Re: [EXT] Re: AI and the future of Web accessibility Guidelines

I’ve been watching this conversation with a lot of interest, and most of what I think about it is in agreement with many folks here, but Gregg, you said one thing that I absolutely can’t let go unchallenged:

AI will absolutely be able to do a better job of doing text alternatives than 50%,   then  60%,  then 70%, then 80%, then 90 %, then 95% etc of authors.
It is not a question of if - just when.

The main thing that prevents AI or any computer-generated tooling from being truly successful at producing alternative text is context. Now, there are many situations where context is simple enough that I agree that as LLMs and computer vision models progress, we’ll see some levels of context enter description. Not today, but definitely in the next few years. This is where image descriptions in places like ecommerce will truly benefit.

Where I don’t think we’ll ever see these models achieve adequate context is images with high-context meaning. What do I mean by “high-context”? Images that are placed with specific author intent. I like to give the example of someone writing a blog post and including a photo of a cat. AI models are going to view the image of the cat and likely provide something in the range of “A photo of a cat” to “A photo of a black, white, and orange cat” (currently). Now if the models get smarter, which I believe they will, they may be able to figure out that my blog post where I have included the image is about my cat, and the image is to give an idea to the reader of what she looks like. If it’s really smart, it may pull out key information from the post to make the image description more informative to the reader: “Phoebe, the author’s cat, is black, white, and orange”.

Where I believe it will still fall down, is “why” did I include the image. What if my blog post is about cat behaviour, the image I have included is meant not only to show the world how cute my cat is but also to demonstrate a specific display of cat behaviour? Now, if I am a good writer, I probably explain this in detail in the text of my blog post, but what if I rely a little too heavily on the image to make my point?

Images are added to books all the time, particularly history books. A photo description of “two men in suits shaking hands” is woefully inadequate when the image is actually documentation of two world leaders agreeing to a monumental peace accord. AI may well progress to being able to identify the people in the image, or even pinpoint the exact moment the picture is documenting, but we also know it often hallucinates and those hallucinations, depending on the context, could be deeply problematic. It’s easily possible to imagine AI being run over a book on the history of WWI and confusing King George V and Tsar Nicholas II, two men who looked remarkably alike but had very different stories.

I’m “AI-neutral”, I think there are some strong potential benefits, but equally strong harms, and I get really nervous when people get too confident about where AI is going. Being cognizant of the challenges and limitations mean we can work with them, and develop standards and processes that counterbalance them.

-Wendy

From: Jennifer Strickland <jstrickland@mitre.org>
Date: Friday, April 12, 2024 at 4:29 PM
To: Michael Gower <michael.gower@ca.ibm.com>, Nat Tarnoff <nat.tarnoff@levelaccess.com>, Roberto Scano <mail@robertoscano.info>, Todd Libby <toddlibby@protonmail.com>
Cc: Patrick H. Lauke <redux@splintered.co.uk>, WCAG list (w3c-wai-gl@w3.org) <w3c-wai-gl@w3.org>
Subject: Re: [EXT] Re: AI and the future of Web accessibility Guidelines

[EXTERNAL] This message comes from an external organization.
From what I have seen to date of AI-generated alt text there hasn’t been a prompt for a user to review.

THAT is the reason this conversation was raised. We need to document guidelines for AI’s use for accessibility so that people have some guidance. We need to address the inevitability that it will be used.


From: Michael Gower <michael.gower@ca.ibm.com>
Date: Friday, April 12, 2024 at 3:30 PM
To: Nat Tarnoff <nat.tarnoff@levelaccess.com>, Jennifer Strickland <jstrickland@mitre.org>, Roberto Scano <mail@robertoscano.info>, Todd Libby <toddlibby@protonmail.com>
Cc: Patrick H. Lauke <redux@splintered.co.uk>, WCAG list (w3c-wai-gl@w3.org) <w3c-wai-gl@w3.org>
Subject: Re: [EXT] Re: AI and the future of Web accessibility Guidelines
The difference might be "a bouquet of roses" versus "Our vibrant red and yellow roses are the perfect gift for your someone special". While technically accessible in the first count, the second is much more meaningful and

  *   The difference might be "a bouquet of roses" versus "Our vibrant red and yellow roses are the perfect gift for your someone special". While technically accessible in the first count, the second is much more meaningful and effective.

But the second one has turned the ALT text into a marketing phrase which exceeds what is visually portrayed by the image.  Unless there is more to this image than a picture of some flowers, I would argue that that is not an appropriate alt text.

In some ways, this conversation adds to my conclusion that despite a lot of effort and documentation, images are frequently either not labelled, or they are labelled in such a way that it would be very difficult to have a group of people (even self-described experts) agree on the relative quality. I’m not even sure if we would arrive at consensus on pass/fail.

