The accessibility risk hiding in AI-generated content

Welcome and Today’s Agenda

Josh: Good afternoon, everyone. Thank you so much for joining us today. We are truly grateful to have you here. Yesterday was the 36th anniversary of the Americans with Disabilities Act, so this is a particularly meaningful time for me and for so many others. It reminds us how far we have come, where we are going, and why this is such an important moment to have a conversation about accessibility.

Accessibility is not the responsibility of one team. It touches content, design, engineering, legal, marketing, and really anyone who helps shape a digital experience. AI is now working its way into all of those areas. That creates tremendous opportunity, but it also creates accessibility risks that can be easy to miss. Today, we are going to make those risks easier to recognize and, more importantly, give you practical ways to respond.

Here is where we are headed. First, we will look at how AI has become core infrastructure for many businesses. Then we will connect that growth to accessibility and examine why speed can outpace review. We will also share the results of our own in-house testing to show what AI actually produces when it is asked to build or evaluate digital experiences. We will leave time for questions at the end, so please share them in the chat throughout the presentation. These conversations make the session stronger.

About the Research

Josh: Before we get into the findings, I want to be clear about where the data comes from. The research we are sharing today is accessiBe’s own research, conducted with Qualtrics to better understand how businesses are using generative AI and how leaders are thinking about accessibility risks. We have studied broader accessibility trends for years. This research goes deeper into generative AI because both the technology and the way teams use it are changing incredibly quickly.

Meet the Speakers

Josh: For those I have not met, I am Josh Basile, the Community Relations Manager at accessiBe. I have been here for more than five and a half years and love being part of the accessiBe family. I am also a disability rights advocate, a trial attorney, and a person who relies on assistive technology every day.

I use voice-dictation software, an on-screen keyboard, and a QuadStick that I control with my mouth to operate different devices and navigate the internet. This subject is meaningful to me professionally, but it is also deeply personal. Without accessible technology, I lose my independence. With it, I am able to participate in the world, work, and engage with businesses. I am thrilled to be joined by Sharon Uda, our Vice President of Research and Development. Sharon, I will let you introduce yourself.

Sharon: Thank you, Josh. Hi, everyone. I am Sharon Uda, and I have been leading the engineering team at accessiBe for a little over a year. Before that, I spent my career leading engineering organizations across companies of different sizes, from startups to larger enterprises, and across several industries. My previous two workplaces were focused on cybersecurity.

This is my first role in the accessibility space, and it has been an incredibly rewarding experience. It is not every day that you get to work on cutting-edge technology while also doing good and helping make the web more inclusive and accessible.

AI has become a major focus for us. We use it extensively across product design and engineering to accelerate our entire software-development lifecycle, and we are also building AI-powered products for our customers. We experience AI from both sides: as developers building software and as a company creating AI solutions. One thing we have learned is that while AI can generate code incredibly fast, it does not automatically generate accessible code. That is exactly what we are going to explore today.

AI as Core Infrastructure

Josh: AI is quickly becoming core infrastructure for e-commerce businesses. Seventy-eight percent of the brands we surveyed now use AI to generate at least one quarter of their content. That is no longer a small experiment. It means AI is actively shaping the customer experience.

For someone like me who navigates technology with assistive tools, the quality of that output really matters. It can determine whether I can experience something or not. AI may write a product description, generate alt text, power a chatbot, personalize a page, or help build a checkout flow. We are using it throughout the digital experience.

When it works, AI can make information easier to create and find. When accessibility is missing, it can create barriers. That is why it is so important to get this right and to put the right steps in place. The good news is that accessibility, search optimization, and answer-engine optimization often reinforce one another. Clear structure and meaningful descriptions help both people and technology understand content. The question is not whether businesses can use AI. The question is how we use it while protecting accessibility.

The Hidden Accessibility Risk

Josh: The encouraging part is that teams already recognize the concern. More than 82% of the leaders we surveyed believe AI may be creating unseen barriers for shoppers with disabilities. ‘Unseen’ is the key word.

