• Skip to main content
  • Skip to footer
Equalize Digital Home

Equalize Digital

Website Accessibility Consulting, Training, and Development

  • My Account
  • Support
  • Checkout
  • Software
    • Accessibility Checker
      • Documentation: Accessibility Checker
      • Live Demo
      • Start Free
      • Buy Accessibility Checker
    • ArchiveWP
      • Documentation: ArchiveWP
      • ArchiveWP Demo
      • Buy ArchiveWP
    • BoardScribe
      • Documentation: BoardScribe
      • BoardScribe Demo
      • Free on WordPress.org
  • Services
    • Accessibility Audits
    • User Testing
    • Accessibility Remediation
    • VPAT & ACR Preparation
    • Accessibility Monitoring
    • Web Accessibility Training
    • Accessibility for Agencies
  • Courses
    • NVDA Screen Reader Testing
    • VoiceOver Screen Reader Testing
    • How to Position and Sell Accessibility Offers
  • Resources
    • Accessibility Meetup
    • Articles & Webinar Recordings
    • Accessibility Craft Podcast
    • Upcoming Events
    • Office Hours
    • Custom Accessibility Training
  • About
    • About Us
    • Our Team
    • Industry Expertise
    • Accessibility Statement
    • Contact Sales
    • Become An Affiliate
  • Contact Sales
  • My Account
  • Support
  • Checkout
Home / Learning Center / Ethical AI & Benevolent Friction: Rethinking Alternative Text in WordPress: Deneb Pulsipher

Ethical AI & Benevolent Friction: Rethinking Alternative Text in WordPress: Deneb Pulsipher

Article PublishedOctober 7, 2026Last UpdatedOctober 7, 2026 Written byEqualize Digital

Rethinking Alternative Text in WordPress Deneb Pulsipher

Generative AI tools are flooded with promises of automated accessibility shortcuts, but fully automated, “hands-off” alt text often fails blind and low-vision users by missing context or introducing inaccuracies.

In this presentation, Deneb Pulsipher uses a real-world case study—the open-source plugin Polly Alt—to explore how we can ethically leverage generative AI as an accessibility aid by intentionally keeping a human in the loop.

Deneb looked at how we can design user interfaces that use benevolent friction within Gutenberg, Elementor, and the WordPress Media Library to make it completely frictionless for site administrators to write or edit alt text, while making it gently difficult to publish empty image descriptions.

Additionally, Deneb discussed how AI shouldn’t just do the work for lazy administrators, but should act as an accessibility training tool—using multi-subject angles and screen-reader context boxes to teach content creators how to draft high-quality visual descriptions on their own over time.

Thanks to Our Sponsor

Kinsta provides managed hosting services for WordPress. It is powering 120,000 businesses worldwide and based on the user reviews it is the highest-rated managed WordPress host on G2. It has everything you need, including an unbeatable combination of speed, security, and expert support.

Powered by Google Cloud and the fastest C3D and C2 servers combined with CDN and Edge Caching. Your sites are secured with Cloudflare Enterprise, protecting you from DDoS attacks. All plans include free migrations, and the first month of the starter plans is completely free, so you can try the service risk-free.

Watch the Recording

If you missed the meetup or would like a recap, watch the video below or read the transcript. If you have questions about what was covered in this meetup please post a message in our Facebook group for WordPress Accessibility.

Read the Transcript

[00:00:00] Meetup Introduction

Amber: Welcome to WordPress Accessibility Meetup: Ethical AI and Benevolent Friction: Rethinking Alternative Text in WordPress with Deneb Pulsipher of SeaMonster Studios.

A few announcements that you should be aware of if you haven’t been before.

The first one is that we have a Facebook group you can use to connect between meetups. You can find it if you go to facebook.com/groups/wordpress.accessibility. It’s a great place to just chat with other folks who are interested in creating more accessible WordPress websites, share things you’re working on, give feedback, and help others.

Everyone always asks, “Is this being recorded?” Yes, it is being recorded. You can find upcoming events and past recordings in one place on our website if you go to equalizedigital.com/meetup. We have somewhat recently updated our process around the recordings, and you can expect to see the recording posted on our website sometime early next week, so they’re coming a little bit faster now.

The other way to get notified when the recording is available is if you join our email list. We send out news related to web accessibility every Thursday, and then we also send event announcements. And you can get that information as well as a notification when the recording is available if you go to equalizedigital.com/focus-state.

And finally, we do put the audio for this on our podcast, and you can find audio-only versions in any podcast player or if you go to accessibilitycraft.com.

If you have any suggestions for the meetup or you need any additional accommodations to make the meetup work for you, please contact us at meetup@equalizedigital.com.

I have not introduced myself yet. If you’re new here, my name is Amber Hinds. I am the CEO of Equalize Digital, and we are the organizer of WordPress Accessibility Meetup.

Equalize Digital is a mission-driven organization that is focused on WordPress accessibility. We make WordPress plugins. We have an accessibility checker plugin that helps you find and fix problems on your website.

We also have some micro plugins, ArchiveWP and BoardScribe, that help you to support remediation efforts. We offer online courses for NVDA and VoiceOver screen reader testing, as well as selling accessibility, and we do a number of accessibility services like audits, remediation, and consulting. You can learn more about all of that if you visit our website, equalizedigital.com.

I do want to thank our sponsor today. Kinsta is generously covering the cost of live captioning and transcription. We do have a professionally trained captioner here with us today to make sure everyone gets accurate captions, and that is thanks to Kinsta.

Kinsta provides managed hosting services for WordPress. It is powering 120,000 businesses worldwide, and based on user reviews, it is the highest rated managed WordPress host on G2. It has everything you need, including an unbeatable combination of speed, security, and expert support.

Powered by Google Cloud and the fastest C3D and C2 servers, combined with CDN and edge caching, your sites are secured with Cloudflare Enterprise, protecting you from DDoS attacks. All plans include free migrations, and the first month of the starter plan is completely free, so you can try the service risk-free.

You can learn more about Kinsta if you go to kinsta.com, which is K-I-N-S-T-A.com. I always ask if you are willing to find our sponsors on whatever social media platform you prefer to use and comment on something they’ve written, tweet at them, send them a message, and say thank you for sponsoring captions for WordPress Accessibility Meetup. That helps to tell them that this matters, and it helps to encourage them to want to continue sponsoring.

I want to talk about a couple of events, and the biggest one that is coming up next week is WordPress Accessibility Day. This is a totally free 24-hour virtual conference. It’s gonna start at 10:00 AM US Central Time on October 7th, and run the full 24 hours until 10:00 AM US Central Time on October 8th.

If you cannot attend any sessions or all the sessions, don’t worry. If you register, you will be able to continue logging into Zoom events and catch all of the replays through the end of the month, so you can watch them on whatever schedule works for you. So I would encourage everyone to go register because there’s 24 phenomenal talks.

They’re all pre-approved for IAAP continuing education credits, if that’s something that you need. And we have some really great speakers, including our keynote speaker, who is the head of accessibility at GitHub. So please go register. You can do that if you go to wpaccessibility.day/2026, and I think we have a link for that in the chat

Other upcoming meetups to be aware of, on Tuesday, October 20th at 10:00 AM, we will have Courtney Robertson here who will be presenting on using Accessibility Checker Pro and Claude AI to remediate a large WordPress website.

And then on Thursday, November 10th, Cam Coulter will be presenting I Didn’t Know WCAG Meant That!

I’m gonna add a spotlight here for today’s speaker. I am so excited to welcome back Deneb Pulsipher. Deneb, who is known as Captain Accessible at SeaMonster Studios, works to make sure that the sites SeaMonster designs and builds work for everyone. When he’s not doing that, he’s working on tools to make the processes easier for accessibility testers and content creators. The more tools are available for accessibility testers, and especially for content creators, the easier they are to use. Of course, that means the more likely it is that people will know to do the right thing.

Deneb has spoken for us a few times in the past, and always gives phenomenal presentations, so you all are in for a treat.

Welcome back, Deneb.

Deneb: Thank you, Amber.

Amber: While you share your screen, I’ll just put a quick note for attendees that we do have the Q&A module available in Zoom. I’ll come back at the end to pass questions along to Deneb. If you can put them in the Q&A instead of in chat, that is super helpful so that they don’t get lost.

Thanks, everybody, and I’ll let you take it away, Deneb.

Deneb: Okay. You can see my presentation?

Amber: Yes.

[00:07:18] Presentation Introduction

Deneb: Okay. Perfect. Ethical AI and Benevolent Friction: Rethinking Alternative Text in WordPress.

We’re gonna talk a fair amount about alt text and what it should be. But using that to get at this broader question of how to use AI ethically and if you want to access the presentation, it’s an HTML version of this presentation. You can go to bit.ly/Polly, P-O-L-Y, hyphen P-R-E-S, Polly-pres, or use the the QR code there.

AI is everywhere, and it’s pretty helpful in many different ways, but only if you approach it with the right kind of intent and the right understanding.

I hope in this presentation we can get to that what it means to use AI ethically and well, and one way to do it. And also best practice for alt text, as an underlying theme.

[00:08:15] Philosophy

Deneb: Here’s my my ultimate philosophy, which I think might be a little bit controversial. And if it is do hit me up. I’d love to get into long, lengthy, email conversations or debates about this.

I think it is better to overdo alt text than to underdo it. I think that, that ensures an equivalent experience for all users. The other side of the question is, of course many people who say no.

If you overdo the alt text, if you write all 125 characters or whatever and you describe these images to the hilt you’re gonna… it’s gonna be problematic for your screen reader users. It’s just gonna bog ’em down, right? And I understand that that view as well and that’s why it is a question.

But I think overdoing it is better than just assuming many of the images are decorative, and maybe just putting the subject in the alt, so really brief. Because, if you only do that, if you only put the minimal alt text, I think you end up excluding your users. You’re withholding some vital information and some vital purposes of why you included those images in the first place.

And I think we’ll, I think we’ll get into that a little bit, but I’d love to discuss it more with people personally as well. So you can always email me or find me on LinkedIn. Okay.

[00:09:32] Beyond Minimal Alt Text: Pros and Cons

Deneb: Let’s talk about some of these benefits of going beyond just the minimum with alt text.

