Artificial Intelligence and Accessibility: Unlocking Potential, Knowing the Limits and Risks

At the end of July 2026, I will be running a workshop at the Dresden University of Applied Sciences exploring the question of what contribution artificial intelligence can make to digital accessibility and inclusion – and where its current limitations still lie. I have already engaged extensively with this topic over the past few years, yet I felt it was nonetheless appropriate to conduct a thorough review of the current literature in preparation for the event. It seemed like the perfect opportunity to share those insights with you here, and to develop some practical recommendations further down in the article – particularly if you are also facing the question of how to approach the topic of accessibility within your own organisation. I hope you enjoy the read! And if you are interested in a workshop on this topic for your team, feel free to get in touch. ;-)

Since ChatGPT first burst onto the scene at the end of 2022, artificial intelligence has been on virtually everyone’s lips. What followed was – and continues to be – a wave of hype that has extended far beyond the specialist community: economic disruption caused by shortages of AI chips, billion-pound investments in data centres, and a broader societal debate about which professions might be displaced by automation. Excitement and anxiety have gone hand in hand, and continue to do so.

It is only natural, then, that every discipline and every industry asks itself: what does AI mean to us – and what can it do for us? Can it solve problems that were previously difficult to address? Can it accelerate processes or make them more cost-effective? These are questions being asked by professionals in medicine, education, and law – and, of course, in the field of digital accessibility as well.

The Starting Point: Digital Accessibility and Its Challenges

Those working in the field of digital accessibility are usually all too familiar with its central challenges. First and foremost, there is a lack of awareness: many of those who develop digital systems or create content have yet to develop a sufficient understanding of the barriers they inadvertently introduce. A basic grasp of the subject may be present, but it is rarely enough to reliably identify and avoid accessibility issues. Accessibility is frequently perceived as a cost factor rather than – as it ought to be – a mark of quality. Unsurprisingly, this leads to low intrinsic motivation to implement it in any meaningful way. The fact that legislators have now stepped in and made accessibility a legal requirement for many digital offerings is an important step forward, yet it does not resolve the underlying dilemma: obligation without understanding tends to breed resistance rather than foster genuine commitment.

On top of this come very real practical hurdles. Identifying barriers within digital systems is a demanding undertaking: it requires expertise, time, and money. Automated testing tools can help, but in practice they only surface a portion of the barriers that actually exist, and they are rarely capable of assessing the quality or severity of a given problem. Whether a missing alternative text truly constitutes a barrier, or whether a particular structure is problematic for screen reader users, cannot yet be reliably determined by automated means alone. Genuine accessibility therefore remains dependent on human expertise and the direct involvement of those affected – both of which are scarce and costly resources, not to mention time-consuming.

This naturally gives rise to the question: can artificial intelligence help to reduce this shortfall? Not necessarily by replacing specialists entirely, but by supporting them – and, ultimately, by making the topic of accessibility more approachable and appealing.

A Look at the Literature: The Current State of Research on AI and Accessibility

The academic community has, of course, already taken up the subject: anyone searching today for research literature on artificial intelligence in the context of digital accessibility will find a considerably broad and growing body of publications. That in itself is an encouraging sign – and, in a sense, quite remarkable when one considers how young the field still is and how lengthy publication processes can sometimes be. The sources I draw upon in this article are, without exception, from the past three years. The most recent were published only last month.

Particularly illuminating are what are known as Systematic Literature Reviews (SLRs), which provide a structured, comprehensive overview of a given research field. Such reviews consolidate individual findings, identify patterns, and reveal where research is concentrated and where gaps remain. For the purposes of this article, the following questions are therefore of particular relevance:

• In which areas can AI already make a demonstrable contribution to accessibility?
• Where does it not yet function reliably?
• And which topics dominate current research to such a degree that they may point to especially high expectations or significant potential?

SLRs are, by their very nature, only produced after a subject has been meaningfully researched for some time – since there must first be a sufficient body of findings worth reporting. It is therefore all the more pleasing that I was already able to identify two such works that offer us a useful overview. Both studies are explicitly dedicated to the intersection of AI and digital accessibility. Chemnad and Othman [4] analysed 43 peer-reviewed articles published between 2018 and 2023, providing one of the first comprehensive overviews of the field. Campoverde-Molina and Luján-Mora [2] conducted a mapping study encompassing 53 studies up to mid-2025. The publication figures alone speak volumes: whilst only isolated studies appeared in the years up to 2022, as many as 26 papers were published in 2024 alone, with a further increase already becoming apparent in 2025. The key insight: academic interest in the topic of AI and accessibility has multiplied in a remarkably short space of time.