A page may look polished while a keyboard user cannot reach a button. An image may have alt text that sounds complete but describes the wrong product. A chatbot may work visually but fail to announce new messages to a screen reader. These problems may be invisible to the person publishing the content and immediately obvious to someone like me who is trying to use it.

Awareness is a meaningful first step. The next step is building a process that catches those barriers before the customer does.

Legal Exposure and Shared Responsibility

Josh: Let me put this into context from my perspective as a lawyer. E-commerce is one of the industries most frequently targeted by digital-accessibility legal action, and that has been true for many years. AI did not create that exposure. What AI has changed is the speed and scale at which inaccessible content can be produced.

Among surveyed brands that faced accessibility legal action, AI-generated content was the most frequently cited factor. Ninety-six percent of those matters ended in a site fix, a settlement, or both. The practical lesson is not to wait for a demand letter before beginning the work. Start early, document the steps you take, and be able to show your testing, remediation, and ongoing improvement. As a trial attorney, I can tell you that evidence of those efforts can go a long way.

This work cannot sit with one person. Developers, e-commerce leaders, content teams, agencies, and legal teams all have a role in the process.

Audience Poll: Pressure to Publish

Josh: We would love to hear from you. How pressured do you feel to publish AI-generated content quickly? Please answer the poll, and if you have seen that pressure affect your review process, share an example in the chat.

Here are the results: 13% reported no pressure at all, 29% reported slight pressure, 23% reported moderate pressure, 29% reported significant pressure, and 6% reported extreme pressure. The results are distributed across the spectrum, but AI-generated content is here, and it is coming at us quickly. Sapir and Sharon, is anything jumping out at you from those results?

Sapir: I am excited for you to share how these results compare with what 300 e-commerce leaders told us in the survey.

Josh: The results reflect what we found in the research. Sixty-seven percent of respondents feel at least moderate pressure to publish AI-generated content quickly. This is not an anti-AI message. We are excited about AI and use it extensively ourselves. The challenge is that AI increases production speed while the time and process available for accessibility review may not increase with it.

Sharon, your team works with AI while continuing to make accessibility a priority. What does that balance look like in practice?

Balancing Speed and Quality

Sharon: That is exactly what we are seeing in engineering. AI has fundamentally changed expectations around delivery speed. We all see stories about products or prototypes that once took months to build and are now delivered in a few weeks. AI is raising the bar for how quickly teams are expected to move, and I have those expectations for my teams as well.

At the same time, as an engineering leader, my job is to make sure we do not confuse speed with quality. We absolutely embrace AI. We use it extensively across engineering, design, and product, and we are building AI-powered products ourselves. But faster delivery does not remove our responsibility to ship secure, reliable, and accessible experiences.

Accessibility is an area where it is easy to assume that AI will handle everything. In reality, it often does not. That is why accessibility needs to be intentionally built into the development process rather than inspected only at the end.

Josh: We can take the extra steps, review the output, and put processes in place. That makes all the difference. If you do not build that mindset and rhythm, relying only on AI becomes a form of cutting corners.

Sharon: It is the same as security or any other code review. You need to verify that it has been handled.

Josh: The more often you do that, the faster you become. You start speaking the language, recognizing where barriers are likely to appear, and getting good at preventing them. Then you are not simply serving customers; you are welcoming all customers. No business wakes up wanting to block potential customers. Accessibility is smart business.

The Review Gap

Josh: The tension appears clearly in the data. Eight in ten brands publish at least some AI-generated content without review. To be fair, most teams cannot manually review every word, image, and component before it goes live. The goal is not to create an impossible standard.

However, publishing with little or no review creates a predictable gap. Our earlier research found that accessibility legal exposure was already widespread before generative AI became embedded into everyday workflows. AI did not create the underlying problem. It added speed and scale. The answer is a review process designed for that scale, using the right combination of technology and human expertise.

Test One: Can Better Prompting Produce Accessible Code?

Josh: What happens when we test the output ourselves? Sharon, I will hand this over to you.

Sharon: We wanted to answer the question I mentioned earlier: can AI handle accessibility end-to-end so that a team can simply forget about it? To investigate that, we conducted two tests.

