AI to Support Neurodivergent Learners in Vocational Education and Training

3. Barriers to effective use of AI and other advanced technologies to support neurodivergent VET learners

AI and other advanced technologies hold promise for supporting neurodivergent learners in VET, however a range of barriers prevent the full potential from being realised. Learners, teachers and employers can be overwhelmed by the growing number of available tools, while VET teachers often lack the necessary skills, resources and encouragement to fully engage with new technologies. While desk- and computer-based work can easily accommodate new tools, the diversity of work and learning environments within VET presents practical and technical challenges. Affordability remains a significant barrier, limiting access and driving inequalities. At the same time, many promising technologies fail to reach the market, and those that do often do not align with the real needs of VET institutions, teachers and learners, and do not integrate well with existing systems. Furthermore, deployment of promising tools within VET will be uneven as long as gaps in infrastructure and connectivity persist, and as long as AI tools serve certain languages and cultural environments better than others.

Learners, teachers and employers are overwhelmed by the number of available tools

An overload of choice can make it difficult for learners, teachers and employers involved in VET to choose the right tools for their needs and environments. While interviewees generally welcomed recent advances in AI and other advanced technologies, including free and easily accessible generative AI chatbots, some mentioned the challenge for learners and teachers to keep pace with the emergence of new tools and new versions of existing tools. Christopher Patnoe (Google) explained that learners and teachers have a limited capacity to take in a vast amount of information, with the consequence that they are often unaware of useful accessibility features and that existing tools are not used to their full potential. Discoverability is a challenge that prevents new companies in the accessibility space from breaking through, as highlighted in previous OECD research (Touzet, 2023[46]) on the use of AI to support people with disability in the labour market.

It requires an enormous effort for a learner to keep up to date with all the assistive tools available. As an undergraduate student, Alisdair Gurling started researching tools to support people with ADHD, building complex systems to map and categorise them. What started as a hobby ended up being the starting point of his PhD, which focusses on mindset shifts around neurodivergent adoption of assistive technologies and led to his current work as a researcher at Wonderlab, Monash University. He described how learners with ADHD could be discouraged by having to navigate new websites and processes, set up multiple accounts and manage passwords, and keep track of syncing of various systems. His research has identified negative mindset as a further “hidden hurdle” that can lead to indifference and inertia around potentially useful tools.

“If you're telling yourself that you are broken and that you can't be fixed, then it produces a sense of indifference and inertia around adopting new tools. Why fix something? Why spend any time attempting to fix something that's irrevocably and permanently broken?” – Alisdair Gurling, researcher at Wonderlab, Monash University

Misinformation can thrive where new tools and features outpace robust and independent evaluation. Rajesh Ananda (Ultranauts) and Cristina AnaMaria Costescu (Associate Professor, Special Education Department, Babes-Bolyai University) expressed concern about companies making claims about their tools that are not backed by evidence, while Neil Milliken described how peer groups and social media can diffuse and amplify misinformation. The consequence is that learners and teachers are steered to tools that are not right for them.

Choosing the right tools is particularly important when resources are limited. Given the time required for users to adapt to a new tool in VET or in the workplace, it is not as simple as choosing the latest available tool, according to Hélène Chinal, Capgemini’s Southern and Central Europe Head of Transformation. The selection process itself can be resource intensive and inefficient. Piia Jokelainen, Educational Digital Specialist at Luovi Vocational College, explained that in many European countries, each institution is responsible for managing its own selection process, with the result that hundreds of institutions might each have to spend a couple of days checking whether a given tool is safe.

“Things move at such speed that the tools of one year ago are no longer the tools of today, so one needs to be constantly vigilant” – Hélène Chinal, Capgemini’s Southern and Central Europe Head of Transformation

VET teachers often lack the capacity and support to use AI and other advanced technologies to support neurodivergent learners

VET teachers need adequate preparation and institutional support to identify and respond to diverse learner needs, including through the use of AI and other advanced technologies. However, these elements are often lacking, according to interviewees. As a result, many VET teachers (including trainers) lack the skills and confidence to use AI and assistive technologies pedagogically for neurodivergent learners. Additionally, rigid curricula, risk-averse cultures and unclear guidelines limit experimentation, so even when devices are procured, they remain underused.

The VET teachers interviewed as part of this study generally used AI and other advanced technologies in their teaching and/or administrative tasks to support neurodivergent learners. Their interest in and knowledge of this topic is why they were invited to participate in the interviews for this project. In many cases, these teachers experimented with and developed skills in AI and other advanced technologies on their own personal initiative. However, this group is not representative of the full VET teacher population.

In reality, the level of skill in AI and other advanced technologies and the attitudes towards them among VET teachers varies greatly. In the most recent OECD Teaching and Learning International Survey (TALIS) of (non-VET and VET) teachers, three in four say that they lack the knowledge or skills to teach using AI while approximately half say that they do not believe AI should be used in teaching (OECD, 2025[47]). Additionally, a quarter of teachers reported a need for training on SEN.

