No, ‘AI’ Will Not Fix Accessibility

In recent years, a series of new technologies have provided better experiences and outcomes for disabled users. Collectively branded “Artificial Intelligence”, the two biggest breakthroughs have been in computer vision and large language models (LLM).

The former, computer vision, allows a computer to describe an image based on extensive training on massive image data-sets, human tagging, and ongoing tweaking. This is well beyond the relatively ancient technology of optical character recognition (OCR), which allows your computer to transcribe the text it finds in a picture. Images that were empty voids to blind users could finally be useful.

Microsoft’s Seeing AI is just one example, allowing people to use an app on their phone to identify people, differentiate currency, navigate a refrigerator, and more. Be My Eyes, which has relied on sighted volunteers for the same tasks, is trialing its Virtual Volunteer. This would give users more privacy and immediate responses.

Similarly, LLMs can help readers distill a complex or verbose article to an abstract or different reading level. They can also help writers when anxiety, language barriers, dyslexia, and more might make working difficult. We can generate reasonable quality captions and transcripts on our own computers, often a significant improvement from the “craptions” of yore.

But…

Large language models are habitual liars. Meanwhile, automated image descriptions aren’t much better. To give them the benefit of the doubt, perhaps these tools simply lack context.

A chocolate Rice Krispie treat sits on a saucer in a darkened room, a single lit birthday candle sticking out of it. Also on the table and in shadow to the right is a drill, to the left is a pack of candles, and behind it a set of drill bits. As image tools get better at describing every detail of a picture, as language models do a better job of conveying an emoji-laden tweet in actual words, they are still not the authors of that content. They have no sense of why it was created. They cannot tell you that a series of vertical lines is meant to signify a wall in a meme. They don’t know that the gag of a photo of a Rice Krispie treat with a birthday candle is the power drill in the background.

Now let’s consider the code that holds all that content.

We have witnessed accessibility overlay vendors for years claim they are deploying AI technology and still failing users miserably. With their influx of funding and lofty promises, it should be done by now.

Forgetting the snake oil of overlay vendors, consider tools like GitHub Copilot, which claims to be “your AI pair programmer”. These work by leaning on the code of thousands and thousands of projects to build its code auto-complete features.

When you copy broken patterns you get broken patterns. And I assure you, GitHub, Google, Apple, Facebook, Amazon, stacks of libraries and frameworks, piles of projects, and so on, are rife with accessibility barriers.

Let’s assume those coding tools eventually get cleaned up to output idealized, conformant, accessible code. How will they account for platform bugs? No tool is stopping you from using CSS display properties on tables today. No accessibility overlay is rebuilding them dynamically to visually look the same while avoiding these long-documented bugs.

This isn’t an issue of not good enough yet. These tools simply won’t know the context or intent of the author (let alone the needs and expectations of the user). Certainly not while navigating the sea of bugs in nearly every learning source dataset (on the input side) or rendered screen to be fixed (on the output side).

Because…

Accessibility is about people. It is not a strictly technical problem to be solved with code. It is not the approximation of human-like ramblings produced by the complex algorithms generally branded as “AI”.

LLM chatbots and computer-generated images are already creating an uncanny valley of spam. We don’t need them to contribute to the analogous uncanny accessibility already extant on the web.

Creating a broken MVP because “AI” will be able to fix it later is lazy. Believing something is accessible because the “AI” said so is shoddy.

When we target output versus outcomes, we are failing our friends, our family, our community, and our future selves. We are excluding fellow humans while we try to absolve ourselves of responsibility by laying it at the feet of a black box its very makers warn will ruin the world.

Update: 8 September 2024

Flattered that final paragraph made it into a conference talk:

YouTube: Ein kritischer Blick auf KI-basierte Accessibility Tools | Ariane Rosing

13 Comments

Reply

Adrian, as always, the expertise and insights your blogs offer us are gold to us. That said, your insights are even more valuable than ever, because so many think that AI (and in particular, ChatGPT) can fix anything. In the accessibility world, we know that is not true, for all the reasons you outline — but it helps all of us in the accessibility world when we can point to a resource such as yours. Sometimes, when I say “no that won’t work for accessibility” it is so hard to say “why” it won’t work, at least in making complex content accessible. This blog outlines why — and supports my efforts in accessibility. Thank you so much for this, and all the considerable things you do for accessibility.

Reply

Do you want websites to never be fixed or do you want to have accessibility? You are making your statements about the capabilities of GPT 3.5. GPT 4 is out and GPT 5 is being trained. These errors and habitual lying will be fixed.

Here’s the point. No matter how much you need it to be, the web will never be fully accessible. Tons of content will remain inaccessible forever. It’s too much manual effort to convert every document.