So first off, it it provides a richer context, right? You give a fuller understanding to screen reader users of the image and also of kind of the reasons that you use that image are more fully utilized or they come to fruition that way, right? It supports inclusion by conveying more nuanced information and reduces ambiguity, right? Also as an added benefit the last one here on the benefits side is it improves SEO by offering descriptive content that then search engines can index, right? So that’s kind of a side benefit, not the most important that we’re looking at. But it, it can be helpful.

Though, going beyond the minimal alt does come with some challenges, right? Lengthy alt text if you exceed 125 characters, and even if you bump up against that limit, you take it right to 125 characters it can slow down screen reader navigation, right? If a screen reader user is listening to everything on the page, which they almost never will, but if they do, it’s gonna be a longer process to listen to all of those images that you’ve described a little more fully. And in that case, too much detail may overwhelm or frustrate some screen reader users. They may just be tapping their toe and saying, “Hey, get to the point. I don’t need to know all of this about this image. I just need to know kind of what’s the gist of it.” And I think that’s a valid concern.

Also it can cause cognitive overload, right? But I think that’s only if it’s not structured or prioritized effectively, and we’re gonna talk about how to do that and what that means in just a minute. But then with this last challenge that I have written here on the screen long descriptions can be cumbersome to write and to maintain. And when something is cumbersome to write, the content creators, it just takes longer to do, and it’s less likely that they will do it.

We’ve got some benefits. We’ve got some challenges. Let’s see how we can resolve those.

[00:11:41] Overcoming the Cons: Screen Reader Capabilities

Deneb: Moving on to the next slide, we’ve got overcoming the cons, right?

One big way to overcome all of these cons, or not all of them, but some of these problems, these challenges with extensive alt text is by just knowing that the screen reader capabilities can overcome that.

The ability is that screen reader users, they can skip around, right?

Once they have the gist of that image, they can fly on to the next element that they’re interested in. They don’t have to listen to the whole thing. So that’s the screen reader ability of the actual program allows them to skim the surface a little bit if they want to.

But there is a gotcha, and the gotcha with this ability is that this only really works well if the alt text is written and if it’s prioritized.

And we’re gonna talk about what it means to prioritize this alt text and how to do that and why it can be a tricky process or it can be. Might not be so bad. Okay.

[00:12:43] Enhancing User Control with Skippable Verbose Alt Text

Deneb: To illustrate this skippability of verbose alt text. If you put alt text that is a little more on the verbose side, here’s how a screen reader users may view that or how the screen reader itself allows them to look at that.

And I’ve created an image on this side. It’s a view of a webpage or a part of a webpage that has three images. And then it has, above these images, it has the alt text. But this alt text is cut off after, I don’t know, five or six or seven words. Just illustrating if that’s all you have, and you jump after those five or six words, you still can know what this alt text is about, right?

The first alt text says… It’s a picture of a dog looking up at its master, and the first alt text says, “An eager brown dog looks up happily at its owner,” and then it went on. There was more to this alt text, but it’s cut off there, right? But that’s enough for the screen reader user to know what that image is about. And it’s enough to convey kind of the feeling that image was probably added in order to do, right? You could easily say “No, that’s that image there is just gratuitous. It’s just decorative. It doesn’t really build up to the meaning of this site,” but I think it does.

I think if somebody puts an image there, even if someone might say “That’s just a decorative image,” if it adds to the feeling of the site, I think it’s needed because we wanna create those same feelings for screen reader users as we do for non-screen reader users visual web surfers, right?

Excuse me. Oh, my. Okay. Screen reader users or screen readers allow users to skip long alt text, giving them control about how much detail they want. If you provide essential information up front, that means that it’s gonna be available for everyone, even if they skip after just a few words, right?

Doing this respects the user preferences as well as cognitive load management, right? So what we need to do in order to accomplish this is you gotta make sure that you clearly structure your alt text with the most important information at the beginning, realizing that there’s a pretty good chance that advanced screen reader users will only listen to the first part of your image.

They may come back later on and listen to the whole thing, but at least at the first glance, they’ll probably just listen to the first little bit. So you gotta make sure that the essential information is put up front. And that for content creators who are not used to making alt texts, and even for those of us who are, right?

That can be a tricky process. ‘Cause I gotta now think, “Okay I gotta be descriptive but I also need to prioritize. I gotta make sure that I’m putting the most important first and everything in the, after the last, six or seven words, I gotta think that it might just be skipped,” right? So this is a lot going on in the content creator’s mind which makes it difficult. And that’s gonna be a barrier for some content creators, ’cause now they’re thinking, “Oh, man, if I have to think about all those things when I’m adding this alt text it’s better to just call this image decorative and not deal with it, and just pretend that leaving it off the website is an okay thing to do,” when really I think it’s not.

[00:16:10] Front-Loading Alt Text: Pros and Cons

Deneb: All right, moving on. Looking a little bit more about these the pros and cons of this process of front-loading the alt text with the most important stuff.

The advantages, it allows users to skip if they want to, so they get the essential information, then they can skip.

It improves their efficiency and reduces cognitive load and navigation time, right?

It helps them to potentially get through the site more quickly. And there are some screen reader users who can evaluate the entire page almost faster than, or maybe definitely faster than I can, and they understand it better than I can as a sighted user, right?

It’s pretty amazing some of the capabilities that they have.

Another thing that this advantage may be, or another advantage this may have of front-loading this alt text is it prevents confusion about how the image relates to the other content on the page, right? Once again, that’s one more thing that the writer needs to prioritize or the content creator needs to prioritize is how does this connect with the surrounding page. And so now they’re thinking about that and they’re think… so it is a lot to think about, but it does support a better user experience for visually impaired and sighted users using screen readers, right?

It also encourages concise and clear descriptions that focus on the essential information first, right?

That’s the point of front-loading is to do that. And it does bear these really great advantages.

Unfortunately, it does also come with some challenges. Some of those challenges are that even when I front-load my most important information screen reader visitors might skip before the subject is fully announced or fully developed.

So you may have figured out and crammed all your most essential information into that first half of the alt text, but what if they only listen to four words, right? And then they’re off. Maybe they do miss the subject before it’s been fully articulated.

Also, another challenge is it requires careful wording to balance clarity without sacrificing the necessary detail.

So just a lot of things go into trying to do this process. And as an alt text composer, the more things you have to think about when you’re writing that alt text, the harder it is for you.

And my final point here is, the harder that process is the less likely it is that it will actually get done. And if it does get done, then it’ll get done well, right?

And so we’ve got all of those advantages and challenges.

[00:18:47] Overcoming the Cons: Ethical AI Can Help

Deneb: But let’s see if we can address some of these challenges. And I think we can with ethical AI. I think ethical AI can help with this process. So let’s talk about what this is.

My definition of ethical AI or ethical artificial intelligence is humans using AI not to cede their responsibilities over to the AI, but instead to enhance their abilities to complete them in the best possible way. I’m not giving over my responsibility to AI to do my job of creating that good alt text or the effective alt text. I’m not giving that responsibility over. I’m not abdicating it. Instead, I’m just using that as a tool to do what I should do, what I need to do, right? And some of the points I have on this slide are that one of AI’s greatest temptations for its users is for them to think that AI can take care of things without them, right?

We see the promise of this in agentic AI, where you say, “Okay, hey AI, go do all these things, all these steps that I want accomplished. Go do them.” And it does it without asking you, “Is this okay? Is this okay?” It just does it. Unfortunately, in the end, even when you can tell AI to do that, and it will do something, you’re not really able to give up your responsibility there, right? It’s still your responsibility as the human to make sure that it’s doing what you need it to do, right? As long as you understand that it’s still your responsibility, that this is just a tool, that it’s just helping, but that you are ultimately the responsible party. If you understand that, I think you can use AI in a bunch of wonderful, ethical, and interesting ways. And alt text, I think, is one of those ways. But I don’t think it’s ethical, or it’s an ethical use of AI to just have AI scan your whole site and add alt text to all your images. That’s not it, because that is abdicating your responsibility, and I think that’s unethical, an unethical use of AI.

And you can find tons of plug-ins, tons of services out there in the WordPress space, and others, in Shopify, elsewhere, that will do that for you. It’ll use AI to just generate in bulk all your images, and I don’t think that’s an ethical use.

[00:21:21] AI For Alt Text: A Bit of a Mixed Bag

Deneb: All right, so let’s talk about AI for alt text and why just having AI generate everything isn’t ethical.

It may be fast so that’s a pro, and AI can follow directions pretty well in terms of length requirements if you give it to them, and the level of detail that you want. So AI’s pretty good at following things, and so it… And it can do things really fast compared to humans. And because of that, it will reduce the cognitive load on the content creators.

So the alt text is more likely to get written, and because I think that some alt text is, in my opinion, is better than no alt text, right? I would rather have alt text on an image that creates a feeling in the visitors than just call it decorative and not give it alt text. I’d rather that.

But still it’s not done the right way, I don’t think. However, it can enhance the inclusivity by ensuring alt text at least captures nuanced image information consistently, and AI is getting better at describing things.

It does have some cons, right? Some of the cons of using AI for alt text is AI-generated alt text may sometimes miss contextual subtleties important to users.

So the content creator knows why they included that image, and the detail in that image that might be the essential detail might or might not get described by the AI as it’s generating that alt text. So that’s one con or one problem. Another is risk of over-reliance on AI. So if the more AI is used and successfully the more people depend on it, and they, then they start assuming that it can do things that it can’t or shouldn’t do.

And so this, the risk of over-reliance could reduce the human oversight or editorial judgment as we do tend to just cede our responsibility to the AI. Another problem is that the AI alt text that it generates is only as good as its training data, which because its training data is like the sum total of alt texts out there on the web that is part of its data set, its training data set. There’s some good alt out there, but there’s a whole lot of bad alt too, right? And because that training data is fairly poor I think overall, you tend to … If you let AI just do its own thing, you tend to get more of that.

So I’ve got a kind a link here at the bottom of this slide. It’s to the Pleated Jeans blog from about a year ago called Funny Times AI Failed At Describing Weird Images. Now, if you happen to go there, if you have some time, and just go and look at those images, you can see, oh man, this is bad. Now, that was a year ago. And I think it’s possible that AI has gotten better at figuring out weird images.

