11 min read
The Slopnami is an accessibility problem
Automation that saves the sender time usually spends someone else's. Disabled people have the least margin to absorb it.
As part of my series on AI, automation, and inclusive design, I want to acknowledge that many people, including disabled folks, are required or need to use AI in their work. I don't want to use this as a forum to discuss the ethics or legality of using generative AI, because I believe that discussion deserves its own space and isn't germane to this issue; it just compounds it. If the underlying issues in generative AI were solved, slopamation would still be an issue to solve.
I pause, my hand hovering over my laptop, as if opening it would release an agitated spirit from its vessel. I got a notification, but I need to steel myself before going into the breach of attention and intention that peppers my day: the low-effort outreach that becomes high-effort for me.
Before I can get to the one message I actually need to take action on, I have to wade through everything else. I'm no longer upset about the noise in my email or SMS inbox, but it's rapidly spreading everywhere. These days, I really only talk to my friends and loved ones on Signal or Discord. But the time cost is starting to wear on me as more tools present more interruptions or sorting tasks. (And unfortunately, automation and AI haven't provided successful solutions on my end, so fighting fire with fire is out.)
LinkedIn connection requests from financial consultants I've never met; seven in one week from the same firm with the same copy-paste template. A "quick question" message obviously built from details scraped from my profile and designed to sell to me. A comment on my last accessibility post that could have been left on anyone's: "thanks so much for writing this, it's important to consider all sides." Each one took its sender about a second of consideration, if that. Each one takes me longer than that, and I have to evaluate it just to find out it was nothing.
Then the stakes grow: there's the support chat that won't hand me to a person, the meeting summary that is longer than the meeting, and the article that takes six paragraphs to reach its one critical point. Then there's the website whose "accessibility" overlay is meant to help, but makes the page harder to use.
None of these is a disaster on its own. Together, they add up to a daily tax, and that tax compounds with other disability taxes.
I've worked in design for the last 16 years, and some of that work comes down to one question: who does the work, and what does it cost in time, effort, and money?
A well-made product can handle complexity so the person using it doesn't have to. AI slop runs that in reverse. It saves the sender time by automating the qualifying step, and spends the receiver's time instead. Disabled people have the least margin to absorb that cost, whether we pay it in screen reader listening time, energy, cognitive load, or the chance to reach a human who can actually help.
- The sender automates. Outreach, writing, images, captions, or code get generated instead of done.
- Volume goes up and quality goes down. More messages, longer text, more markup, fewer human checks.
- The receiver absorbs the difference. Someone has to read, triage, verify, correct, or route around it.
- Disabled receivers pay the most. Every item costs more time with assistive tech, more energy, and more cognitive load, and the barriers it creates fall on us first.
The taxonomy of slop and its cost
I sorted what I have observed into a few categories. Some are annoyances. Some decide whether a person gets a job, a benefit, or a straight answer.
Inbound noise
This is the one that started this post. Automated outreach that feels entitled to my attention for free.
| Type | General impact on users | Impact on disabled users |
|---|---|---|
| Outreach and connection spam | Inboxes fill with low-value messages, real contacts get missed, and people stop trusting cold outreach. | Every message has to be triaged. For screen reader users, that means listening to each one in full. Time and energy go to deleting junk before any real professional contact can happen. |
| Bot comments and fake engagement | Comment sections and feeds fill with filler, and real conversation gets harder to find. | Disabled creators and community spaces get buried in generic replies. Real peers get harder to find, and moderation of slop and hate becomes one more unpaid job. |
| AI robocalls, AI lead hunters, and voice spam | Unknown calls become noise that most people ignore or block. | People who rely on the phone for healthcare coordination often can't ignore unknown numbers, making AI robocalls and leads a particular vulnerability for disabled folks. AI tools that hunt leads often aren't correct, and I receive outreach about jobs I left years ago: I spend my time talking my way off the phone two minutes at a time, several times a week. |
Bloat
Length and density have a cost, and someone always pays. Generated text is cheap to produce and expensive to read.
| Type | General impact on users | Impact on disabled users |
|---|---|---|
| Padded, verbose writing | Readers spend longer finding the point and skim past what matters. | The point is buried under filler. That adds cognitive load for people with dyslexia, ADHD, brain fog, or fatigue, and listening time for anyone using assistive tech. |
| AI meeting notes and transcripts dumped unedited | Notes stop saving time, and decisions and action items get lost. | People who rely on notes because they can't follow live audio, or need to review at their own pace, get a longer and less accurate record. |
| SEO slop articles | Search results get worse, and answers take longer to find. | Paragraphs of preamble before one answer. That's costly with magnification or switch access. Usually these articles are padded with that verbose writing as well. |
Fake accessibility
This category makes my job harder, partially because it gives people a false sense of security while being opaque about what the tools can actually do. Anyone who has run an automated checker like axe knows it only catches part of the picture. AI "fixes" have the same shortcomings, and they get advertised as being able to do the whole job. Each of these lets an organization say the work is done without accountability or growth.