What Can AI Already Do?

Both review studies demonstrate that AI is already making a demonstrable contribution to accessibility across several areas. Particularly well-researched and well-tested are applications centred on automatic image description: systems that use computer vision and large language models to generate alternative texts for visual content account for nearly 30% of all studies analysed in the mapping study by Campoverde-Molina and Luján-Mora [2]. In the area of accessibility testing, AI-assisted tools are also showing considerable promise: language models such as ChatGPT can analyse HTML code for accessibility issues and provide suggestions for remediation, which can offer a meaningful improvement over purely manual review processes. Equally promising applications are emerging in the automatic captioning of videos, speech recognition, and the development of assistive technologies for blind and visually impaired individuals. Ferebee [5] describes in her narrative review a number of illustrative examples, including OrCam MyEye (Link: https://www.orcam.com/en-ie/orcam-myeye) – an AI-powered wearable camera that describes the surrounding environment in real time through audio output – as well as Microsoft’s Seeing AI app (Link: https://www.microsoft.com/en-us/garage/wall-of-fame/seeing-ai/), which makes similar functionality accessible on a smartphone.

Where Does Further Progress Still Need to Be Made?

At the same time, clear limitations are also emerging. Campoverde-Molina and Luján-Mora [2], for instance, note that whilst AI systems can detect whether an alternative text is present, they cannot determine whether it is actually appropriate in terms of content. Current models continue to fall short, particularly when dealing with complex graphics, charts, or data visualisations. A similar picture emerges with so-called overlay tools, which promise to automatically “fix” websites to make them accessible: Reins and Scheer [12], along with professional associations, are sharply critical of these tools, as in practice they frequently interfere with the use of assistive technologies such as screen readers rather than facilitating it – a rather paradoxical outcome. Automatic translation into Plain Language or German Sign Language also falls short, at present, of the quality achievable by human specialist translators. Fisseler [6] further highlights a structural problem: AI systems deployed in educational settings – for instance, to assess learning progress or analyse behaviour – can inadvertently disadvantage learners with disabilities if their atypical usage patterns, such as those arising from the use of assistive technologies, are flagged as anomalous by the underlying algorithms.

Which Topics Dominate?

A striking pattern in both review studies is the strong overrepresentation of applications designed for people with visual impairments. Chemnad and Othman [4] identify this as a significant weakness: although disability encompasses a broad spectrum – ranging from motor and cognitive impairments to hearing and speech difficulties and autism spectrum conditions – the vast majority of research remains focused on visual accessibility. Cognitive disabilities are particularly underrepresented in AI-based solutions, despite accounting for a considerable proportion of those affected. Furthermore, Campoverde-Molina and Luján-Mora [2] demonstrate that across the studies analysed, only 47 of the 86 success criteria defined in WCAG 2.2 are addressed at all. More than half of the normative requirements therefore remain largely overlooked within the current research focus.

A further recurring theme is the problem of systemic bias in AI systems: AI models are predominantly trained on data in which people with disabilities are underrepresented. Fisseler [6] references the widely cited Gender Shades study, which demonstrated that commercial facial recognition systems misclassified dark-skinned women at an error rate of up to 35%, whilst the error rate for light-skinned men was below 1%. This phenomenon carries over into accessibility applications. Put simply: those who are absent from the training data are served less well by the system.

Taken together, the literature paints a nuanced and instructive picture: AI holds considerable potential, already visible in practical applications, yet the research landscape remains comparatively young, somewhat uneven in its thematic coverage, and unresolved on key questions of quality. This in itself offers a useful indication of where development could – and arguably should – be heading.

Using AI for Accessibility Testing

One area that features with notable frequency in the literature is the use of AI to support accessibility testing. This comes as little surprise: it is precisely here that the strengths of generative AI meet the existing limitations of conventional testing tools.