But it’s still, it’s still only as good as its training data is. So we got a potential problem there. Okay.

[00:24:45] The Wrong Way to Use AI for Alt Text: Automatic Generators

Deneb: Let me get into a wrong way to use AI for alt text. And this is a specific example. And the specific example is Shopify, right? Since 2023, Shopify has included an alt tech suggestion ability in their Shopify “Magic” release, right?

And so what you have here on the left-hand side of this slide, we’ve got an image of a… kind of a screenshot of the place in Shopify where you can use AI to set the the alt text for the image. In this case, it’s, I think, a product image. And so you click a button, and it applies that to the alt text. The suggestion is decent. So this particular image is of a snowboard with a pretty cool image on it, on the snowboard. And the alt text that the Shopify Magic recommends is, “Blue and white snowboard with liquid branding on a white background.” Which is it’s all right. Maybe that’s what the content creator or the curator of this website would want known about this, but maybe they who actually like this product, would have something better to say about it than just that.

But so this this process has some… the Shopify Magic process has some problems, right?

First, the AI doesn’t know which aspects of the image to include, right? It’s just describing the whole image. Maybe that’s not the important things that it’s picking out though. The AI gives a single option which is a problem because then the human doesn’t get to choose the best one.

It’s just you’re stuck with it, right? The AI doesn’t follow length recommendations. There are so many times when I’ve had Shopify generated an alt for some of the images, and they’ll come out, 140, 150 characters long, which is a little beyond what the recommended length is. The AI doesn’t prioritize the most salient feature of the image. Doesn’t prioritize it, right? It doesn’t front load that. It doesn’t get it into the front because it doesn’t know what the most salient feature is, right? It’s AI. It’s dumb. It’s never ridden a snowboard. And the final problem that I look at here is the AI usually doesn’t have any context for this image. Now, this image might be used in a page someplace, or it could be used as a product photo, or it could be as … But AI doesn’t know where it’s used. Or at least this version doesn’t show any indication that it knows. And so we’ve got some, some troubles. This is the wrong way to use AI.

This leads you to abdicate your responsibility to the AI model that’s doing it.

[00:27:31] Challenges to Effective AI Alt

Deneb: All right.

We’ve talked about some of these challenges already. On this slide, I have a little just a table saying, “Okay, challenge. What’s the challenge?” And then solution column of this table is empty.

Our challenges so far are poor training data, not prioritized or front-loaded, it’s got no context, and finally an over-reliance on AI. And then my question on this slide, the challenges to effective AI alt, my question is how can AI in a WordPress plugin overcome these problems? And so that’s what we’re gonna get to throughout the rest of the presentation.

[00:28:12] Poor Training Data

Deneb: Looking at some of these problems, poor training data. All right? This is a core limitation of historical AI alt text or AI generated alt text, right? The training data is not that great. So that’s a problem. The data or the training set that the AI is trained on has a bunch of slop, right?

It’s got filings sometimes end up in alt quite a bit SEO keywords get stuffed into alt a lot of times. Unnecessary words get put there. Length is all over the board, right? So AI replicates these mistakes, right? Because they’re in its training data. That’s what it was trained on, and so it’s gonna do more of the same.

If we want to avoid having it do more of this problematic the same, i f we leave it to its own devices, it’s not gonna overcome those challenges.

[00:29:09] We Keep it Honest

Deneb: So how do we do it? We keep it honest, and the way we keep it honest is by continually feeding it instructions to tell it not to do those things, right?

And the ability to continually feed those instructions to the AI is one thing that a plugin is actually really good at, right? You can make it a WordPress plugin so that it always feeds your certain standards to the artificial intelligent model, telling it, “Oh, do this, don’t do that, don’t do this.”

And as it does that, the model is pretty good at following instructions. So if you tell it every time, “Keep this at an appropriate length. Don’t say image up,” right? That sort of thing it’s gonna follow that. And so as long as we continually feed those instructions, we do overcome this poor training data problem.

Okay, moving on to, okay, so as I realized this, it was actually because I had a client, we had a client at SeaMonster Studios that they redid their website and then or we redid their website for them. As they got their hands on that website, they started moving over their thousands of images, and so every day they’d add 50 more images and I just saw their image, their media library get fuller and fuller, and I realized they didn’t give alt texts to a single one of them. They maybe did, but not many of them, right? There were just dozens and hundreds eventually of images that didn’t have AI, and I said, “That’s a problem. Maybe I can help.” Or that didn’t have alt, sorry. “That’s a problem. Maybe I can use AI to-” To help them fix that problem. And as long as I can keep the AI honest by feeding it this, these instructions maybe it’ll be an okay thing to use the AI to go through all these dozens and dozens of images every day.

[00:31:16] Polly Alt

Deneb: And I found that it did. But in order to do that, I made first just a local plugin that I called Polly Alt, right? So play on the word Polly, because I wanted to, I wanted it to give the humans who were doing this alt text, give them some options, that they were in charge of setting that alt, not just rolling through it.

But, in this plugin, I had to say “Okay, how am I going to make sure that my model doesn’t do the wrong things?” So I started with some, some pretty basic instructions and these are just a part of it, right? Alt text should describe what sighted visitors see in the image. Aim for around 125 characters.

Avoid starting with “image of” or “photo of”, right? Which are some of the big problems with the alt text out there in the world that would serve as the training data for this AI. So I said, “let’s start with some base instructions,” and then the instructions now have gotten more involved.

But that’s… initially, that’s what I started with. These instructions centered the AI model on the task at hand, and got it looking for what to have what we are looking for, which is what we want, and also tell it to avoid what we’re trying to avoid.

There’s things that we can put in, into that instructions telling it to avoid keyword stuffing, right?

Avoid use of “image of” or “picture of”, right? So since the plugin feeds that instruction with every single image that goes to the model, it’s giving it some good, some guidelines, right? Okay.

[00:32:56] Generation in a Vacuum

Deneb: What’s another problem with using AI to generate alt text? It is that as we saw with the Shopify, the generation of that AI, of that alt text would happen in a vacuum, right?

The image would be sent to the model, and a description of the image would come back, and that’s all you’d have, right? And often with most of the plugins, I think that’s kinda, that are out there, I think that’s what happens, right? They roll through your media library, they get assign everything in alt regardless of where it’s used.

And that’s a problem, ’cause you’re missing that crucial context. So without a reference to the surrounding paragraphs or the intent of the image the AI is just gonna generate descriptions in isolation, and it’s not gonna produce things useful to accessibility. The AI isn’t the one deciding to put the image in there. It’s the content creator, and the content creator is putting it in there for a specific purpose. If that purpose isn’t discovered, which it somewhat can be in terms of the context if you can include that context, but if that purpose is never discovered, then the AI is just gonna give poor alt.

It’s just gonna give very… It’s descriptive, right? So what do we do? We say context is king. We’re gonna give the AI the context of the images, right? In a WordPress plugin kind of environment, the plugin can actually go and find the paragraph context from within Gutenberg or Elementor or Classic Editor, and it can feed that context along with the image and the instructions that we give right to the model.

Now, the model does have the context, and so there’s a much greater chance that it will produce alternative text that that benefits that and in the context that it is.

[00:35:02] Giving Polly Context

Deneb: And so if we make context king there we have it. So how do we do it? We gotta give the Polly the context.

So at the point of insertion in a Gutenberg block or in an Elementor page, and you’re using the image block or in the classical editor, it’s easy for Polly this the plugin, to grab the context of the surrounding paragraphs and to feed that along with the image and the instructions to the model, right?

And so I have here on this slide some of the instructions I’m giving to the model, which is at least the first variation should connect what’s visible in the image and what the surrounding text is about. And if the image relates to surrounding text, that connection belongs only in the explanation, right? So I’m telling the plugin what to tell the AI model to make it do what I need it to do, which is to take the context into account when writing the alt text.

And then there’s the second part. If the image relates to the surrounding text, that connection only… it should only be in the explanation field.

I’ll show you that in just a minute when we get into what the what the interface actually looks like. But there’s a space that describes or explains to the content creator what the model thinks its alt text is doing. And they can look at the image, they can see the alt option, they can read the content or the explanation that the AI is giving for it, and make a good choice.

[00:36:36] Polly Alt Interface

Deneb: Now we have the page content is fed to the AI and to the human editor. So here we have our first view of the actual, Polly interface, right? Polly Alt Interface. It says, “Okay, here’s your image. You need to choose your alt text now,” right? It has a spot in this popup, in this dialogue I guess where it’s says the page content before and after.

In this case, on this it’s an image of a ship. And what comes right before this image in the page is, “Ready to jump in.” And what comes after it is, “Argosy Cruises is the best way to jump into the Sound without actually getting wet!” right? The plugin pulled that from the actual page where this image was gonna be used.

And so then it gives me some options, right? The first option says, “Passengers sit on the front deck of a white boat safely observing the water from the center of a narrow concrete lock.” And then right below that, this popup has the image focus, which is written by the AI explaining why it offers this option as an option and thinks it’s good.

And its image focus says, “Passengers on deck. By foregrounding the passengers, this description helps the reader visualize the jump-in experience mentioned in the surrounding text without needing to get wet.” That doesn’t show up in the alt text itself. That’s just to the reviewer. So the reviewer can think, “Oh, is that a good alt text? Should I choose that one?” And you can see from this view there are several other options below, right? I think by default, Polly offers three choices, can do as many as five. But it justifies each alt it suggests with this image focus that explains how it kinda connects or how it thinks it connects to the context.

All right, so I think that’s potentially pretty helpful.

[00:38:32] Every Usage is Different

Deneb: All right, moving on to if I can. Ah, another challenge for AI to produce alt is that every usage is different, right? You can have an image in your media library that is used in four or five different places in, throughout your site, and truly with alt text, each one of those usages, the alt text ought to be individualized, right? Because you’re using those images for different purposes in different places. Unfortunately, with WordPress, one thing that has always frustrated me is that you can’t really change… you’ve got your alt text set in the media library, and that’s what gets pushed out every place.