The first question was practical. If teams are already using AI to build websites, can they improve the outcome simply by telling the model to pay attention to accessibility? We asked three leading AI models to generate the same simple website. The only thing we changed was the prompt.

We started with a standard prompt that essentially said, ‘Create this website and make it accessible.’ Then we added explicit and detailed WCAG guidance provided by our accessibility experts. Finally, we used a dedicated accessibility skill. To keep the evaluation unbiased, every generated website was reviewed manually by one of our accessibility experts rather than only by an automated tool.

The trend was clear. Better instructions consistently produced more accessible code. Every model improved when we gave it more accessibility context. But the improvement had a ceiling. No matter how much guidance we provided, every model still generated websites with significant accessibility issues. None came close to producing a fully accessible result. Prompting is a lever, but it is not a complete solution.

Test Two: Can AI Reliably Audit a Production Website?

Sharon: The first test looked at AI as a builder, but that is only half of the story. Many organizations already have production websites. The next question is whether AI can reliably audit an existing site.

For the second test, we took a real, more complex production website and asked leading AI models to identify its accessibility issues. We then compared their findings with deterministic accessibility engines: our own AccessFlow engine and axe-core, an industry-standard scanner included as an additional reference point. All findings were again manually reviewed and validated by accessibility experts.

We identified three patterns. The first was coverage. AI found only a fraction of the accessibility issues identified by the deterministic engines. Even the strongest prompt left a significant gap.

The second pattern was consistency. We ran the same test against the same website under the same conditions ten times. Each AI run produced a different set of findings. By contrast, the deterministic engines produced identical results every time. If you audit the same site twice and receive two different answers, you do not simply have inconsistency; you have a reliability problem. You also cannot measure improvement over time.

The third pattern was accuracy. AI generated significantly more false positives. That means engineering teams could spend valuable time investigating issues that are not actually present.

Taken together, the conclusion is clear. AI alone is not a reliable accessibility auditor for production websites. It misses too many real issues, produces inconsistent results, and creates unnecessary work through false positives.

What the Testing Revealed

Sharon: Let me summarize the findings in five points. First, baseline accessibility was low. Every model we tested produced real, unresolved issues when it received no accessibility guidance.

Second, prompting helps, but only to a point. A better prompt reduces the number of issues, but it does not bring the number to zero.

Third, AI on its own remains unreliable. It gives different answers when the same test is run twice, and close to half of what it flags may not be a real issue.

Fourth, deterministic engines found roughly ten times more issues than AI on the same site.

Fifth, complexity increases the challenge. We did not test this directly in the same experiment, but the websites used were relatively simple. As the complexity of the experience increases, the challenge becomes greater.

The takeaway is not ‘do not use AI.’ It is that AI needs a reliable layer underneath it: a process that reviews the output, catches what AI misses, and produces the same answer every time you check.

A Layered Accessibility Strategy

Josh: Thank you, Sharon. The research makes one thing clear: there is no single accessibility solution for every organization. The right approach depends on your technology, your team, and where you are in your accessibility journey.

That is why the accessiBe platform supports multiple paths. Runtime remediation can address high-volume, real-time needs at scale. Developer tools can help teams identify and fix issues at the source. Expert-led remediation supports complex experiences where human judgment and hands-on work are essential.

The goal across all three is the same: an accessibility program that is measurable, ongoing, and defensible, and an experience that works better for real people.

Audience Q&A

Josh: We have reached the question portion. Sapir, do we have questions from the chat?

Sapir: I sent you a question earlier. Can you check that one?

Josh: Absolutely. The first question is: what does a proper accessibility review actually look like? For me, the answer is that it should be layered. Automated scanning can reveal many technical issues, but it cannot catch everything. Expert review and user testing bring the process to the next level. It is also important to distinguish between something being technically accessible and being genuinely usable for visitors. Both are part of the accessibility journey.

Another question asks: if no model gets to zero issues, is better prompting still worth it? Yes, absolutely. Prompting will not get you to zero, but our research showed that it can reduce issues from 74 to 35, a 53% reduction. That is meaningful. Better prompting is a strong first filter, but it is not a substitute for testing and verification by your team.

