
Making accessibility easier and more effective, not just automated
October 1, 2026
It is better to overdo alt text than underdo it, ensuring an equivalent experience for all users.
VS
Decorative Images
Subject-only Alt
Minimal alt text risks excluding users by withholding vital information, whereas detailed descriptions empower greater accessibility and inclusion.
Screen reader users can skip once they have the gist.
Only works if alt text is written well and front-loaded with key information.
Humans using AI not to cede their responsibilities over to it, but to enhance their abilities to complete them in the best possible way.
One of AI’s greatest temptations for its users is for them to think that AI can take care of things without them, ie “Agentic AI”.
In the end, though, humans are the ultimately responsible parties, and they can’t abdicate the ultimate responsibility to the AI.
Ethical AI is only using AI in a context where this is fully understood.
(and how AI in a WordPress plugin can overcome them)
| Challenge | Solution |
|---|---|
| Poor training data | |
| Not prioritized/ front loaded | |
| No context | |
| Overreliance on AI |
AI alt text generation ability has historically only been as good as its training data (which was pretty poor). There’s a ton of lazy, sloppy, misguided alt out there (filenames, SEO keywords, unnecessary words, etc), meaning AI often replicates these exact mistakes when generating new descriptions.
A plugin can feed consistent instructions to the AI on how to avoid the pitfalls the training data may contain. By establishing structured boundaries, we ensure the generative process remains accurate, preventing the replication of lazy, sloppy, or misguided alt text patterns commonly found in historical datasets.
“Alt text should describe what sighted visitors see in the image. Aim for around 125 characters. Avoid starting with ‘image of’ or ‘photo of’.”
Clear instructions center the AI model on the task at hand:
AI-generated alt text may sometimes miss contextual subtleties important to users. Usually, the AI doesn’t have access to the surrounding page context, meaning it is forced to generate alternative descriptions entirely in a vacuum. Without reference to the surrounding paragraphs, template structures, or intent, even high-quality visual outputs fail to produce useful digital accessibility.
Easy to pull paragraph context from Gutenberg or Elementor when about to insert images, and for after-the-fact, it's possible to search the database to find where the image is used and its context.
At the point of insertion (Gutenberg, Elementor, Classic Editor) it’s easy for Polly to grab the context of the surrounding paragraphs and feed that along with the image and instructions to the model:
“At least the first variation should connect what's visible in the image and what the surrounding text is about.”
“If the image relates to the surrounding text, that connection belongs ONLY in the 'explanation' field.”
The AI highlights the specific focus of the option and its relationship to surrounding text so the human can evaluate appropriateness.
A challenge with alt text in WordPress generally is that the alt attribute of an image is singular whereas an image may be used in a variety of posts or pages, with a variety of different contexts, so should really have a variety of different alts. Most AI alt generators are unaware of context, so wouldn’t provide more than one alt, even if WordPress would allow them to.
Give AI the context of images from various pages, then have plugin rewrite default tags to replace global alt attributes with specific ones.
Polly writes the page-specific alt the user chooses directly into the page by rewriting the default alt:
$img_tag = $img_match[0];
$new_tag = preg_match( '/\salt="[^"]*"/', $img_tag )
? preg_replace( '/\salt="[^"]*"/', ' alt="' . esc_attr( $alt_text ) . '"', $img_tag, 1 )
: preg_replace( '/<img/', '<img alt="' . esc_attr( $alt_text ) . '"', $img_tag, 1 );
For images used on multiple pages, choose which specific instance or global setting you want to add alt text to.
Image-links need a special type of alt indicating where the images lead to rather than just visual appearance.
If a plugin can edit alt per instance, and detect its context, it can easily suggest and set a different appropriate alt for an image-link, writing it into the markup of the page directly.
At the point of insertion, notifications inform editors when a different kind of alt text is needed for links, and gives them choices to meet that need.
From the Media Library, after selecting the instance, instructions indicate what kind of link alt text is required.
(and how AI in a WordPress plugin facilitates them)
Over-reliance can lead to decreased quality
Just have to make doing the right thing (providing the oversight in the creation of alt text) easier than avoiding doing it. This is actually pretty easy at the point of insertion.
In Classic Editor, Elementor, or Gutenberg, inserting an image without alt text triggers a modal alert:
const modal = document.createElement('div');
modal.id = 'polly-enforcement-modal';
modal.className = 'polly-modal-alert';
modal.innerHTML = `<h3 id="polly-alert-heading">🦜 Just a moment, Captain!</h3>`;
You, the human, can choose to do the wrong thing, but not without thinking about it, and Polly’s there to help you do the right thing
Polly squawks at various key moments:
If you choose to not add alt, Polly squawks and squawks and squawks! Eventually, it’s just easier to add it…
Polly gives 3+ choices of alt, essentially requiring human oversight. Choices each have a different focus
“Generate exactly ${config.choiceCount} distinct alt text variations following these rules:
Write in plain language, present tense, active voice, include only what is visible — no interpretation or assumptions Each variation must foreground a DIFFERENT visible subject or element from the image as its opening focus. The thing named first in the alt text should differ across all variations.”
Modal presents a zoomed-in image that you can scroll to see everything without having to deal with bidirectional scrolling.
The AI gives 3 or more options so you, the human, exercise your human judgment to pick the one that fits, or you can edit one
AI alt text generation ability has historically only been as good as its training data (which was pretty poor)
There’s a ton of lazy, sloppy, misguided alt out there (filenames, SEO keywords, unnecessary words, etc)
Continually feed instructions to the AI to keep it honest
A plugin can feed consistent instructions to the AI on how to avoid the pitfalls the training data may contain
AI-generated alt text may sometimes miss contextual subtleties important to users.
Usually the AI doesn’t have access to the context, so it’s generating the alt in a vacuum
Give AI the context of the images (page, post, template, or block)
Easy to pull paragraph context from within Gutenberg or Elementor when about to insert the image, and possible to search the database to find where image is used and its context
Risk of over-reliance on AI could reduce human oversight and editorial judgment.
Over-dependence on AI may cause important nuances to be missed, diminishing accessibility quality and user experience.
Benevolent friction rather than prioritizing easy
Just have to make doing the right thing (providing the oversight in the creation of alt text) easier than avoiding doing it. This is actually pretty easy at the point of insertion.
Deneb Pulsipher
Captain Accessible at SeaMonster Studios
deneb@seamonsterstudios.com
Polly Alt Plugin: https://wordpress.org/plugins/polly-alt/
GitHub repo: github.com/SeamonsterDeneb/Polly-Alt/