And there’s a limited ability to change the alt for different instances. But as we move on to how to overcome this the plugin can propagate page-specific alt. We give the alt or the AI the context of the images from the various pages to suggest unique alt on the different pages, but then the plugin itself can rewrite those page-specific instances of that alt that need to be different from the whatever is set by default in the WordPress database.

But it can replace the specific instances in the various pages with the context of where one generated for that instance. So we kinda overcome this problem that every usage needs to be different, and in WordPress kinda can’t be different with this plugin approach.

[00:40:14] Giving Polly Context

Deneb: All right moving on.

So here’s how we … I have the code here on this page. It says, “Giving Polly the context. Polly writes the page-specific alt that the user chooses directly into the page by rewriting the default alt.” And I’ve got the code here just showing that it, it uses some pretty complex replacements and things.

But, if we faithfully give it that context per instance of the image and overwrite the general alt text with that, then then we come to a place where it’s actually pretty helpful. So let me show you how Polly Alt handles multiple instances.

The first from the media library, if you click on the preview and generate button for some alt, the first thing that’s gonna pop up is gonna say, “Which spot on the ship?” So in this popup it has a little dropdown selector saying “Which of these spots are you gonna … are you looking to modify?” In this case it just has two options here because this image was only used in two pla- actually … Anyhow if it were used in five different places throughout the site, there would be five.

And so you first choose which instance you want to modify from the media library. So as you’re going through and you’re saying, “Okay let’s change it here. Let’s change it there.” And then it’ll tell you the context of that particular instance to remind you where you used it, and then you can choose from the choices it gives you.

This is one of the things that I kinda like most about Polly Alt, is that it gives you this chance to do this, which is something that I don’t think you really could do, at least not in any sort of systematic way, in with alt text in WordPress generally. It’s one of my complaints has been that it, you can’t do this. So Polly Alt lets you do it.

[00:42:14] Links Need Special Treatment

Deneb: All right. Another problem or another challenge, that AI could run into is that alt images, or I guess, images serving as links they need special treatment, right? Because you shouldn’t be giving a descriptive alt text for an image that’s a link that sends you to some site, some other site or some other page, right?

No, the image link, it needs a very focused alt text explaining where it’s going so that when a screen reader user gets to that image link, they know where they’re going. That’s the more important facet of the alt text in that case. And so if the links need a special treatment typically in WordPress, they can’t be given that special treatment. They can’t get that special treatment in any sort of programmatic way. But with Polly Alt since it can detect not only the context of what the words around it say, but also where if it is a link and if it’s going someplace, where that is, it can detect all of that just from what the information it has available to it from WordPress, from the backend of WordPress. The plugin can get the page content, context. It can get the DOM context, whether it’s a link or not a link, and then it can feed both of those into the AI, and the AI will suggest a different type of alt, right? If it’s a link, it’s gonna suggest the alt text that describes where the link goes, basically.

Then the human chooses the best option, and the plugin writes the appropriate alt to the image link on that specific instance. So that’s how we overcome that challenge.

[00:43:54] Links at Point of Insertion

Deneb: Okay Here I have on this page, it’s called links at point of insertion. I have an image of the Polly Alt modal that will pop up at the point of insertion.

If you’re in a Gutenberg page, for example, and you’re using this image as a link, once you assign it as a link, you come back to this part it’s gonna tell you, “Oh, this is a link,” right? And so it has a little bit of instructional text after it. It says that it’s a link. So this image functions as a link, so Polly’s gonna come up with a concise alt text for this link without using filler words like “link” or “to go” or “go to”.

So you as a reviewer, if you’re using Polly Alt, you’re going to be able to read that kind of description of what the goal is here. And then when it gives you, you can preview and generate, and it’ll give you some options. You can choose the best one that does what you need it to do, which in this case is tell what where the link goes essentially.

[00:44:58] Links from the Media Library

Deneb: But here we have a separate view, ’cause now this is from from the Media Library. If you chose preview and generate button in the Media Library for this image, it’s gonna pop up and it’s gonna say here, “Hey, this is a link.” And it’s gonna tell you the same information.

It will give you some page context, but then in, in its options, it’ll give you three options as it always does, three to five. And you’ll now be able to choose which of these options best meets this goal of telling the link destination. And so I think that’s a pretty helpful process that I don’t know a better way to do or even really a good way to do in WordPress failing this or without a tool like this.

[00:45:50] Solutions to Effective AI Alt Text

Deneb: So all right. So we’ve talked about the problems. We’ve talked about the solutions, right? We come to this table from before, right? With the challenge was poor training data. Our solution is human oversight and clear AI instructions, right? That’s how we overcome this poor training data problem.

When the challenge was no context, our solution is to feed the context programmatically to the agent, to the the model, the AI model so that it can actually give us context aware alt.

Our problem was not prioritized or front-loaded. This kind of goes back to the first problem of the poor training data. If we consistently instruct the AI to front-load the most important part then it’s gonna do that consistently.

And I guess the thing that we just talked about the image links, that kind of falls into this no context challenge. So but anyhow these are… the top three challenges are those.

[00:46:48] Over-Reliance on AI

Deneb: But then our last challenge, if you’ll recall from slide 11, I believe it was, the last challenge was this over dependence or over-reliance on the AI to do your job, right? And this tendency to, once you see that AI has done a decent job a couple times the tendency to just say “I’ll let it do it. I’ll let it do its thing,” right? Let’s talk about how this is getting into the benevolent friction part of the talk, right? How do we introduce a little bit of friction to make sure that doesn’t happen, that you’re not just abdicating your role? So the challenge, over-reliance on AI can lead to decreased quality, right?

It can reduce human oversight and editorial judgment. It can make it so that you miss important nuances, and overall, it’s just gonna diminish the accessibility quality and user experience, right? ‘Cause if machines are doing the whole work of writing your alts, they’re gonna… it’s… the quality’s just gonna go down.

[00:47:49] Benevolent Friction

Deneb: So how do we overcome that? This is what I call benevolent friction, right? This is prioritizing doing the right thing over the easy thing, right? So you can find dozens of plugins at, in the WordPress repository of plugins that that’ll assign AI generated alts.

But they don’t make you do your job, right? And in fact, they make it impossible for you to do your job, ’cause they just roll through the list and do all of the reassignments of that, that alt. You’ve got no oversight. You can’t. There’s nothing you can do to change anything. You could go back after it has assigned all the alts and re-change it, but that’s not gonna happen.

[00:48:32] Polly’s Squawk

Deneb: What we need to do is we need to put in some benevolent friction to make sure that humans are actually doing their job. Okay, so the way Polly Alt plugin does that is by squawking, right? I say a Polly squawk is when in Elementor or Gutenberg or the Classic Editor, the moment you try to add an image without alt, Polly squawks at you, gives you a modal alert, right?

And I’ve got the code here on screen just showing, here’s how it constructs that squawk modal. But here’s what it looks like. I’ll show you. On this image we have, I guess this is in the medial… No, this is in the insert. So when you go to insert an image and you choose an image that doesn’t have alt and click insert, the interface Polly is gonna squawk, and there’s a little popup or a little modal that pops up saying, “Just a moment, captain. This image has no alt text. Blind and low vision visitors won’t know what it shows.” So it has two options: let’s add some alt or insert anyway. And so this means that, yes the human can choose to do the wrong thing, which is to insert this image without alt. You can do that, and you can click that insert anyway button, but you could also click the other button, which is let’s add some alt.

And then it will bring up the main modal that shows the context and the different options for you. And so Polly’s there to help you do the right thing, and if the choice is between clicking the wrong button, which is insert anyway, and clicking the right button, which is let’s add some alts, right? It’s just as easy to add the alt. Not quite as easy, but pretty close. And it forces you… The friction part is that it forces you to choose. So you’re choosing deliberately to do the wrong thing if you do the wrong thing.

Some other squawks that Polly will give you. It’s gonna squawk at you at various moments throughout the process. When you try to insert without alt, when you save a post with an image without alt, when you navigate away from a post that has altless images or a page, right? When you leave the Media Library with images that don’t have alt text set. None of these things, of course, happen in WordPress.

[00:50:49] Polly’s Power of Annoyance

Deneb: Now the Equalize Digital Accessibility Checker probably maybe does some of these things at least alerts you that there’s no alt, I think. But but Polly actually helps you, right? It tells you that these things are doing. And the result is, it gets pretty annoying, right? I call this slide Polly’s Power of Annoyance, and it just has three views of different squawks that Polly may throw up there for you telling you that you gotta do the right thing.

So the first one says, “Just a moment, captain!” Right? “There’s no alt here,” blah, blah, blah. The next one says, “Steady as she goes, Captain! Hold fast. There’s one hidden image that doesn’t have alt.” And then, another time it squawks, it just says, “Blargh! Three images are missing alt. Let’s fix them,” right?

And then in each of these modals that pops up alerting you to these problems, there’s a choice. The correct choice is let’s add it, right? Guide me through them. Let’s add some. And so you can always make that choice. And eventually with all these squawk so if you try to add some images and ignore the first squawk you say insert anyway, and then you try to save the post, and it squawks at you again, right?

You could say, “All right I’m gonna ignore that again.” You can publish it any,” right? You can do that, but the more times it pops up, eventually you’re just gonna give up. It’s just easier to add it, especially when the actual method for adding it is fairly easy, fairly minimal.

[00:52:22] Polly Forces Oversight Without Even Trying

Deneb: You just choose among some of the best options. All right. This slide I’ve got a slide that says Polly forces oversight without even trying, and that is because with the way the Polly Alt plugin is structured, it gives you three choices. So you can’t just say, assign alt to all these images. No. With each image it has three choices you should read. Now, you could just always choose the first image, and that makes it pretty quick. But if it only takes an extra five seconds to read through the three different choices and as the content creator you decide, “Oh yeah, this alt does what I’m looking for best,” or, “No, that one does.” It’s pretty easy. And each of these options has a different focus, right? So another part of the instructions that I give to Polly is that it should in each of these variations, it should highlight a different or focus a different aspect. And so if you’re using the image for something that connects with one part…

Anyhow, it’s hard to explain how valuable this is, but I think it’s pretty valuable having different options because instead of focusing on a part of the image that really doesn’t have anything to do with what you’re trying to accomplish, one of the three choices will hopefully focus on the thing that you do want highlighted.