It’s difficult believing that a decent machine algorithm applied indiscriminately would produce worse results than the present state, frankly. Even with hallucinations.

But more importantly, I don’t think that is primarily how AI is going to be utilized in the near term. Most AI functions are in the form of suggestions that are acted on by a worker (in our case, an author). Further, the trend seems to be to try as much as possible to alert end users to the use of AI. I suspect curation of AI generated content – and transparency on where AI content has been adopted -- will become even more the case as lawsuits compound where unattended  generative AI has provided inaccurate information.

Mike


From: Nat Tarnoff <nat.tarnoff@levelaccess.com>
Date: Thursday, April 11, 2024 at 6:34 AM
To: Jennifer Strickland <jstrickland@mitre.org>, Roberto Scano <mail@robertoscano.info>, Todd Libby <toddlibby@protonmail.com>
Cc: Patrick H. Lauke <redux@splintered.co.uk>, WCAG list (w3c-wai-gl@w3.org) <w3c-wai-gl@w3.org>
Subject: [EXTERNAL] Re: [EXT] Re: AI and the future of Web accessibility Guidelines
TL;DR - Until humans as a whole have the capacity to think of others before themselves, AI will not be trust worthy to do the job of a human where emotions are involved. "Saying that something can’t be done in the future because it failed
TL;DR - Until humans as a whole have the capacity to think of others before themselves, AI will not be trust worthy to do the job of a human where emotions are involved.

"Saying that something can’t be done in the future because it failed in the past or can't be done now is like saying that "man will never be able to fly" in the 1800’s when they couldn’t….yet. "

This statement is accurate for things like planes, cars, computers, but I don't think it applies to AI. The reason I say this is that AI requires learning and training which is done by humans exposing the AI to human creations. Humans are fallible. We've seen this in cameras that can't focus on darker skin, in AI that was trained off Twitter and turned harmful in 24 hours, and in people taking over or creating companies that promote and thrive on hate speech. That point is made by your P.S. mentioning how bad the state of alt text is currently. Alt text isn't new. With decades of its existence and people still haven't learned.

I also think that while we have ML intelligent tools that can determine some of what is in a picture, giving alt text to an image that describes the image is often the wrong alt text. It won't deliver the meaning or emotion the image generates when in context. The difference might be "a bouquet of roses" versus "Our vibrant red and yellow roses are the perfect gift for your someone special". While technically accessible in the first count, the second is much more meaningful and effective. On the off chance AI can give more in depth information, I don't think it is going to get the feeling a human can provide until it is Turing complete.


Nat Tarnoff (they/them)


________________________________
From: Jennifer Strickland <jstrickland@mitre.org>
Sent: Thursday, April 11, 2024 8:01 AM
To: Roberto Scano <mail@robertoscano.info>; Todd Libby <toddlibby@protonmail.com>
Cc: Patrick H. Lauke <redux@splintered.co.uk>; WCAG list (w3c-wai-gl@w3.org) <w3c-wai-gl@w3.org>
Subject: Re: [EXT] Re: AI and the future of Web accessibility Guidelines

CAUTION: This email originated from outside of the organization. Do not click links or open attachments unless you recognize the sender and know the content is safe.


First, I agree with Todd and Patrick that AI will not be able to do the work of accessibility any time soon if every.



Second, the reality is that people will look for any shortcut and for many it will be easier to say let AI do it instead of learning to deliver accessibility.



We need to be clearer in our guidelines to make it easier for people to understand how to make things accessible.



We also need to address the use of AI, because people will jump to the seemingly easy route to meet accessibility without learning how to do it.





From: Roberto Scano <mail@robertoscano.info>
Date: Thursday, April 11, 2024 at 2:06 AM
To: Todd Libby <toddlibby@protonmail.com>
Cc: Patrick H. Lauke <redux@splintered.co.uk>, WCAG list (w3c-wai-gl@w3.org) <w3c-wai-gl@w3.org>
Subject: [EXT] Re: AI and the future of Web accessibility Guidelines

Agreed. Same issue of Google is happened with an e-commerce web site with mistake for gluten-free and with gluten. In this case the damage can be higher. . . Il gio 11 apr 2024, 03: 22 Todd Libby <toddlibby@ protonmail. com> ha scritto: I agree

Agreed. Same issue of Google is happened with an e-commerce web site with mistake for gluten-free and with gluten. In this case the damage can be higher...







Il gio 11 apr 2024, 03:22 Todd Libby <toddlibby@protonmail.com<mailto:toddlibby@protonmail.com>> ha scritto:

I agree fully with Patrick:

> Once you bring in the "AI will do it" line of thinking, we may as well
> just remove any author requirement, and WCAG becomes just a list of
> requirements for AI user agents to massage any old web content into
> something accessible.