Automated testing tools have been a staple of the accessibility testing toolkit for years. They do a solid job of identifying a broad range of potential issues, and they do so quickly, delivering reproducible results whilst reducing the burden of manual effort. Their current limitations lie where quality judgement is required – that is, where the actual impact of a barrier needs to be assessed. The classic example, once again: an automated tool can reliably determine whether an alternative text is present for a given image. What it cannot assess, however, is whether that alternative text actually captures the essence of the image – whether it is informative, precise, and genuinely helpful within the context of the page. That final evaluation remains a manual step.

Much the same applies to structural issues on web pages: when a tool flags a particular arrangement of elements as a potential barrier, it often remains unclear whether this constitutes a genuine accessibility problem, or whether the chosen design might nonetheless be fit for purpose in that specific context. What is consequently missing is an assessment of severity – in terms of both persistence and impact: how frequently does the problem occur? To what extent does it prevent users from achieving their goal? Automated tools currently provide no reliable answers to these questions.

It is in precisely these gaps that AI, when applied thoughtfully, can offer real added value. A multimodal language model capable of processing both image and text has at least the potential to compare an existing alternative text against the actual content of the image and provide an assessment of whether the description does justice to it. It can take context into account, sometimes recognise nuance, and formulate actionable recommendations – capabilities that conventional testing tools have so far lacked entirely.

One significant caveat remains, however: AI does not replace human expertise either. User studies, expert review, and the direct involvement of people with disabilities continue to be indispensable for assessing the genuine accessibility of a system. In this context, AI is a powerful aid – one that can accelerate the testing process, provide richer information about the quality of barriers, and enable testers to work with greater precision and focus.

Using AI to Develop Accessible Content

Identifying barriers is the first step – remedying them is where the real challenge lies. Here, too, the question arises as to whether and how AI can provide support: for instance, by offering concrete suggestions for correction, or by preparing content in such a way that barriers do not arise in the first place.

At first glance, this sounds promising. A language model that understands HTML code can, in theory, directly address identified barriers – adding missing ARIA attributes, correcting a flawed heading hierarchy, or adjusting inaccessible form elements. For (at least static) web content, this is a natural use case, given that the underlying code is well-structured and readily interpretable by language models.

Yet, even with the next media type, things can become considerably more complex. PDF documents, for example, consist of multiple layers: a semantic content layer comprising tags and reading order, and a visual layer that determines what the user actually sees on screen. Accessibility here means explicitly more than a screen reader being able to read the text correctly – it also requires that aspects such as font size, typeface, and contrast ratios meet accessibility requirements. This multidimensionality makes automated correction significantly more demanding. The semantic content layer is broadly comparable in complexity to that of static websites, but the associated visual component raises the level of difficulty considerably.

A further step up in complexity are interactive applications: apps, software systems, and publicly accessible terminals. Whilst so-called vibe coding approaches now make it possible to formulate requirements in natural language and have AI systems translate them directly into code, this is, in itself, a remarkable development. Accessibility cannot, however, be fully realised in this way, because accessible interactive systems do not emerge from correct code alone – they emerge from user-centred development, that is, from the active involvement of the target audience. This process cannot be replicated by AI through the simulation of real users, though AI can provide meaningful support at specific points along the way.

What becomes apparent across all media types is this: the more complex the system, the more AI depends on human guidance and oversight. It can lay the groundwork, generate proposals, and accelerate routine tasks – but responsibility for the quality of the outcome remains with the human. And in some cases, it is not even desirable to remove the human from the development process altogether.

Using AI to Generate Alternative Texts

Generating alternative texts for visual elements is perhaps the most obvious use case for generative AI in the field of accessibility. On the face of it, the idea is straightforward: AI can analyse images and describe in text what it detects within them. This would, in theory, allow one of the most common barriers on the web to be addressed automatically – missing or inadequate alternative texts, upon which blind and visually impaired people depend in order to access visual content.

Here, however, lies the problem: a good alternative text does not simply describe what is visible in an image. Ideally, it conveys what the image means within its particular context – what information it carries, what purpose it serves, and what message it communicates. A photograph of a person might be a decorative element, the portrait of an author, part of a news item, or an illustration of a concept. Depending on its intended use, the ideal alternative text can differ considerably. The same applies to charts, pictograms, data visualisations, and screenshots: each follows a different descriptive logic, one that cannot be derived from the image content alone, but only from the context in which the image is used.