Sapir: We have another question from Alexandra for Sharon. Aside from accessiBe, what tools are available to help make AI-generated code more accessible? What should teams look for, especially in vibe-coding tools?

Sharon: We are seeing accessibility skills appear in the market. We used one of the most popular skills in our test. We are also seeing platforms such as Lovable, Bolt, and Replit mention accessibility capabilities.

We conducted a separate, smaller test specifically with Lovable, and the outcome was similar. Even in a small test, the result still contained more than ten accessibility issues, four of which were critical. Perhaps this will change in the future, but today AI by itself does not provide a bulletproof solution. That makes sense when you consider that AI is trained on code from the existing web, most of which is inaccessible.

Sapir: That is a powerful point. Only about 4% of the internet is accessible, and AI is being trained on the rest. At the same time, organizations are under pressure to add more AI to their workflows. We will see how those forces balance out in the coming years.

Josh, Brett asks whether certain industries or business sizes are receiving more intense accessibility scrutiny, or whether everyone is feeling the pressure.

Josh: We are seeing it across the board. E-commerce receives a great deal of scrutiny because substantial amounts of money are exchanged and the industry can become an obvious target. But we also see legal pressure in financial services, healthcare, education, and manufacturing.

Any organization that is not bringing accessibility into its framework, documenting its efforts, and showing the steps it is taking may face exposure. The pressure is broad, but companies that take the necessary steps have been successful in responding and continuing their accessibility journey.

Sapir: We are also seeing preliminary data suggesting that businesses that address accessibility earlier improve both SEO and AEO, or answer-engine optimization, because those systems rely on many of the same accessibility best practices. That can indirectly affect revenue and website drop-off rates. There are important business benefits as well as risk considerations.

Josh: That makes sense. When content is accessible to a person with a disability, search crawlers and AI systems can often understand the site more effectively as well. If information is blank or blocked for a person with a disability, those systems may also have difficulty understanding what the site is about.

The disability community is also extraordinarily brand loyal. When we are treated well and have a good experience, we come back as repeat customers and tell friends and family. I am genuinely grateful when I can move from the beginning of a website journey through checkout without being blocked or needing to wait for a family member to help. Every business wants customers to have that kind of experience.

Sapir: We have one more question from Miguel. Is the widget enough for runtime protection? The question notes that monthly reports can feel reactive rather than preventative, and this applies to more than AI-generated content.

The widget is a valuable first step in an accessibility journey. It provides runtime support so users can make adjustments that improve usability, while helping mitigate compliance risk. When teams are moving quickly with AI, however, they also need to document their efforts and treat accessibility as a shared responsibility across marketing, content, product development, and product information.

The widget is a strong step in the right direction, but we recommend a holistic solution. We also offer manual accessibility services, including PDF remediation.

Josh and Sharon, do you have anything to add about runtime protection and the difference between reactive and proactive accessibility?

Josh: There will never be one switch that fixes everything. What organizations can do is be proactive and approach accessibility from several angles, both for the business and for the end-user experience.

The more you do this work, the easier and faster it becomes. It becomes part of how you conduct business. Search engines and AI systems can understand and communicate your story more effectively, and there are benefits across the board.

What I particularly value about the widget is the usability component. Accessibility is essential, but usability takes the experience to another level. Depending on whether I am using a computer, phone, tablet, or another device, and whether I am shopping, learning, or researching, the widget helps create a customized experience for that particular website. I love being able to have a better experience on a site.

Sharon: I think you phrased it perfectly.

Closing Remarks

Sapir: Thank you, everyone, for joining today. I have shared several resources in the chat, including information about AEO and the generative-AI research released today. Thank you for being part of the research launch. We also have industry reports covering government accessibility, higher education, healthcare, and e-commerce.

Please follow up with us if you have additional questions, and look out for the recording of this meeting. We are grateful for your time and your thoughtful questions, and we look forward to seeing you again.

Josh: Thank you, everyone, for being here, especially so close to the anniversary of the ADA. We are honored to be part of your accessibility journey. You do not have to do this alone. Sharon, thank you for your wisdom and for everything you do.

Sharon: Thank you, Josh, and thank you, everyone.