Okay and because, once again, because they have to choose which of the options, that’s oversight, right? Now, that is putting the human, keeping the human in the driver’s seat and not letting them abdicate their authority.

[00:54:05] Benevolent Friction: Forcing a Choice

Deneb: Okay, next. Our next one talks about benevolent friction, right? Forcing a choice. So here is just a view of this the modal, the selection modal, the top of the selection modal has a zoomed in view of the image that you’re assigning alt to. It’s zoomed in as much as it can be which gives a, I guess the little scroll bar in only one direction, right?

At least you don’t have to pan around the image and look to the top and, right? No, it’s just up or down in this case, or left or right, if it were a really tall image. Or a really wide… I don’t know. Anyhow you have this unidirectional modal that you can pan around. And then on the bottom the AI gives you three or more options so that you, the human, exercise your human judgment to pick the one that fits. Or there’s also on each of these AI options, there’s an edit button. So if you like one of AI’s suggestions but just wanna tweak it a little bit, you can do that.

So once again, giving ways to keep humans in charge. And to force them to be in charge by making them, forcing them to make a choice.

[00:55:25] Overcoming the Challenges of AI Generation in a WordPress Plugin

Deneb: Okay, let’s see. So to review the, some of the challenges of AI generation of a WordPress plugin. AI alt text generation ability has historically only been as good as its training data, which has been pretty poor, right? Up to this point. Now, I think the AI is getting better at overcoming that on its own. So it can realize how poor some of the things are. But we can help it along too by continually feeding the instructions to the AI to keep it honest.

Another challenge is the AI generated alt text may sometimes miss contextual subtleties important to users. And this is usually because the AI doesn’t have access to the context, so it’s generating that alt in a vacuum. But the way we overcome that is to give the AI the context of the image, which we do through the plugin by feeding that context in, whether the context comes from a page, a post, a template, a block, whatever. So we fix that challenge that way.

Next challenge is the risk of over-reliance on AI that reduces human oversight or editorial judgment which we’ve talked about a fair amount. The way to overcome that is this benevolent friction rather than prioritizing easy, right? You get many other plugins that offer to generate your images with alt prioritize the easy way, right?

They say, “Okay, should we roll through your library and set alt for everything? And it’ll be good, it’ll be descriptive,” and you say, “Sure,” but by doing that, you’re just abdicating, right? Let’s make it so that there’s a little friction in there. First, the friction to make you do the right thing. Second, the friction of actually choosing so that you are actually discharging your duty.

Okay now we’re basically to the the part of the show where you guys tell me what questions you’ve had. I haven’t looked at the Q&A or the chat at all. I know lots of people try to answer questions as they go, but I just run.

[00:57:19] Q&A

Amber: Yeah, no problem. Thank you so much. I actually am curious. I have a question of my own. There are a few in the Q&A, and I’m gonna run through those, so if anyone wants to add any more or if you wanna up vote, that will help me organize my questions.

But I was curious because it seemed like during your presentation, Deneb that you were talking a lot almost like there’s very few images that should not have alt text.

And I know that’s a thing that we discuss internally a lot, is this decorative, is it not? And I’m curious if Polly has integrated with the new feature that came out in WordPress 7.1, where there is a checkbox where content authors can intentionally say, “This is decorative,” and if Polly will then stop yelling at them. Squawking. Wait, squawking is the term-

Deneb: Yeah, squawking.

Amber: Which I love, by the way.

Deneb: Yes. Amber, you bring up a wonderful point, and it is… I had meant to include that in the presentation. There is a way to say, ” This image is decorative.” But you’re right that this is where I expect to get some blowback from some people.

Because my opinion, which is not as valuable as a real screen reader user’s opinion, but my opinion is that if I’m saying- I’m gonna mark this image decorative. This means I don’t think screen reader users would need or want this image. I don’t wanna make that judgment call. I’d rather have … a front-loaded…

Amber: Okay, out of respect, time to-

Deneb: I’ll finish and then

Amber: I wanna hear you … because there are, I’ve seen people still build websites the old way where it’s like a divider, it’s like a PNG of a line.

Deneb: That’s true.

Amber: That’s true. Or like a little icon that is used as a background image but some builders actually put them in as images.

Deneb: Yeah.

Amber: And then you just make them not, like they’re semi-transparent. Like-

Deneb: Well, that’s true …

Amber: There are definitely images that are purely decorative. Yeah. Not photos as much, right? But…

Deneb: Yeah, that’s true. A point well taken. There are images that are purely decorative. And so there, there is actually in Polly there’s a box that you can check for that that says this image is decorative.

Amber: Does it recognize the core, the one that came out in 7.1, do you know, or not yet?

Deneb: I don’t know that it does.

Amber: Okay …

Deneb: In fact, I think it probably doesn’t. This is actually one of the features that I need to work on a little bit. Like it’s a little bit inconsistent right now.

So I think-

Amber: All plugins are starting points, right?

And that’s the fun of them.

Deneb: Exactly.

Amber: You work and you iterate and all that.

Deneb: Yeah.

Amber: Yeah.

Deneb: Yeah, for sure. And then I wanted to

Amber: Is it…

By the way, I forgot to ask. Is it available like on your GitHub account or somewhere if anyone wanted to look at the code or play around with it?

Deneb: Yep. So I did finally put this this last screen up the contact and-

Amber: Oh

Deneb: get involved, right? And

Amber: There we go.

Deneb: Okay.

Amber: Yes. Perfect.

Deneb: All right. So- Contact and get involved, right? I would love for some help, right? I’m not the most amazing developer. I, I do okay. But here’s my contact, so all of you who wanna scream at me and get into this big philosophical debate about which images should have alt and not, that is something I, I would welcome.

I’d love to hear what you guys’ opinions are particularly from actual screen reader users. But my standard is I’d rather err on the side of giving more and not deciding that a screen reader user isn’t gonna wanna see that image, right?

But but anyhow my, my email, deneb@seamonsterstudios.com.

The Polly Alt plugin on the WordPress plugin repository at wordpress.org/plugin/Polly-alt. And then if you wanna get involved, which I do hope you will even if it’s just trying it on your site and then sending me a list of bugs because I’ve I know there are a lot of things that are that still need a lot of help. So it’s at github.com/seamonsterdeneb/Polly-alt. Please do yeah. And, if you wanna report the bugs, that’s great, but if you wanna fix the bugs, do a pull request even better. I’d love it.

Amber: Yeah. Awesome. Thank you so much for sharing that, and I think-

Deneb: Yeah

Amber: We’ve also, Paola’s put at least a WordPress plugin in the chat as well-

Deneb: Great.

Amber: … For folks. Let’s see. And then I’m gonna run through Q&A. So on the topic of decorative images, someone asked, “How would AI know if it’s decorative if even a human is sometimes not sure?”

“Mood setting images, for example.” Do you wanna share some thoughts on that?

Deneb: Yeah. Okay, so the AI is gonna describe every image. There is a way in Polly for you to say, “No, this is decorative.” So you check that box, and then it’ll stop squawking at you, like you said. So you are the one who decides that.

The AI, I wouldn’t trust it to say if an image were decorative or not. It’s just gonna describe everything.

Amber: I think that… I think I’ve seen this in some of the lawsuits against websites that are using the accessibility overlays. Because one feature of accessibility overlays is they’ll add alt text when it’s empty, or-

Deneb: Oh, yeah.

Amber: The alt attribute is missing. And I’ve noticed in reading the lawsuits they’ll say things like, “The alt text wasn’t correct,” and they’ll also be like, “This image was definitely decorative, and it had alt text added to it-” … because the overlay was doing it. So I think you’re right. AI is just gonna assume every image should be described.

Deneb: Yeah. Yeah, it is. And, my personal opinion is I would err on the side of describing more images than not, but once again, it’s… that’s a human judgment call that humans should make.

Amber: Yeah. We’ve been on user testing sessions where an image that we were like, “This is decorative, leave it empty,” screen readers still know it’s there.

They’ll say, “Screen readers won’t announce it.” Okay, it stops it from reading the full URL to the image, right? But they can still tell there’s an image there, and I’ve had them be like, “There’s an image here. It doesn’t have a description.” And then we’ll be like, “Oh, that’s because it’s this,” and we’ll describe it, and they’re like, “Okay, yeah, that’s fine.”

But they know, and if they don’t have … a sighted person telling them what it is, then they might be like-

Deneb: Interesting.

Amber: … They can tell there’s images without description on page.

Deneb: Huh.

Amber: And then they might think that it’s a mistake. So it is… I see what you’re saying about erring on the side of give them some information so they can choose.

Deneb: Yeah. And then I also have… One other thing is I met a guy at AccessU last year who said that he uses the native describing ability of his screen reader, right? So he’ll actually ask his screen reader what the image… Some features of the image, and it’ll tell him things about the image. And so that might be an argument to keep the alt a little more basic, but I still think it’s a better idea to be a little more descriptive. And let screen reader users jump away if they want.

Amber: Yeah. So there are two questions here about length-

Deneb: Okay.

Amber: … From Jacque. And Jacque asked one, “Is there a technical limit to the number of alt text characters?”

And then two, “Does overly lengthy alt texts cause the user’s device to quote crash?”

Deneb: I-

Amber: So what do you want to say about alt text length?

Deneb: Yeah, so the the standard recommendation is 125 characters, it’s that way because I think it was JAWS back in the day, that was as long as it would read, right?

It would just stop reading the alt text after that. And now I think it’ll read until the cows come home. I’m not sure. Maybe it stops at some point. But I’m… It could be really long, I think. However, if you get lengthy if it’s beyond, I think about… I think 125 is a pretty good standard even though it’s not because of a restriction.

The problem is the alt text does not give a way for the screen reader to reread a sentence or go back and reread a part of it again. It’s all or nothing. And so if you have… If you want to include a really long description you should do it in a different element, right?

That they can navigate to and then they can skip around sentence by sentence, word by word instead of putting that really long description into the alt text. At least that’s my understanding. What do you think, Amber?

Amber: Yeah, so I find this, this is always really interesting, and it’s one that came up when we first created our…

We have a Needs Review item in Accessibility Checker for long alternative text. And we had set it to, I think, 150 characters, ’cause that’s what we- … had decided as a threshold. And then NASA started using our plugin.