It is precisely this distinction that AI systems still struggle to make. The development of these systems can be well illustrated by a concrete example that you may well have encountered yourself: Microsoft Office – PowerPoint in particular – was among the early adopters of an AI-assisted function for the automatic generation of alternative texts. The initial results, however, were barely usable. The system produced simple enumerations of detected objects: “desk, laptop, plant, window”. This falls far short of a helpful alternative text. Quality has since improved noticeably – systems now describe images in a more coherent manner – but context-sensitivity remains elusive. One step is still missing: an understanding of the purpose of an image within its specific editorial context.

This does not mean, however, that AI has no value here – let us be clear on that point. Quite the contrary: as a support tool (once again!) for authors who write alternative texts themselves, AI can provide a useful starting point – a first draft, so to speak, which can then be refined as needed. What it should emphatically not do is take over this task entirely without human review. Alternative texts that no one has checked are, at best, imprecise and, at worst, misleading. For people with visual impairments, this risks creating a new barrier where none need have existed.

A Brief Aside: What Makes a Good Alternative Text? The W3C Decision Tree

Incidentally, authors are always best advised to write image descriptions themselves, rather than delegating this step to a third party. This is the surest way to ensure that the intended information is accurately conveyed. At the same time, authors need to have a sound understanding of what a good text alternative actually looks like. To this end, the W3C has published a decision tree on its website (Link: https://www.w3.org/WAI/tutorials/images/decision-tree/) designed to guide authors through the process. Do feel free to take a look – see Figure 1.

AI-Based Assistive Technologies

Assistive technologies – those tools that enable people with disabilities to use digital environments in a self-determined manner – stand to benefit from artificial intelligence in two distinct ways: existing technologies can be made significantly more capable through AI, and entirely new technologies can emerge that would not have been realisable without it. For me, this area ranks among the most exciting in the entire field of AI and accessibility.

A compelling example is the ongoing development of screen readers. Modern screen readers now integrate AI-assisted image description directly into their functionality – for instance, through the integration of dedicated plug-ins. This means that even where a website provides no alternative text for a graphic, the screen reader can independently generate a description, giving the user at least some indication of what the image is or might be. This shifts the dependency to a certain degree: accessibility is no longer entirely contingent on whether content creators have fulfilled their duty of care. A degree of autonomy that simply did not exist before. A word of caution against excessive enthusiasm, however – the limitations of automatic image description have already been discussed, and they apply here just as much.

More striking still is the range of assistive technologies that would not exist at all without AI. A well-known example is the OrCam system: a relatively unobtrusive pair of glasses with an integrated camera that analyses the surrounding environment in real time and describes it through audio output. Whether the wearer’s gaze falls on a text that can be read aloud, whether a person is recognised in front of them, or what colours and objects are present in their field of vision – for blind and severely visually impaired individuals, this opens up a form of environmental perception that was previously inaccessible.

Equally, far-reaching are applications designed for users with cognitive disabilities or neurodivergent needs. In the course of our AutARK project, which aimed to improve the employment situation of autistic individuals in the workplace, we encountered in numerous conversations just how concrete this need actually is. Many workplace situations present particular challenges for autistic people: correctly interpreting emails and implicit communication norms, recognising priorities, structuring tasks, or navigating direct conversational situations. AI-assisted systems can make a targeted contribution here – for instance, by processing incoming messages into a more usable form, identifying tasks and responsibilities, helping to formulate responses, or providing real-time support during live conversations. Speed is also an important factor for the technologies described here. Current AI is delivering meaningful advances in this regard, though it remains a challenge when additional requirements such as data protection and privacy come into play.

At this point, a conceptual clarification is important: not all of these applications are based on generative AI in the strict sense. Many assistive technologies draw on broader methods of machine learning – for instance, for image and speech recognition, facial recognition, or real-time classification. The term AI therefore encompasses considerably more here than language models and LLMs alone.

A summary overview of the application areas discussed, along with the possibilities and limitations described and any relevant recommendations for action, can be found in Table 1 below.