| Type | General impact on users | Impact on disabled users |
|---|---|---|
| Auto-generated alt text | Images look described in audits and code review, so nobody writes real descriptions. | "A person smiling" technically fills the attribute and conveys nothing. Screen reader users learn nothing about the image. |
| Unreviewed auto-captions | Anyone watching muted, watching a translated caption, or in a noisy place gets the wrong words. | Names, jargon, and accented speech get mangled, yet checked off as "captions provided." Deaf and hard of hearing viewers get errors. |
| Overlay widgets sold as "AI accessibility" | Organizations pay for compliance they don't get, keep their legal risk, and their understanding of accessibility stays low. | They often break screen readers and keyboard navigation while letting organizations claim compliance. My hot take? Any non-disabled person who works at an overlay company should have to use assistive tech at work so they can experience the UX of overlays. |
| AI-generated code and UI shipped untested with real accessibility tools and assistive tech users | Bugs and bloated markup ship faster than review can keep up. | Missing labels, div buttons, and broken focus get reproduced at scale, faster than anyone can audit them. |
Gatekeeping
With purchasing, most of the time you can switch to another product or provider. In hiring and benefits, there's rarely an alternative. When automation of any kind sits between a person and an opportunity, it becomes a barrier. I mentor upcoming accessibility practitioners and this is one area that concerns me most for my mentees going into the field.
| Type | General impact on users | Impact on disabled users |
|---|---|---|
| Chatbot-only customer service | Anyone with a problem the bot can't solve gets stuck in a loop. | Reaching a human means getting past a bot first. That's a barrier for people with speech, cognitive, or language differences, and the chat widgets are often inaccessible themselves. |
| AI hiring screens | Qualified candidates get filtered out, and recruiters drown in generated applications, so fewer get read closely. | Résumé rankers and video or voice analysis can penalize disability, atypical speech, eye contact, and résumé gaps. |
| Automated benefits and claims processing | Errors get decided at scale and take a long time to appeal. | The appeal burden lands on the people with the fewest resources and the most at stake. |
Bad information
Disabled people do a lot of research to get through systems that weren't built for us. Confident wrong answers make that work more dangerous.
| Type | General impact on users | Impact on disabled users |
|---|---|---|
| Hallucinated health, benefits, or legal information without confidence scores or explicit warnings | People act on confident wrong answers. | People navigating disability benefits, medications, or accommodations act on it, with real consequences. |
| Fake product reviews | Shoppers can't trust ratings. | Assistive tech is expensive and hard to trial. Synthetic reviews make it harder to know what actually works. |
| Summaries that replace the source | Nuance and errors get lost between the source and the reader. | People who rely on summaries for access get the distortions, with no easy way to check them. |
Representation and erasure
This is the one that makes me saddest about the AI slopnami. Even with strong art direction and examples, AI aims for the most common, broadly appealing, and non-disabled. When discussing disability and accessibility, the intersections and complexity are both the challenge in the space and also what generates energy for problem-solvers.
Disabled people have been the ones bringing that complexity to the table, in research sessions, in hallway conversations, and as the colleague people ask, "does this work for you?"
HBR recently described "relationslipping," where people turn to AI instead of their colleagues, lose track of what skills live on their team, and settle for AI's "good enough" guess. When that happens, lived experience is the first thing to go. AI hands back that good ol' average; disability becomes a rounding error.
| Type | General impact on users | Impact on disabled users |
|---|---|---|
| Idealized images that erase disability | Generated images flatten everyone toward a narrow default of bodies, faces, and ages. | Generators, editors that "enhance," and beauty filters remove wheelchairs, prosthetics, hearing aids, facial differences, and limb differences, even from real photos and prompts that ask for disability. Disabled people disappear from stock imagery, avatars, and marketing, and are told their actual body is an error to be fixed. As these tools learn from their own output, the erasure compounds. |
| Reductive stereotypes when disability is considered | Audiences see the same tired tropes repeated. | When models do depict disability, they tend to collapse it into one sad, isolated, wheelchair-shaped image. |
| Synthetic disabled personas and "inspiration" content | Audiences can't tell real voices from manufactured ones, and the resources that design, product, and engineering rely on are flattened to averages. | Disability gets depicted by people and systems with no lived experience, which crowds out disabled voices and recycles stereotypes. |
What the research says
The same pattern shows up in hiring tools, image generators, language models, and the accessibility industry itself.