Deneb: Oh.

Amber: And we had all these conversations with content creators at NASA because…

And there have been articles in mainstream news about their alt text game. They are really … on and… Because if you were gonna… This is a perfect example of why AI can’t always do good alt text, right? Because they would always just be like, “Stars in the sky.” Whereas they really… they can name the stars.

They’re- Yeah … scientists, right? They can tell you all kinds of stuff about what the gray cloudy thing to the side of the night sky actually means, right?

And so they were like, “We can’t, of course we can’t generate this,” and we really, we don’t want it to just be like, “Here’s another picture of the Milky Way,” right?

Yeah. It needs to be more descriptive to explain what it is and why we’re showing it and all that kind of stuff. And- Yeah … we ended up making our alt text length warning, and this is horrible that I don’t know. It’s in our documentation. I don’t remember. But I wanna say it’s now 250 characters.

So we made it quite a bit longer. It is filterable, so any developer for any specific website could change.

Deneb: Oh, cool.

Amber: And I think they even made it bigger for their specific use case.

Deneb: Oh, okay.

Amber: It might be like 500 characters or something long.

Deneb: Do you-

Amber: But that said, I do think that to your point of if it gets very long, it’s difficult to navigate back through parts of it, and so I do agree that more concise alternative text is better because having more detailed descriptions like that explaining what the image is probably actually beneficial to sighted users, too.

Deneb: Yeah.

Amber: And when you put it only on the alternative text, the sighted users miss out on the explanation of what this image of a star in the sky actually is, right?

Deneb: Yeah, for sure.

Amber: Or graphs is another one where I’ve seen people try to, if it’s a PNG of a graph, and then they try in their alternative text to explain everything that’s happening in the graph that might have multiple bars or lines and bars-

Deneb: For sure.

Amber: And, like, all that stuff. And my answer is how about we have a short thing that says, graph of stars and their location, like bar chart of stars and their location. And then we’ll even say in the alternative text you can write something like, “Tabular data follows image.”

Deneb: Yes.

Amber: And then right below the image you have maybe it’s an accordion, if you don’t want to show it visibly on the page that someone could expand to get a table that has the same data that would be found in the image of the graph. Instead of trying to list it all out in an alternative text. And then that’s really useful for people, because there’s a lot of sighted people too, that they understand information better in a table form than a graph.

Deneb: Yes.

Amber: Some people like graphs, some people like… so and you can do the same thing with those longer descriptions. You could have an accordion or a collapsible component. I know Joe Dolson’s WP Accessibility plugin does something like Twitter. There’s a setting you can turn on that would make a little button on the image that you can then expand and anyone can read the alternative text.

Deneb: Great.

Amber: So I do think there are arguments for making it available to sighted people too.

Deneb: For sure. But then like you said, including that information in the alt itself. So your alt is a little more basic, but you say, “Here’s where you go for the more expanded.”

Yep.

Amber: Yep. Let’s see, Alice said “What other aspects of accessibility with AI do we need to consider when selecting AI tools?” So this is like going beyond, and I know we don’t have a lot of time, but do you just generally have anything, any thoughts beyond alternative text and AI that you would like to share?

And my lights just flickered because we’re having a thunderstorm, so if I disappear Paola will pop on here. It’s been happening a few times.

Deneb: As far as using AI, I like AI used for alt because it can be very narrowly controlled with the instructions that you give it. And it can be, your instructions can be pretty uniform and can be consistent.

If you’re using AI on different tasks, which I do, right? I use AI all the time. But the more detailed you are in your instructions to it, the better. And so if you want it to be accessible in certain… you gotta probably think what are the ways that it needs to be accessible that its training data might get wrong?”

And explain in your instructions that it needs to do this and that. Then I think it’s gonna do a, it’s gonna do a better job for you. It’s gonna help you, And if it gets it wrong the first time, right? And oh, no, it did all these things wrong. First of all, you as the person have to check it, what it did, and actually see if it is inaccessible or whatever.

But then feed it the instructions back in. Say, “Hey, no. Okay, you gotta fix it. Fix this and that.” And then you gotta double-check it, ’cause sometimes it’ll do regressions in other ways, too. As long as you as the human are always in charge and as explicit as possible to this dumb servant that you have, telling it exactly what you’re looking for, I think it can do a pretty good job.

Amber: Okay. So Caderius had a really great question. They asked, “In an editorial environment where authors upload dozens of photos a day, how do you prevent benevolent friction from turning into cognitive fatigue? For example, I can imagine a scenario where a writer’s just clicking the first AI option to get past the modal instead of genuinely reviewing the content.”

Deneb: Yeah. There’s always gonna be a gotcha, and that’s kind of the gotcha. Because even with that, when Polly pops up its modal, it’s pretty easy if you did wanna still give up your responsibility to pick the first one or whatever, right? It is easy to do that. But I think if you just talk with people about it, say, “Hey, okay, it’s your responsibility to assign this, this alt text.

This plugin is gonna help you, but it’s still your responsibility,” right? This’ll take it from instead of taking you one minute to do the alt, it’ll only take you 20 seconds to scan through and decide which is best, right? But you should still do those 20 seconds. And you just gotta, you just gotta be real with them.

If they’re uploading a dozen images and Polly is giving them three choices for each image, right? That’s only 36 choices that they’ve gotta- there’s 36 things they have to read in order to assign good alt to each of those, to those images, and so that’s definitely worth it.

That’s an amount of work that I think is definitely worth it.

Amber: Yeah. So it’s like a training. Yeah. And helping people understand why this is important.

Deneb: Yep.

Amber: Yeah.

Deneb: And I like I tried to put into Polly information about, so for people who haven’t thought a lot about alt, information in there about why it’s important and about how to do it right and stuff so that it’s not just, they’re not just going through the motions to go through the motions.

If they read those informational texts, it’ll help them to understand.

Amber: Awesome. I think we have time for maybe one more. So this is also on the lines of cognitive, but going to the AI side of it. Jac had asked, “Does a tool exist to help website designers and developers optimize alt texts in particular for people with cognitive disabilities?”

And now, that’s got me thinking, have you ever thought maybe this is a feature in the future for you to test, like, reading level of the alternative text-

Deneb: Oh.

Amber: … To make sure it’s simple?

Deneb: Yeah. Interesting.

Amber: Or maybe you could put that in your prompts, like make sure it has a Flesch-Kincaid grade level of eighth grade or lower.

Or something like that, right?

Deneb: Yeah. Good. That could definitely be a… and particularly if that were something that were important in one particular website, right? That’s valuable. I could see having an option in the settings prioritize simpler alt for or a simpler, a lower read- whatever.

That’s a great feature request. He should, or this person should come and and do a pull request or and see what we can do to add that. I’d love it.

Amber: Awesome. Thank you so much. Thank you, everyone, and luckily it seems like my internet and electricity have held out here-

Deneb: Woo-hoo!

Amber: Despite a few flashes. I don’t know if you all have observed those. But thank you again, Deneb. And I do want to call out I love your WordPress Accessibility Day T-shirt that you’re wearing there-

Deneb: Yep.

Amber: … From a prior year. I noticed Adrianne… Yep, it says, “Focus: It’s a feature, not a bug,” and it shows a, a focus rectangle.

Deneb: Love it.

Amber: I noticed Adrianne mentioned that in the chat as well. I hope everyone has registered for WordPress Accessibility Day next week. It’s gonna be wonderful as usual. But please make sure-

Deneb: Yep!

Amber: … You go register to get access to all the talks. Thank you, Deneb, and we’ll see you next week.

Deneb: All right.

Access Ethical AI & “Benevolent Friction”: Rethinking Alternative Text in WordPress Web Presentation.

About the Meetup

The WordPress Accessibility Meetup is a global group of WordPress developers, designers, and users interested in building more accessible websites. The meetup meets twice per month for presentations on a variety of topics related to making WordPress websites accessible to people of all abilities. Meetups take place on the first Thursday and third Tuesday of the month at 10:00 AM U.S. Central (5 PM CET).

Learn more about WordPress Accessibility Meetup.

Summarized Session Information

Alternative text should communicate why an image matters in the place where it appears. Writing it well requires decisions about context, detail, wording, and purpose. AI can help with that work, but content creators remain responsible for the descriptions they publish.

In this session, Deneb Pulsipher explored an approach to AI-assisted alternative text built around clear instructions, page-specific context, and deliberate human review. Polly Alt, a WordPress plugin, provided a practical example of benevolent friction: small interruptions and choices designed to keep accessibility decisions in human hands.

Session Outline

  • Introduction
  • Philosophy
  • Beyond Minimal Alt Text: Pros and Cons
  • Overcoming the Cons: Screen Reader Capabilities
  • Enhancing User Control with Skippable Verbose Alt Text
  • Front-Loading Alt Text: Pros and Cons
  • Overcoming the Cons: Ethical AI Can Help
  • AI For Alt Text: A Bit of a Mixed Bag
  • The Wrong Way to Use AI for Alt Text: Automatic Generators
  • Challenges to Effective AI Alt
  • Poor Training Data
  • We Keep it Honest
  • Polly Alt
  • Generation in a Vacuum
  • Giving Polly Context
  • Polly Alt Interface
  • Every Usage is Different
  • Giving Polly Context
  • Links Need Special Treatment
  • Links at Point of Insertion
  • Links from the Media Library
  • Solutions to Effective AI Alt Text
  • Over-Reliance on AI
  • Benevolent Friction
  • Polly’s Squawk
  • Polly’s Power of Annoyance
  • Polly Forces Oversight Without Even Trying
  • Benevolent Friction: Forcing a Choice
  • Overcoming the Challenges of AI Generation in a WordPress Plugin
  • Q&A

    Introduction

    AI can assist with alternative text, but the quality of the result depends on how the tool is used and what the person using it understands about their responsibility.

    Alternative text provided a specific way to explore a broader question: what does it mean to use AI ethically? Deneb connected that question to everyday publishing decisions, including how much detail an image description needs, which information belongs first, and how to prevent automation from replacing editorial judgment.

    The central approach combined three elements: useful descriptions, informed human choices, and tools that support those choices.

