When AI Writes the Web, Millions of Users Are Left Out

Written by: A. Elizabeth O.

Ai is everywhere. In digital marketing I am seeing more "AI generated content than human written content. Trust me I get it. There is an appeal to create something fast and cheap. However, in the rush to automate content creation, businesses and creators are leaning heavily on Large Language Models (LLMs) like ChatGPT, Claude, and Gemini. While these tools, when used correctly, are a great help to your marketing team, they are just not good enough to be used for digitally accessible content.

There is a significant, often overlooked flaw with the AI generated content. While AI is excellent at mimicking human speech, it is surprisingly poor at understanding the humans it is generating content for. After all, humans have words we tend to use that do not sound like they were written for your resume. There is also the issue of emojis and em dashes not being screen reader friendly and more.

Digital accessibility is the practice of ensuring digital content is usable by people with a variety of disabilities. Some people are visually impaired, others with hearing loss, and there are also those with mobility issues. Each of these factors can change the way that your potential clients and customers use your digital information.

In 2026 that means that relies on structure, context, and semantic meaning. AI, by its nature, relies on statistical probability. This fundamental difference creates a gap where AI-written content frequently fails to meet Web Content Accessibility Guidelines (WCAG), leaving millions of users behind.

Here is why AI-written content is rarely accessible by default, and why human intervention remains non-negotiable.

1. The "Wall of Text" Problem & Semantic Structure

Screen readers (software used by blind or low-vision users to read text aloud) rely heavily on semantic HTML. This means using proper heading tags (<h1>, <h2>, <h3>) to create a navigational hierarchy.

The AI Flaw: When you ask an AI to write a blog post, it often outputs a flat stream of text. Even when it uses bolding or bullet points visually, it frequently fails to tag them correctly in the underlying code unless explicitly instructed to generate HTML.

Visual vs. Structural:

AI might give you a "Heading" that is just bolded text (<b>Title</b>) rather than a structural heading (<h2>Title</h2>). To a screen reader user, bold text is just text; it cannot be used to skip sections or understand the document outline. Even within LinkedIn you are able to create the proper heading chronicle under the "style" section although they are not called out by <h2> and <h3> headers they are called headings and subheadings, (I have used both in this article.)

Run-on Paragraphs:

AI models tend to be verbose. They generate "walls of text" that are cognitively demanding for users with reading disabilities like dyslexia or ADHD.

2. The Alt-Text Hallucination

One of the most critical aspects of digital accessibility is "Alt Text"—descriptions of images for those who cannot see them. Accessibility tools use alternative text on images to explain the purpose of the image on the content.

The AI Flaw: While multimodal AI models (which can "see" images) are improving, they still struggle with contextual relevance.

As an example here is a screenshot from a post on LinkedIn from the IAAP. The post advertised a document accessibility course. On the left you see the image that was attached to the post. On the right you can see that it included great descriptive alt tag that reads:

"White backdrop with a decorative light blue circle behind a person in a wheelchair typing on a laptop. Text reads, Document accessibility course. Sign in or Register now. List below with text reads, This course includes - general accessibility, -Microsoft word, -accessible forms and more. top right is IAAP vertical logo."

Right click and open the image in a new window to view the embedded alt text.

This is a good example of an alt text for an image that includes important information found in text overlayed on the image. Without that context, users who have visual impairments may miss important details about the course being offered on document accessibility.

The AI Flaw: While multimodal AI models (which can "see" images) are improving, they still struggle with contextual relevance.

Over-description:

An AI might describe the photo in the example above as "person in wheelchair with laptop." It may skip over the text in the image entirely because the text itself is embedded in the file, it does not make it readable for computer generated alternative text.

The Accessibility Reality:

If the image is decorative, the alt text should be null (empty). If the image is a chart, the alt text needs to summarize the data, not the colors of the bars. AI lacks the judgment to know why the image is there, leading to cluttered, noisy audio experiences for screen reader users.

3. Ambiguous Link Text

For users navigating via keyboard or screen reader, links need to make sense out of context. A user might pull up a list of all links on a page to find what they need. This is also true of many websites. Generic buttons for different links often lead to confusion for those who are using screen readers.