Limitations and Risks

As promising as the potential may be, a sober assessment of the limitations and risks of deploying AI in the field of accessibility is absolutely necessary. These limitations are often closely intertwined with the risks: those who are unaware of the boundaries of current systems run the danger of placing more trust in them than is warranted. This, in turn, gives rise to various risks in practice – born, one might say, of a certain naivety that is, however, readily addressed.

The Limitations of Current AI Systems

Current AI systems are capable, but not infallible. One of the most well-known structural problems of LLMs is what is referred to as hallucination: language models occasionally produce statements that are factually incorrect yet sound linguistically convincing. For some use cases, this is tolerable to a degree – provided a human reviews the output. It does, however, preclude the full automation of processes. In the field of accessibility, this can have concrete consequences: for instance, when a generated alternative text contains false information about an image and is adopted without review, leading to misunderstandings.

Beyond hallucination, AI systems also encounter substantive limitations: some tasks require more context or domain-specific knowledge than a model can reliably provide on its own. Assessing whether a web page is genuinely accessible to a particular target group, for example, cannot be derived from code alone – it emerges from real-world user experience. This experience is something AI cannot reliably simulate.

This, in turn, has a practical implication: the desire to dispense with specialist expertise in accessibility through the use of AI is unlikely to be fulfilled in the foreseeable future. Awareness-raising, methodological knowledge, and the involvement of people with disabilities remain indispensable. AI can, however, meaningfully complement this expertise – for instance, as a tool that enables faster and more efficient working. Those who deploy it as a substitute, on the other hand, risk producing outcomes that give the appearance of accessibility without actually being accessible.

Data Protection and Trust

A second risk – one that tends to be underestimated – concerns the handling of data. Public awareness of this issue is growing, though it frequently manifests as anxiety about using these tools, or avoidance of them altogether. Neither, of course, serves anyone well. To be clear: anyone using AI systems to review or improve digital content for accessibility is transmitting that content to the system in question – and potentially sensitive information along with it. This becomes particularly relevant when the content in question includes internal documents, personal data, or confidential material.

As a general principle, LLMs are not designed to permanently store inputs: the model processes a query and returns a response, without any learning process or data retention taking place. However, the providers of such systems are at liberty to log the data transmitted to their servers and to use it for the further training of their models. This is, in principle, a legitimate interest – but one that users must be aware of and factor into their decisions. This also opens up a fundamental tension: the most capable AI systems run on the external servers of large providers and are accessed via the internet. They deliver the best results, but require trust in the respective provider. Alternatively, AI models can be run locally on one’s own device, with the advantage that no data is transmitted externally. The drawback: local models are limited in their capability, as they must operate within the resources of standard consumer hardware and are consequently smaller in scale.

The good news is that not every task requires the full capability of a large external model. Many routine tasks can be handled to a sufficient standard by local models as well. The guiding principle should therefore be: as soon as sensitive data is involved, local systems should be preferred. For this to work effectively, the tasks assigned to the AI must be scoped accordingly. External systems should only be used where the provider’s data protection terms have been carefully reviewed, and no sensitive information is being transmitted.

Recommendations for Action: What Does This Mean for Your Organisation?

Three things you can do right now:

1. Introduce AI as a complement – but do not treat it as the one solution.

Speak with your IT department or service provider: which existing tools already support AI-assisted accessibility testing? Are there alternatives? By deploying AI systems thoughtfully, you can work towards meaningful and efficient improvements in accessibility.

2. Clarify data protection before deployment.

Before integrating AI tools into accessibility-related processes, make sure to establish which data is being processed and where it is going. For internal documents and personal content, the rule is: prefer local models, or scrutinise provider contracts carefully.

3. Do not replace specialist expertise – but deploy it more strategically.

AI saves time on routine tasks. Redirect that saved time towards the work AI cannot do: involving people with disabilities, assessing the quality of user experience, and raising awareness within teams. Accessibility remains a fundamentally human endeavour. When used correctly, AI makes it considerably more efficient.

A Word on the German Accessibility Strengthening Act

The Accessibility Strengthening Act (Barrierefreiheitsstärkungsgesetz, BFSG) has been in force since June 2025. Those who still have no accessibility strategy in place are late to the table and risk penalties – but above all, reputational damage. AI can help to make up lost ground. The first step is always the same: understand where your own barriers lie. From there, it becomes clear which tools – AI-based or otherwise – are appropriate.