But….a capable LLM can and will address 99.999% of these issues. And provide additional insight. Not only will it work like a perfect screen reader, it will do it in the voice of your choice and you will be able to ask it questions and interact with the site.

It’s the holy grail for online accessibility. You shouldn’t be so quick to dismiss it.

In response to Frow. Reply

Frow,

Forgive me ignoring your opening question. It felt rhetorical.

[…] You are making your statements about the capabilities of GPT 3.5. GPT 4 is out and GPT 5 is being trained. These errors and habitual lying will be fixed.

When? Are you suggesting they are fit for purpose today? Because what they are doing today is filling the web with bad information and misleading users.

Here’s the point. No matter how much you need it to be, the web will never be fully accessible. Tons of content will remain inaccessible forever. It’s too much manual effort to convert every document.

I don’t disagree that the web will never be fully accessible. There is simply too much legacy content and too many broken tools creating yet more soon-to-be legacy content. I am certainly not going to wash my hands of trying to fix those tools (and developers), however. That would be laziness and dereliction of responsibility as a human building for other humans.

But….a capable LLM can and will address 99.999% of these issues. […]

When? Where did you get that 99.999% number, or is that hyperbole? Which issues?

[…] And provide additional insight. […]

Indeed, computer vision can provide more detail for an image than an author provided, regardless of whether it is relevant to its use. As far as I know, that is not LLM on its own, however.

[…] Not only will it work like a perfect screen reader, it will do it in the voice of your choice and you will be able to ask it questions and interact with the site.

Even screen reader users disagree on what is the perfect screen reader. But I am curious how you think a LLM will itself replace a screen reader. One is simply a conversational interface the other is assistive technology. I am also curious how much experience you have working with screen readers to warrant making that assertion.

It’s the holy grail for online accessibility. You shouldn’t be so quick to dismiss it.

“Holy grail” is indeed a good analogue in that it is mythical (except to True Believers as evidenced by their great effort and expense).

I want you to note that I am not dismissing these tools you collectively call ‘AI’. I am dismissing their ability to understand what the author or developer intended to create. I am wary they will ever be able to address bugs in real-time. I am doubtful their training data are (or will be) adequate. And I am definitely pushing back on the idea you can stop making accessible stuff today simply because you believe that tooling will appear sometime in the future to make unspecified adaptations for users (and also screw them until then).

None of your response addresses those.

In response to Adrian Roselli. Reply

“When? Are you suggesting they are fit for purpose today? Because what they are doing today is filling the web with bad information and misleading users.”

I never said it’s ready today. I said “will be fixed” not “is fixed”. The capabilities in current state is enough to show the potential. AI is super effective now. We are seeing exponential improvement happen in real time. GPT4 is miles better than GPT3.5. etc… Your opinion that they are filling the web with bad information just shows your disdain for these tools and neglects their incredible power and potential. They are game changers and vastly different than the overlays you are used to railing against. The overlays are still code fixes.
There is no interaction with an overlay like you get from GPT. The AI will read and interpret the page. It can not only read out loud word for word but it will also describe the content.

” I am dismissing their ability to understand what the author or developer intended to create.”
I don’t see this as an issue. AI will describe the image. In as much detail as you want and you can ask it about the image. The original human author “under compliance” is only required to add a text label to the image. There’s nothing in there that says the author will put any more effort to label that “rice krispy treat”. They can put in “treat candle” and it satisfies compliance. Most fixes to content are done to satisfy compliance, there’s not that many heroes out there who are willing to go the extra mile and create meaningful alt text for every image on their site.

“I am wary they will ever be able to address bugs in real-time.”
Do screen readers or some other compliance address bugs in real time? They don’t have any capability to do so. AI is the only potential solution that *might* be able to address bugs in real time. You are wary that we will go from no way to address them, to something that might?

“I am doubtful their training data are (or will be) adequate.”
We are seeing leaps and bounds improvement on models with larger training data. There are proposed models for AI self training which could improve that process exponentially. Every new iteration we’ve seen so far has improved accuracy and been more impressive. Your doubts aren’t necessarily rational or based on facts.

” And I am definitely pushing back on the idea you can stop making accessible stuff today simply because you believe that tooling will appear sometime in the future to make unspecified adaptations for users (and also screw them until then).”
I’ll agree with this point, in that we aren’t there yet. But we will never be there if we need humans to modify and update their content. Humans are modifying and fixing the sites because of regulation. Regulation means you will get minimal effort possible to keep the auditors happy. Mediocre websites and slow adoption and top 5 content fixed while 95 isn’t fixed, is what you have now.

AI has so much potential and will provide game changing access to someone who needs it. The accessible community should be all in on AI technology.

