-defined "prompt" (to carry the metaphor forward) and hope it works. For each visit / page load.

Generative AI, where it works, should fix issues as infrequently (and consistently) as possible and without putting the burden on the user (and their device). At code creation or on the server would address that. For additional customization that the user wants, that would be best provided by generative AI tools on their OS or in their browser so it can apply to all sites (and potentially retain and learn from those changes).

In other words, better outcomes can be achieved with generative AI by moving it out of overlays and to the sites themselves and/or to the OS/browser. Overlays are almost literally the worst place to deploy the features being described in 6.1 User Interface Generation section.

In addition, I think adding Daniela Kubesch's research, The Impact of Web Accessibility Overlays on the Usability and User Experience for People with Permanent Visual Impairments would be appropriate in 8. Reference List.