Five Common Misconceptions

“AI will automatically make my website accessible.”

A very clear no. AI can identify problems and suggest corrections, but it cannot assess whether a website is genuinely accessible to real users. That always requires human judgement.

“Overlay tools solve the problem at the touch of a button.”

Quite the contrary: many of these tools interfere with the use of assistive technologies more than they help. Professional associations and those affected largely reject them.

“We are a small company. The BFSG does not apply to us.”

Partly true, but not as a blanket statement. Micro-enterprises are currently exempt, but the definition is narrower than many assume. Anyone pursuing public contracts or offering digital products and services should examine their own situation carefully. Likewise, It is conceivable that the exemption for micro-enterprises could be removed in the future. And ultimately: please consider what motivation you want to bring to the topic of accessibility – is it purely a matter of compliance, or does it stem from a sense of social responsibility?

“If no user has complained, there is no problem.”

People with disabilities simply leave offerings that demand too much effort of them. They often do not complain. This means only that you are receiving no feedback – not that no barriers exist.

“With AI, we no longer need accessibility expertise.”

The opposite is true: deploying AI requires people who can assess, contextualise, and correct its outputs. Without a basic understanding of accessibility, it is impossible to judge whether what AI produces is actually good. This applies to virtually every field in which AI is used – and accessibility is no exception.

Conclusion

To return to the term used in the introduction: artificial intelligence is, in the field of digital accessibility, far more than a passing trend. It genuinely has the potential to bring about real improvements, opens up new possibilities, and is already increasingly changing how accessibility is tested and implemented. That is the good news. But the full picture also includes this: the very high hopes that some attach to AI – the idea that accessibility might suddenly become self-sustaining, cheap, fast, and requiring no specialist expertise – cannot yet be fulfilled.

What AI can already achieve today is nonetheless quite remarkable: it accelerates testing processes, provides useful assessments of identified barriers, supports the remediation of content, and is driving a new generation of assistive technologies that enable greater autonomy for many people with disabilities. What matters most, of course, is how it is used – namely, as a support for human expertise.

At the same time, it is worth looking ahead: the areas in which AI systems still fall short today – contextual understanding and reliable quality assessment – will continue to develop. The pace at which the field has evolved in recent years gives grounds for hope that some of today’s limitations will no longer be limitations in a few years’ time. Academic interest is strong, the literature is growing rapidly. All of this is cause for optimism.

For those working in the field of accessibility, this means one thing above all: now is a good time to become familiar with these tools, to build on one’s own expertise, and to be well prepared for what lies ahead. And perhaps one thing more: in the course of my research, I repeatedly encountered descriptions of challenges that could likely be resolved if a more consistently human-centred approach were taken. So here, too, a call to action: wherever possible, involve your target audience in your development processes in a structured and systematic way. This is almost always a gain for everyone involved.

Let’s Continue the Conversation

If this topic is of relevance to you – whether as a developer, a communications professional, or an organisation that takes accessibility seriously – I am happy to serve as a point of contact. As a consultant and trainer with many years of experience in digital accessibility and human-computer interaction, I help organisations integrate AI-assisted approaches into existing processes in a meaningful and sustainable way.

Please don’t hesitate to get in touch directly – for a no-obligation conversation, a consultation, or a workshop tailored to you and your organisation.