The AI Flaw: AI loves generic phrases. It frequently generates calls-to-action (CTAs) like:

"Click here"
"Read more"
"Learn more"

If a screen reader user hears a list of 10 links that all say "Click here," they have no idea where those links go. UX designers today are learning to eliminate buttons and ambiguous language links in web design, but when writing for other applications like blogs, social media posts and others, the same concept applies.

You want your links to have more descriptive text so that the end user knows what will happen when they click the link. An example of well designed accessible content requires descriptive links like "Read our guide on SEO strategies" or "Contact our support team." Unfortunately, AI generated content rarely defaults to this level of specificity without strict prompting.

4. Complexity & Cognitive Load (Plain Language)

WCAG guidelines suggest writing at a lower secondary education level to ensure content is understandable for people with cognitive disabilities or those reading in a second language.

The AI Flaw: LLMs are trained on vast datasets of academic papers, corporate reports, and literature. Consequently, they often default to a "corporate speak" style that is unnecessarily complex.

Jargon Overload:

AI often uses words like "leverage," "synergize," and "utilize" when "use," "work together," and "use" would suffice. This kind of language comes from the LLM, and honestly when you read it, it just sounds like computer jargon. It fails to capture the human element of communication. There is just a natural pattern of speech that the AI is not good enough to capture yet.

Passive Voice:

AI frequently writes in the passive voice ("Mistakes were made"), which is harder to process cognitively than active voice ("We made mistakes"). There are tools you can use to check your passive voice including Yoast SEO designed to analyze your website text, and the Hemmingway Editor for other written applications.

5. Emoji Overuse & Screen Readers

AI models, particularly when asked to write social media captions or "engaging" intros, love emojis.

The AI Flaw: To a sighted user, three rocket ships imply excitement. To a screen reader user, the device reads the Unicode description for every single icon.

The Experience: "Rocket. Rocket. Rocket. Sparkles. Fire. Here is the new update."
This interrupts the flow of information and can be incredibly annoying. AI does not understand the auditory "cost" of the visual flair it adds.

Emojis can really take the fun out of consuming social content written by AI. It also makes it very identifiable to readers of all abilities as AI generated. Today, when we see content that is littered with emojis we often skip it entirely. Did the person who approved the post even read it? Chances are, they didn't. They relied on the AI and did not audit the content it produced.

6. Color Contrast & Data Visualization

When AI is used to generate code for charts or suggest color palettes, it operates on mathematical color theory, not optical perception. Those who are colorblind might not be able to see your graphics with words if the colors blend together for them. Or those with sight impairments. I used examples of this in an article in this newsletter about your logo and color contrast requirements, but the same goes for AI generated images. You can read the article about color contrast HERE.

The AI Flaw: AI might suggest a color palette that looks "harmonious" but fails contrast ratio tests (WCAG 2.1 requires a contrast ratio of at least 4.5:1 for normal text).

The Answer is to Use AI as a Tool, Not the Architect

AI is not inherently anti-accessibility; it is simply does not understand how to prioritize it for digital content use. That means that the person using AI to prompt for the desired content, need to know how to ask for the right parameters for digital accessibility. That is why I wanted to address this on my Ink Blot Newsletter.

You have to train your AI to develop content in order to generate better outcomes. Even with the training, you still need a human audit with a quick checklist. If you are using AI tools to generate content it is important to audit it before you push the "publish" or "schedule" button. Here are my three basic "audits" for humans to use to help with digital accessibility.

Prompt for Structure: Explicitly ask the AI to "Use semantic H2 and H3 tags" or "Write in plain language at an 8th-grade reading level."
Audit the Output: Never copy-paste directly. Check link text. Ensure headings are real headings.
Contextualize Alt Text: Write your own alt text based on the purpose of the image, rather than relying on auto-generation. This also give you an opportunity to include keywords related to the content you are creating.

Accessibility is an act of empathy with intentional design to include everyone. AI can simulate intelligence, but it cannot simulate empathy or natively create digital assets that meet ADA requirements. Until it can, the responsibility for an inclusive web rests firmly with the human hitting "Publish."