Being a strong supporter against AI, because it will never work for accessibility, we have been down this road before. I mean, Google attempted their best shot and look where it got them. In hot water (so to speak).

https://www.theguardian.com/technology/2015/jul/01/google-sorry-racist-auto-tag-photo-app<https://urldefense.proofpoint.com/v2/url?u=https-3A__www.theguardian.com_technology_2015_jul_01_google-2Dsorry-2Dracist-2Dauto-2Dtag-2Dphoto-2Dapp&d=DwMGaQ&c=BSDicqBQBDjDI9RkVyTcHQ&r=GOTy3dcMNQuBB7l1GyFMjdnaO7XtjwzMNdypJV9QRHY&m=HqYmBUG8xxBYOXh3b6AvMPzGffBifgnmGioXtOq2TQfP07YQ4Gt9QJ_Yoc6huCUr&s=CihERo1WHTQOEwu0a5liSMP7M7ZaQYWwM1mYfxf0cVc&e=>

and if we are ever going to learn something as a people (which we have not), instances like this will always be why I vehemently oppose AI in tech and accessibility.

The article may be old, but the fact remains that AI is not the answer now (or ever in my opinion).





---
Best,

Todd Libby





On Thursday, April 4th, 2024 at 8:08 AM, Patrick H. Lauke <redux@splintered.co.uk<mailto:redux@splintered.co.uk>> wrote:

> On 04/04/2024 08:02, Gregg Vanderheiden RTF wrote:
>
> > I think much of our work is not forward-looking.
> >
> > We will soon have AI that can do a better job of text alternatives than
> > humans can for example.
> > And then it is unclear why we would require authors to do all this work.
> > This applies to a LOT of things.
>
>
> As a counterpoint, Gregg ... when does it end? You've stated similar
> when it comes to things like authors needing to provide correct explicit
> markup for headings, since (to paraphrase) "AI will be able to do it".
>
> Captions, audio descriptions ... "AI will be able to do it".
>
> Colour contrast issues? "AI can detect it and change it on the fly".
>
> Once you bring in the "AI will do it" line of thinking, we may as well
> just remove any author requirement, and WCAG becomes just a list of
> requirements for AI user agents to massage any old web content into
> something accessible.
>
> P
> --
> Patrick H. Lauke
>
> * https://www.splintered.co.uk/<https://urldefense.proofpoint.com/v2/url?u=https-3A__www.splintered.co.uk_&d=DwMGaQ&c=BSDicqBQBDjDI9RkVyTcHQ&r=GOTy3dcMNQuBB7l1GyFMjdnaO7XtjwzMNdypJV9QRHY&m=HqYmBUG8xxBYOXh3b6AvMPzGffBifgnmGioXtOq2TQfP07YQ4Gt9QJ_Yoc6huCUr&s=6weOMvBRKM9L4SK4UYAV6pyQzu_Ao6qOFnFQFqvVV_A&e=>
> * https://github.com/patrickhlauke<https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_patrickhlauke&d=DwMGaQ&c=BSDicqBQBDjDI9RkVyTcHQ&r=GOTy3dcMNQuBB7l1GyFMjdnaO7XtjwzMNdypJV9QRHY&m=HqYmBUG8xxBYOXh3b6AvMPzGffBifgnmGioXtOq2TQfP07YQ4Gt9QJ_Yoc6huCUr&s=d4aRCMLn88vTTpg8JKW_UU8bbnsrsGDzw6DCqyN-rvU&e=>
> * https://flickr.com/photos/redux/<https://urldefense.proofpoint.com/v2/url?u=https-3A__flickr.com_photos_redux_&d=DwMGaQ&c=BSDicqBQBDjDI9RkVyTcHQ&r=GOTy3dcMNQuBB7l1GyFMjdnaO7XtjwzMNdypJV9QRHY&m=HqYmBUG8xxBYOXh3b6AvMPzGffBifgnmGioXtOq2TQfP07YQ4Gt9QJ_Yoc6huCUr&s=lkVPzW7oA6Fl8ai1y47cmyCl_0VuCE53ylYS_aThSKo&e=>
> * https://mastodon.social/@patrick_h_lauke<https://urldefense.proofpoint.com/v2/url?u=https-3A__mastodon.social_-40patrick-5Fh-5Flauke&d=DwMGaQ&c=BSDicqBQBDjDI9RkVyTcHQ&r=GOTy3dcMNQuBB7l1GyFMjdnaO7XtjwzMNdypJV9QRHY&m=HqYmBUG8xxBYOXh3b6AvMPzGffBifgnmGioXtOq2TQfP07YQ4Gt9QJ_Yoc6huCUr&s=bJ2T0ocsby8VDyGC78g3rpOk7XhPlWzPYcQISI06SX8&e=>