References

1. Goenawan Brotosaputro, Agung Supriyadi, and Michael Jones. 2024. AI-Powered Assistive Technologies for Improved Accessibility. International Transactions on Artificial Intelligence (ITALIC) 3, 1: 76–84. https://doi.org/10.33050/italic.v3i1.645
2. Milton Campoverde-Molina and Sergio Luján-Mora. 2026. Artificial intelligence in web accessibility: A systematic mapping study. Computer Standards & Interfaces 96: 104055. https://doi.org/10.1016/j.csi.2025.104055
3. Caroline Kopp. 2025. Einsatzpotenziale von Large Language Models zur Verbesserung der Barrierefreiheit im automatisierten Website-Testing. Hochschule der Medien Stuttgart, Stuttgart. Retrieved from https://hdms.bsz-bw.de/frontdoor/deliver/index/docId/7244/file/Bachelorarbeit_Kopp.pdf
4. Khansa Chemnad and Achraf Othman. 2024. Digital accessibility in the era of artificial intelligence—Bibliometric analysis and systematic review. Frontiers in Artificial Intelligence 7: 1349668. https://doi.org/10.3389/frai.2024.1349668
5. Susan Ferebee. 2025. AI and Accessibility: Breaking Barriers for People with Disabilities. Premier Journal of Artificial Intelligence. https://doi.org/10.70389/PJAI.100012
6. Björn Fisseler. 2023. Künstliche Intelligenz, Behinderung und Technoableism. In Künstliche Intelligenz in der Bildung, Claudia De Witt, Christina Gloerfeld and Silke Elisabeth Wrede (eds.). Springer Fachmedien Wiesbaden, Wiesbaden, 313–330. https://doi.org/10.1007/978-3-658-40079-8_15
7. Andreas Hinderks, Artak Davtyan, and Dennis Allerkamp. 2025. Automatisierte Barrierefreiheitsprüfung mit Large Language Models: Eine vergleichende Analyse von GPT-4o und traditionellen Tools zur WCAG 2.1-Konformitätsprüfung. In Mensch und Computer. https://doi.org/10.18420/muc2025-up-362 or https://dl.gi.de/items/811bdef0-2f16-44cd-aa1a-bfaea0094764
8. Selina Hölzl and Christian Langenbach. 2025. Künstliche Intelligenz zur Verbesserung der Zugänglichkeit von Business-Graphiken für Menschen mit Sehbeeinträchtigungen. HMD Praxis der Wirtschaftsinformatik 62, 6: 1312–1331. https://doi.org/10.1365/s40702-025-01222-3
9. Juan-Miguel López-Gil and Juanan Pereira. 2025. Turning manual web accessibility success criteria into automatic: an LLM-based approach. Universal Access in the Information Society 24, 1: 837–852. https://doi.org/10.1007/s10209-024-01108-z
10. Niklas Metzinger. 2025. Einsatz von Large Language Models bei der Generierung von Alternativtexten für Grafiken und Bilder in E-Books. HTWK Leipzig, Leipzig. Retrieved from https://nbn-resolving.org/urn:nbn:de:bsz:l189-qucosa2-988979
11. Achraf Othman, Amira Dhouib, and Aljazi Nasser Al Jabor. 2023. Fostering websites accessibility: A case study on the use of the Large Language Models ChatGPT for automatic remediation. In Proceedings of the 16th International Conference on PErvasive Technologies Related to Assistive Environments, 707–713. https://doi.org/10.1145/3594806.3596542
12. Frank Reins and Birgit Scheer. 2024. Barrierefreiheit von Software, Apps und Web in Arbeitsumgebungen – Grundlagen für Berufliche Inklusion. In Die Rehabilitationstechnologie im Wandel: Eine Mensch-Technik-Umwelt Betrachtung, Vanessa Heitplatz and Leevke Wilkens (eds.). Eldorado, Dortmund, 481–506. Retrieved from http://hdl.handle.net/2003/42377
13. Sabrina Berg. 2025. Large Language Model unterstützte Barrierefreiheitsprüfung für Websites – Analyse und proof of concept. Hochschule der Medien Stuttgart, Stuttgart. Retrieved from https://hdms.bsz-bw.de/frontdoor/deliver/index/docId/7304/file/Masterarbeit_Sabrina_Berg_pub.pdf
14. Vagner Santana, Reinaldo Ferraz, and Jeniffer Vieira De Deus. 2026. On the Limits for Web Accessibility Evaluation Using AI Agents. In Proceedings of the 23rd International Web for All Conference, 187–197. https://doi.org/10.1145/3800424.3800445
15. Diana Schneider. 2025. Künstliche Intelligenz: Ein Motor für Inklusion? EthikJournal, 2: 73.

David Gollasch

Hi, I’m David! As a computer scientist and HCI researcher, I’ve spent many years working in the field of human–computer interaction and digital accessibility. There is significant potential here – socially, but also economically. That’s why I focus on solutions that turn accessibility into a competitive advantage rather than just a legal obligation. At the same time, I address questions of particular societal relevance by developing solutions for highly diverse user groups: neurodivergent individuals as well as the challenges of an ageing society.