    Philosophy

    The starting philosophy favored providing more meaningful image information when there is uncertainty about whether to include it. A minimal description, or a decision to treat an image as decorative, can omit something that contributes to the experience of a page.

    Images may communicate mood, personality, or emotional context alongside factual information. If a photograph helps sighted visitors feel welcome, curious, or connected to the content, that purpose deserves consideration when deciding what a screen reader user should receive.

    An equivalent experience can involve more than identifying the image’s main subject. It may also involve communicating an expression, an interaction, or another detail that explains why the image was selected.

    This was presented as a position open to discussion, particularly with people who use screen readers. More description also has costs, and genuinely decorative images still exist. The question is how to avoid withholding useful information while respecting the user’s time and attention.

    Beyond Minimal Alt Text: Pros and Cons

    More descriptive alternative text can provide richer context, greater nuance, and less ambiguity. It can help someone understand both the image and its relationship to the surrounding content.

    For example, identifying an animal as a dog communicates the subject. Describing the dog’s posture or interaction with its owner may also communicate the feeling that motivated the image’s inclusion.

    Descriptive content may offer a secondary search benefit, but the main purpose remains helping people access the image’s meaning.

    Additional detail also creates challenges:

    • Longer listening time: More words take longer to hear when someone listens to the full description.
    • Unnecessary detail: Information that does not serve the image’s purpose can become frustrating.
    • Cognitive load: A poorly organized description can make it harder to identify what matters.
    • Authoring and maintenance effort: Descriptions that take too much time or thought to produce may be skipped or neglected.

    The goal is to make descriptions informative and purposeful while keeping the writing process manageable.

    Overcoming the Cons: Screen Reader Capabilities

    Screen reader users can often move to another element before a description finishes. This gives them some control over how much information they receive.

    That ability makes the order of information especially important. If the essential meaning appears at the beginning, a person can hear enough to understand the image and then continue navigating.

    Skipping does not automatically solve the problems of long descriptions. Its usefulness depends partly on how the alternative text is written. A description that saves its main point until the end makes it harder for someone to move efficiently.

    Put the information someone needs most before the information they might choose to skip.

    Enhancing User Control with Skippable Verbose Alt Text

    One example used a photograph of a dog looking toward its owner. The opening description—“An eager brown dog looks up happily at its owner”—communicated the subject, action, and emotional tone before any additional detail.

    Those first words offered enough information to understand the image’s basic contribution to the page. A person interested in more detail could continue listening; someone who already had what they needed could move on.

    This approach treats description as something users can engage with at different depths. The opening should work even when the rest is not heard.

    Creating that opening requires judgment. The author must identify the image’s purpose, select its most important features, and organize them clearly. The more difficult that process feels, the greater the temptation to use an overly generic description or classify an image as decorative simply to avoid the task.

    A useful authoring tool should reduce that effort while preserving the person’s role in deciding what the image needs to communicate.

    Front-Loading Alt Text: Pros and Cons

    Front-loading means placing the most important information at the start of the description.

    This can improve navigation efficiency, reduce the amount someone must hear before understanding an image, and clarify the image’s connection to the page. It also encourages writers to identify a clear subject and purpose.

    However, deciding what belongs first is not always straightforward. A description may need to balance the image’s subject, the surrounding text, and a detail that makes this particular image relevant.

    There is also no guarantee that every person will hear the same amount. Someone may move on after only a few words, before the writer has fully expressed the main point.

    The practical challenge is to communicate essential meaning early without making the wording confusing or stripping out necessary detail. That takes editorial effort, and adding more requirements to an already difficult task can make good alternative text less likely to be written consistently.

    Overcoming the Cons: Ethical AI Can Help

    The session’s working definition of ethical AI centered on using AI to strengthen a person’s ability to fulfill their responsibilities.

    For alternative text, AI can help draft descriptions, adjust wording, and organize information. The content creator still needs to determine whether the result is accurate, relevant, and appropriate for the image’s purpose.

    That responsibility remains even when an AI tool can complete a task independently. A system’s ability to generate or apply text does not establish that the text should be published.

    Within this framework, automatically scanning an entire website and assigning descriptions without meaningful review was considered an inappropriate delegation of responsibility. Humans remain accountable for what the website communicates.

    AI For Alt Text: A Bit of a Mixed Bag

    AI offers several practical advantages. It can generate wording quickly, follow instructions on length and detail, and reduce the effort required to start a description.

    That support can make alternative text more likely to be completed, especially in publishing environments with many images. It can also suggest descriptive options an author might not immediately consider.

    The limitations are equally important.

    AI may miss the detail that matters most. A model can describe visible objects without understanding why the content creator chose the image. An accurate description may still emphasize the wrong feature.

    Successful results can encourage overconfidence. After several useful suggestions, someone may begin accepting later output without checking it carefully.

    Poor examples can influence output. The presentation raised concerns about the quality of existing alternative text and the possibility of reproducing weak patterns, such as unnecessary wording, inconsistent length, or descriptions that do not serve the page.

    Examples of AI misinterpreting unusual images reinforced the need to inspect the result. Better image recognition does not remove the need for contextual and editorial review.

    The Wrong Way to Use AI for Alt Text: Automatic Generators

    A Shopify Magic example illustrated the limits of accepting a single generated description without examining its purpose.

    The image showed a snowboard, and the suggested description identified its blue-and-white appearance, branding, and white background. That description captured visible features, but it did not establish which features mattered to someone considering the product.

    The example raised several concerns:

    • A single suggestion narrows the decision to accepting or changing one interpretation.
    • The model may emphasize incidental details instead of the feature most relevant to the page.
    • The important information may not appear first.
    • The output may miss the requested length target.
    • The generation process may lack information about where and why the image is used.

    The broader concern was a workflow that encourages users to treat a generated description as finished. A plausible description still needs a person to assess its usefulness.

    Challenges to Effective AI Alt

    Four recurring challenges provided the framework for the rest of the presentation:

    1. Poor training examples or inherited writing patterns.
    2. Descriptions that do not prioritize important information.
    3. Missing context about the image’s use.
    4. Over-reliance on automation.

    A WordPress plugin can address parts of this problem because it sits inside the publishing workflow. It can supply repeatable instructions, collect surrounding content, and present suggestions for review at the point where the author is working.

    The remaining challenge is designing that workflow so that reviewing the output remains a meaningful part of the task.

    Poor Training Data

    Weak alternative text can take several forms: filenames used as descriptions, strings of search keywords, unnecessary introductory words, or inconsistent levels of detail.

    The concern was that AI may reproduce familiar patterns even when those patterns do not support accessibility.

    Generating text is not the same as applying a useful editorial standard. A system needs explicit direction about what to include, what to avoid, and how to organize the result. The output then needs to be checked against those expectations.

    We Keep it Honest

    A plugin can send the same core instructions with every image-generation request. This makes the intended standards part of the process rather than something an author must remember to type each time.

    Those instructions can address length, unnecessary phrases, keyword stuffing, and the need to prioritize meaningful information. Consistent prompting provides a way to steer the output, while human review checks whether the instructions produced a suitable result.

    The practical motivation for Polly Alt came from a client website with a growing collection of images. Content creators were moving thousands of images into the site, sometimes adding dozens a day, while many arrived without alternative text.

    The task needed to become easier without disappearing from the author’s responsibilities.

    Polly Alt

    Polly Alt began as a local plugin intended to give content creators useful choices.

    Its initial instructions included:

    • Describe what sighted visitors can see in the image.
    • Aim for approximately 125 characters.
    • Avoid routine openings such as “image of” or “photo of.”
    • Avoid keyword stuffing.

    The 125-character figure was used as a writing target in this workflow. The later Q&A clarified that it should not be treated as a universal technical maximum.

    The instructions became more detailed as the tool developed. Their purpose was to establish consistent expectations each time the model received an image.

    The plugin’s central design choice was to offer alternatives for a human to select and edit.

    Generation in a Vacuum

    An image alone does not necessarily reveal its purpose on a page.

    Sending only the image to a model can produce a visually accurate description that misses the point of its inclusion. The same photograph might support different ideas in different articles, product pages, or sections of a website.

    Context helps determine which visible details deserve emphasis.

    Within WordPress, a plugin can retrieve surrounding paragraphs from the editor and send them with the image and the writing instructions. The presentation discussed this approach in Gutenberg, Elementor, and the Classic Editor.

    Providing that context gives the model more information about the image’s role. It also gives the reviewer a stronger basis for deciding whether a suggestion fits.

    Giving Polly Context

    At the point where an author inserts an image, Polly can collect nearby text and include it in the request to the model.

    The instructions ask for variations that take the surrounding content into account. They also separate the proposed alternative text from an explanation of the suggestion’s focus.

    That distinction matters. The published description needs to communicate the image’s relevant content. The reviewer may additionally benefit from an explanation of why the model emphasized a particular detail.

    The explanation supports the editorial decision; it is not automatically part of the image’s alt text.

    Polly Alt Interface

    The Polly Alt interface brings together the image, nearby page content, suggested descriptions, and explanations for the reviewer.

    One example used an image of a boat with surrounding copy about enjoying an Argosy Cruises experience on the Sound without getting wet.

    A suggested description focused on passengers sitting on the front deck of a white boat inside a narrow concrete lock. Its accompanying image focus explanation connected the emphasis on passengers to the experience described in the surrounding text.

    The interface offered several alternatives—typically three, with the ability to provide as many as five. Each option could emphasize a different aspect of the image.

    This gave the reviewer several things to evaluate together:

    • Whether the description accurately represented the image.
    • Whether its focus matched the page’s purpose.
    • Whether the most relevant information appeared early.
    • Whether another option, or an edited version, would work better.

    Every Usage is Different

    A single image can appear in several places across a website, and each placement may require different alternative text.

    An image’s Media Library description provides a general starting point. However, that wording may not fit every page where the image appears.

    The workflow discussed in the presentation addressed the difficulty of managing those differences systematically. Polly could use the context of a particular placement to suggest a description for that instance, then replace the default wording in that location.

    The important distinction is between describing an image as a stored asset and describing what that image contributes in a particular use.

    Giving Polly Context

    When working from the Media Library, Polly could ask which use of an image the author wanted to update.

    A selection dialog identified the available placements. An image used in several locations could therefore be reviewed one location at a time.