In response to Frow. Reply

[…] AI is super effective now. […]

I think many would disagree (on both effectiveness and what ‘AI’ is). LLMs lie (as I link above) and image recognition lacks context (and can have features limited owing to past mis-deeds).

[…] Your opinion that they are filling the web with bad information just shows your disdain for these tools and neglects their incredible power and potential. […]

It is not my opinion. LLMs are writing more and more content, partly for SEO mills and partly as a function of markted CMS features. In the post I also link evidence of outputs from poor training data. I agree in the incredible power (to be believed by those who should know better) and potential (for defamation suits).

[…] The overlays are still code fixes. There is no interaction with an overlay like you get from GPT. The AI will read and interpret the page. It can not only read out loud word for word but it will also describe the content. […]

My browser can already read aloud all the content of the page. I don’t need ‘AI’ to do that. As for describing the content, sure, as I noted above LLMs can create simpler abstracts. But LLMs do not allow for the kinds of interaction that assistive technology does (unless you are telling me ChatGPT hooks into platform accessibility APIs and hardware interfaces).

I am dismissing their ability to understand what the author or developer intended to create.[…]

I don’t see this as an issue. […] Most fixes to [image alternative text] are done to satisfy compliance […]

I don’t disagree that fixes are done to satisfy compliance. And certainly for image alternative text, not all authors do a great job. But ultimately what you don’t see as an issue is the crux of why ‘AI’ is not positioned to solve these issues alone — documents written by humans for humans are best made accessible through humans.

[…] Do screen readers or some other compliance address bugs in real time? […]

Yes. Not all bugs, mind, but they have a long track record of built-in heuristics to address bugs, tweaking how they do so as a function of user (human) feedback.

[…] You are wary that we will go from no way to address [bugs], to something that might? […]

On the contrary, I am wary of their ability to understand a bug that impacts users and then, if recognized, their ability to remediate it correctly (evidence I linked above suggests not).

I am doubtful their training data are (or will be) adequate.

[…] Your doubts aren’t necessarily rational or based on facts.

See links within my post and evidence of poor training data cited within those links.

AI has so much potential and will provide game changing access to someone who needs it. The accessible community should be all in on AI technology.

As I noted, LLMs and computer vision indeed have a lot of potential. While I think the self-described ‘AI’ industry’s fear that this means the end of humanity is a bit much, I also think the argument that these technologies will be a cure-all is naive. After all, the disability community has lots of evidence and experience with that for every hot new technology. Right now ‘AI’ is being treated as just another disability dongle. Hopefully that reality will change and I would love to see it happen.

In response to Frow. Reply

> But….a capable LLM can and will address 99.999% of these issues.

It won’t address 99.999 percent of these issues. HTML is about content; CSS is about presentation; and JavaScript is about behaviour. If you’ve ever programmed, it should be obvious that you can’t fix issues in all three areas effectively.

XHTML was introducing a great step towards quality by throwing errors on invalid HTML. Instead, browser makers decided to fix all issues on the fly; hence, many issues aren’t even noticeable.

Instead of spending energy building tools to automatically fix issues made by software engineers, it would be better to focus on educating them and convincing browser makers to throw errors more often instead of fixing them automatically.

I wrote an article titled Accessibility overlays – demystifying secrets in which I emphasise limitations. I hope this helps.

Cezary Tomczyk; . Permalink
Reply

Great stuff as usual! Lot of potential, lot of hype.
Hey – consider changing the link and h1 colors to something more contrasty?

Gonzalo Silverio; . Permalink
Reply

On the link color – it seems to be random on each page load. I just ran into a very light orange.

Gonzalo Silverio; . Permalink
In response to Gonzalo Silverio. Reply

Yeah, I have an array of 8 WCAG-conformant pairs of highlight colors (for the light and dark themes). A set is randomly chosen on page load.

Reply

I’d love to see this updated to reflect UserWay’s LLM ‘accessible’ code engine. We often refer folx to your posts, and this concerning offering from UW needs a respected voice like yours to debunk their claims.

In response to JULIETTE ALEXANDRIA. Reply

Juliette, I updated my post #UserWay Will Get You Sued with a new section, UserWay Offers New ‘AI’ Code Fixer.

Reply

It’s a lazy overlay. As someone who is an early adopter of AI and deeply interested in accessibility, I agree with your points. When it comes to assistive technology for the blind and visually impaired, it’s not pretty—it’s messy and sloppy—but it works about 70% of the time. Since we don’t have anything better, we make do.
I’m optimistic, do you think it will eventually be good enough?

In response to Alex. Reply

I think it will continue to improve, but the edge cases will still exist. The challenge is that many people already think it’s good enough, say as much, and are probably doing little to confirm it always works, which means users in what are most the vulnerable situations (edge cases) will be forgotten.

Leave a Comment or Response

You may use these HTML tags and attributes: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>