| Area | Finding | Source |
|---|---|---|
| Hiring | GPT-4 consistently ranked résumés with disability-related honors lower than the identical résumé without them. A version instructed not to be ableist improved some results, but bias persisted, notably for autism and depression. | Glazko et al., University of Washington, FAccT 2024 |
| Image generation (erasure) | When prompts explicitly mentioned disability, many generated images showed no visible sign of disability at all. | Exploring Representations of Disability and Sexuality in AI-Generated Art, CSA conference |
| Image generation (stereotypes) | Across several text-to-image models, most images of disabled people showed them as old, sad, and using manual wheelchairs. | Tevissen, 2024 |
| Image generation (lived experience) | In eight focus groups with 25 disabled people, participants found models repeatedly produced reductive archetypes, including many wheelchair users shown sad, lonely, or in pain. | Mack et al., CHI 2024 |
| Language and moderation | Every public sentiment and toxicity model tested showed statistically significant bias against disability-related language. | Venkit, Srinath & Wilson, Penn State, 2023 |
| "AI accessibility" overlays | The FTC ordered accessiBe to pay $1 million over claims its AI widget could make any site WCAG compliant, and alleged the tool created barriers for disabled users. In WebAIM's 2021 survey of accessibility practitioners, 67% rated overlays not very or not at all effective, rising to 72% among respondents with disabilities. | FTC final order, April 2025; WebAIM Practitioners Survey #3 |
| Hiring law | The DOJ warned in 2022 that employers are responsible for ADA compliance of AI hiring tools, including vendor tools, and that screening out qualified disabled applicants can violate the ADA regardless of intent. | DOJ, Algorithms, AI, and Disability Discrimination in Hiring |
Inaccessibility, automated at scale
This number worries me most as a practitioner.
Every year, WebAIM tests the home pages of the top one million websites. After six years of slow improvement, the 2026 report went backwards. Errors rose, more pages failed, and pages got dramatically more complex.
WebAIM's summary named likely causes: heavier reliance on third-party frameworks and libraries, and automated or AI-assisted coding, or "vibe coding." It matches what I see. Generated code produces more markup faster, but if you didn't know about accessibility in the first place, it's hard to check.
| Measure | 2025 | 2026 |
|---|---|---|
| Average detected errors per page | 51 | 56.1 |
| Pages with detected WCAG failures | 94.8% | 95.9% |
| Average elements per page | 1,257 | 1,437 |
| Average ARIA attributes per page | 106 | 133 |
| Pages with low contrast text | 79.1% | 83.9% |
| Pages with missing form labels | 48.2% | 51% |
| Pages with empty buttons | 29.6% | 30.6% |
Source: The WebAIM Million, 2026 report
A few more numbers from the report stood out to me:
- Disabled users can expect to hit an error on 1 in every 26 home page elements.
- ARIA use has grown to more than six times its 2019 level. Pages using ARIA averaged 59.1 errors, compared with 42 on pages without it. WebAIM is careful to say ARIA doesn't necessarily cause the errors, since those pages are also more complex. To me, it's the same pattern as auto alt text and overlays: markup that looks like accessibility, without the understanding or work that makes it accessible.
- 96% of all detected errors fall into the same six basic categories that have topped the list for seven years. Low contrast, missing labels, empty links, and empty buttons all got worse this year. If AI repeats code and patterns already in use, it makes sense that errors get recycled too.
Slopnami relief
- If you're doing outreach, send fewer messages and make them personal. If it isn't worth your time to write, is it worth mine to read?
- If you're shipping AI-generated code, test it with a keyboard and a screen reader before it goes to testers who use assistive tech every day (and it should).
- If you're using AI for alt text, captions, images, or summaries, have a person (content designer, hopefully!) review it. Generated output is always a draft, not a deliverable.
- Be very wary of using AI in hiring, benefits, or support. Always keep a fast, accessible path to a human, and test the system with disabled people before it decides anything for or about them.
Automation isn't neutral. It decides whose time counts. Right now, too much of it is spending disabled users' time and energy, and adding to the disability tax.
Download the take-away for a quick reference!
Sources
- Glazko et al., Identifying and Improving Disability Bias in GPT-Based Resume Screening, FAccT 2024
- Mack et al., "They only care to show us the wheelchair": disability representation in text-to-image AI models, CHI 2024
- Tevissen, Disability Representations: Finding Biases in Automatic Image Generation, 2024
- Exploring Representations of Disability and Sexuality in AI-Generated Art, Canadian Sociological Association conference
- Venkit, Srinath & Wilson, Automated Ableism: An Exploration of Explicit Disability Biases in Sentiment and Toxicity Analysis Models, 2023
- Federal Trade Commission, FTC Approves Final Order Requiring accessiBe to Pay $1 Million, April 2025
- WebAIM, Survey of Web Accessibility Practitioners #3 results, 2021
- U.S. Department of Justice, Algorithms, Artificial Intelligence, and Disability Discrimination in Hiring, 2022
- WebAIM, The WebAIM Million: 2026 report
- Niederhoffer, Rapuano, Liebscher, Robichaux & Hancock, Team Dynamics Weaken When Employees Turn to AI Instead of Colleagues, Harvard Business Review, October 2026