    After the author selected an instance, the interface displayed its context and generated options for that use. The chosen wording could then be written into the relevant page instance.

    Selecting the placement before selecting the description helped connect the editing decision to the actual page experience.

    Links Need Special Treatment

    An image that functions as a link needs alternative text that communicates its destination or purpose.

    A purely visual description may leave a person uncertain about what activating the link will do. In that situation, the link’s function becomes the priority.

    The plugin could use information from WordPress and the page structure to determine whether an image was linked and where the link led. That information could accompany the image and surrounding text in the request to the model.

    The resulting suggestions could then focus on the destination. A human would choose the most appropriate option, and the plugin would apply it to that particular linked image.

    Links at Point of Insertion

    When an author assigned a link to an image, the insertion workflow could identify the image’s linked status.

    The interface explained that the image functioned as a link and that the suggested text should concisely communicate its purpose. The instructions discouraged filler such as “link,” “to go,” or “go to.”

    This helped the reviewer assess the options against the right objective: will someone understand what this link does?

    Links from the Media Library

    The Media Library workflow could also identify an image instance that functioned as a link.

    The reviewer received the linked-image explanation, page context, and several generated options. Those options could then be compared according to how clearly they communicated the destination.

    This allowed authors to make a function-specific decision even when they began in the Media Library rather than in the page editor.

    Solutions to Effective AI Alt Text

    The proposed solutions connected each problem to a specific part of the workflow:

    • Weak output patterns: Supply clear instructions and review the results.
    • Missing context: Send relevant surrounding content to the model.
    • Poor prioritization: Explicitly request that the most important information appear first.
    • Linked images: Include information about the image’s function and destination.

    These measures improve the information available to the model and the standards it is asked to follow.

    Human judgment remains necessary to determine whether the resulting description serves the page.

    Over-Reliance on AI

    Repeatedly receiving acceptable suggestions can make it tempting to stop inspecting them.

    That shift reduces editorial oversight and increases the chance of missing inaccuracies, irrelevant details, or important nuances. A tool may continue producing fluent descriptions while the quality of the actual user experience declines.

    The workflow therefore needs to address more than generation quality. It also needs to encourage people to remain involved after the novelty of the tool has worn off.

    Benevolent Friction

    Benevolent friction prioritizes a useful, responsible action by making it a deliberate part of the workflow.

    In this context, friction could mean pausing when an image lacks alternative text or requiring an author to choose among descriptions before applying one.

    The concern with fully automatic bulk assignment was that review could become an optional cleanup task. Once a tool had filled every field, authors might have little incentive to revisit the results.

    A small decision at the point of creation makes that responsibility more visible.

    Polly’s Squawk

    Polly’s “squawk” was a modal alert triggered when someone attempted to insert an image without alternative text.

    The message explained that blind and low-vision visitors would lack information about the image and offered two choices: add alternative text or insert the image anyway.

    Choosing to add text opened the interface with context and suggested descriptions. The workflow made the next step accessible from the warning itself.

    Other described reminders could appear when saving a post containing images without alt text, navigating away from such a post or page, or leaving the Media Library with missing descriptions.

    The intention was to make an omission visible and require a conscious decision. The later Q&A also addressed the need to recognize images intentionally marked as decorative.

    Polly’s Power of Annoyance

    Repeated reminders were presented as a way to encourage completion.

    An author might dismiss the first alert when inserting an image, then receive another reminder when saving or publishing. Each interruption offered a route to address the missing descriptions.

    The reminder was paired with assistance. The author did not have to leave the workflow and independently work out how to resolve the problem.

    However, the value of repeated prompts depends on whether they lead to thoughtful action. The Q&A returned to the risk that frequent interruptions could instead produce fatigue or automatic clicking.

    Polly Forces Oversight Without Even Trying

    Offering several suggestions introduces a review step into the generation process.

    Different options are intended to emphasize different aspects of the image. The author can compare those emphases with the reason the image was selected and choose the best fit.

    That comparison can be quicker than composing a description from scratch while still requiring a meaningful editorial decision.

    The limitation is that a required choice does not guarantee careful review. Someone can still select the first option without reading it. The interface supports oversight, but training and responsibility remain important.

    Benevolent Friction: Forcing a Choice

    The selection interface included an enlarged image preview above the generated options. This let the reviewer inspect the image while considering the wording.

    Each suggestion also had an editing option. An author could keep a useful starting point while correcting a detail, changing the emphasis, or simplifying the language.

    Human control included both selection and revision. The workflow did not require accepting an AI suggestion exactly as written.

    Overcoming the Challenges of AI Generation in a WordPress Plugin

    The final review brought the approach back to three connected practices.

    Give consistent instructions. Each generation request should carry the relevant writing expectations, including clarity, appropriate detail, and prioritization.

    Supply the image’s actual context. Nearby content and information about the image’s function can help the model produce suggestions suited to a particular placement.

    Keep editorial decisions with people. Prompts, alternatives, and editing controls should help authors review and improve the output.

    The larger principle was that AI can reduce the effort required to create useful alternative text while leaving responsibility with the person publishing it. Benevolent friction supports that relationship by making important decisions harder to overlook.

    Q&A

    Can images be marked as decorative, and will Polly stop prompting for alt text?

    Yes. Polly included a checkbox for intentionally marking an image as decorative.

    Purely decorative images, such as an image used only as a divider or background ornament, do not necessarily need a description. The decorative-image feature still needed refinement, and compatibility with WordPress’s own decorative-image control was not confirmed during the session.

    Where can people access Polly Alt or contribute?

    The session directed attendees to the WordPress plugin directory and the Polly Alt GitHub repository.

    Testing, bug reports, and pull requests were welcomed. Questions and feedback could also be sent to deneb@seamonsterstudios.com, including feedback from screen reader users about the approach to image descriptions.

    How would AI know whether an image is decorative?

    The content creator makes that decision. Polly’s AI generates descriptions; the author can mark an image as decorative.

    Mood-setting images can require judgment because they may contribute to the experience even when they do not add new factual information. Automatically describing every image can create unnecessary text, while automatically hiding images can omit something meaningful.

    Is there a technical character limit for alt text, and can long alt text crash a device?

    The 125-character target was a writing guideline, not a universal technical limit. No device-crash issue was identified in the discussion.

    The main concern was usability: a long description can be harder to listen to, revisit, or navigate. When an image needs extensive explanation, provide concise alt text and put the longer description in another accessible part of the page.

    How should complex images, scientific photographs, and graphs be described?

    Use concise alt text to identify the image and direct people to a fuller explanation or data equivalent.

    NASA’s images illustrated why subject expertise matters: a generic description of stars may miss the scientific information that makes the image important. For graphs, an accessible table can communicate data more effectively than a long sequence of values inside an alt attribute.

    Making detailed explanations available on the page also benefits sighted visitors. A description or accessible expandable section can serve people who understand the information better through text or tables.

    What should people consider when using AI for accessibility tasks beyond alt text?

    Give specific instructions, check the output, and explain what needs correction.

    Consider which accessibility requirements the model may overlook. After requesting changes, check again for regressions, because correcting one problem may introduce another. Human review remains necessary throughout the process.

    How do you prevent benevolent friction from becoming cognitive fatigue?

    Combine the tool with training and clear expectations about why review matters.

    Authors can still click the first option to dismiss a modal. The proposed response was to make review easier, explain its purpose, and retain it as an expected part of publishing. The time-saving example—reducing a task from roughly a minute to about 20 seconds—was illustrative, rather than a measured result.

    The goal is less effort per thoughtful decision. Repeated prompts alone cannot ensure that a person evaluates the descriptions.

    Can AI-generated alt text be optimized for people with cognitive disabilities?

    Simpler wording and reading-level guidance were discussed as possible future improvements.

    One suggestion was to include a reading-level target in the prompt, such as eighth grade or lower, or provide a setting that prioritizes simpler descriptions. This was a feature idea, not a confirmed existing capability, and no specific tool was identified as the answer during the session.

    Facebook0Tweet0LinkedIn0Shares0

    Filed Under: Recorded Webinars WordPress Accessibility Meetup

    About Equalize Digital

    Equalize Digital's team has specialized in WordPress accessibility for more than a decade. We offer accessibility audits, WordPress accessibility remediation, user testing, and build bespoke, accessibility-first websites. Our WordPress Accessibility Checker plugin is used by large and small businesses, nonprofits, higher ed, and government websites worldwide. Try it free today.

    Post navigation

    Accessibility Checker changelog livestream: Issues Explorer Rescans and Better Builder Support.Previous post: Changelog 017: Issues Explorer Rescans and Better Builder Support

    Easier, Faster Accessibility Testing

    Equalize Digital Accessibility Checker gives you real-time accessibility feedback in the WordPress editor. Learn accessibility and make fixes earlier in the dev and content creation process. Full-site accessibility scanning without the per page fees.

    Get Accessibility Checker

    Footer

    Equalize Digital Websites for Everyone

    Your WordPress accessibility team. Accessibility plugins, rapid audits, and consulting to help you make your website usable by people of all abilities.

    • Facebook
    • GitHub
    • LinkedIn
    • YouTube

    Company

    • About Equalize Digital
    • WordPress Accessibility Meetup
    • Accessibility Statement
    • Blog
    • Events
    • Contact Us

    Services

    • Accessibility Audits
    • User Testing
    • Remediation
    • Ongoing Monitoring
    • VPAT & ACR Preparation
    • Accessibility Training
    • For Agencies
    • Website Development

    Accessibility Checker

    • Features
    • Pricing
    • Documentation
    • How to Get Support
    • My Account
    • Affiliate Dashboard
    • Become an Affiliate

    © 2026 Equalize Digital · Privacy Policy · Service Terms · Software Terms · Data Terms

    International Association of Accessibility Professionals member

    Small Business Accessibility Playbook

    Learn how to make your website accessible.

    Free Ebook: The Small Business Accessibility Playbook for WordPress by Equalize Digital and WP Buffs.

    Get a copy of the free e-book via email.

    This field is for validation purposes and should be left unchanged.
    Name(Required)
    This field is hidden when viewing the form
    This field is hidden when viewing the form
    Privacy Policy(Required)
    This field is hidden when viewing the form