VET teacher training can leave teachers without adequate preparation to use new technologies to address diverse needs, according to interviewees. Pierre Dillenbourg explained that in Switzerland, teacher training for VET teachers was unlikely to include much training on learners with neurodivergences or disability and on the tools available to support them. He gave an example of a typical profile of a VET educator, a 45‑year‑old carpenter who transitions to teaching carpentry due to a back injury, and the training they might receive. He imagined such a person receiving at most two hours of training on learning disabilities. He described VET teacher training as always 10 years behind the current technology. Speaking about the United Kingdom, Rohan Slaughter described a lack of training on assistive technologies even among specialist teachers, occupational therapists, speech and language therapists, and professionals who might support a student’s transition into the labour market. For example, the national professional qualification for SENCOs (Special Educational Needs Co‑ordinators or teachers in mainstream schools who oversee the strategic development of SEN policy and provision) was only changed in September 2024 to include intended learning outcomes on assistive technology. In the same vein, Susan Scott-Parker (Founder of Business Disability International) pointed out that initial teacher education and certification often do not require skills in inclusive, assistive and AI tools. She urged education authorities to build these requirements into pre‑service training so trainee teachers know which tools are likely to be used, how to use them, and how to keep up as technologies change. Otherwise, devices bought later go unused. She also noted examples in the United Kingdom where specialist providers such as Microlink train teachers, but said this depends on local initiative, stressing the need for earlier, system-level approach.

Once in the job, VET teachers have reason to engage cautiously with AI and advanced technologies. According to interviewees, they are often aware of the risks discussed in more detail in Chapter 4: that AI could undermine privacy; increase misinformation; encourage cheating; and hinder learning. Ann Kennedy, a tutor working in further education, explained that teachers generally want to ensure that the core principles and values around teaching are respected before introducing new technologies and methods. Freya Bevan, Digital Learning Coach at Gloucestershire College, estimates that 25% of teachers in an institution such as hers might have used AI out of their own initiative, with a further 25% using it once encouraged to do so, while 50% would oppose AI. In her view, opposition to AI is more likely among: more experienced teachers than less experienced ones; among teachers in higher education than in (vocation-oriented) further education; among teachers without industry experience; and among teachers of certain subjects (e.g. mathematics, English) than of others (e.g. computing, game design, forensics). With opinions varying significantly between VET teachers, institutions can play an important role in setting the culture, rules and guidelines for engaging with AI and other advanced technologies.

In some countries and contexts however, institutional culture and rigidities can discourage VET teachers from innovating and experimenting. Many interviewees mentioned that institutional change takes time. Thomas Köhler described an institutional culture in some central European countries which limits teachers’ freedom and mindset to explore. He said that he had seen some non-European institutions (for instance, in Africa) with much fewer resources that were more open in this regard. According to Lorenzo Desideri, some national education authorities dictate rigid curricula – often valorising traditional, pre‑digital methods – which inadvertently stifle teachers’ ability to innovate. In Italy, he explained, official guidance sometimes implies that “classic” approaches are superior, which discourages teachers from integrating AI solutions in creative, context-specific ways. Additionally, bureaucratic hurdles, such as the need for special authorisations to use nonstandard platforms, discourage experimentation.1

One consequence is that teachers often lack a clear policy or guidelines around the use of technology, which could provide practical implementation tips and encouragement, while alleviating concerns about accidentally breaching internal rules or regulations. In Estonia, national teacher guidelines for supporting special education needs learners currently do not include references to AI, reflecting the early stage of AI integration in practice, according to Sandra Fomotškin, an Advisor on Inclusive Education at the Ministry of Education and Research. A lack of ongoing support and unclear guidelines are reasons why instructors often revert to surface‑level use of technologies instead of exploring the full functionality according to Lorenzo Desideri. In his view, schools invest heavily in purchasing devices but underinvest in fostering the skills and institutional flexibility required to leverage these tools optimally. Nwanneka Udeka, a speech-language pathologist highlighted the importance of clear implementation guidelines that show instructors exactly how to incorporate the technology into lesson plans. A lack of guidelines – combined with insufficient continuing professional learning and competing classroom demands – is why she estimates that only about 30% of AAC devices provided to students by speech-language therapists are later used in the classroom by the teacher. Piia Jokelainen reported that only about 10% of teachers in Finland are active users of AI tools, while more than half express interest but find the tools too complex or lack of guidance to explore them. This indicates a latent readiness that could be unlocked through targeted support and simplified implementation models.

Teachers often have few opportunities to observe how inclusive digital tools are being used in workplaces – such as AI-assisted communication, adaptive task management systems, or sensory-friendly technologies – resulting in a disconnect between the classroom experience and the evolving demands of inclusive labour markets. Rohan Slaughter and Pierre Dillenbourg noted that while companies increasingly deploy AI to enhance accessibility and personalise tasks, VET teachers rarely have structured exposure to these developments. Hiren Shukla (EY) and Francesc Sistach (Specialisterne) further highlighted that neuroinclusive employers already leverage AI to tailor onboarding, communication, and productivity tools for neurodivergent staff, yet these approaches seldom inform VET pedagogy or teacher training. Brad Tombling (Bud Systems) added that rigid training plans and slow curriculum adaptation make it difficult for providers to align learning with such innovations.

Both Lorenzo Desideri and Nwanneka Udeka highlighted the importance of ongoing collaboration among all stakeholders around the student, including teachers, parents, peers and support professionals. Without this, AI and other advanced technologies will not be used to their full potential to support neurodivergent learners in VET, because tools will not be deployed in the first place or because useful features of deployed tools will remain dormant or inconsistently applied.

Attitudes and lack of awareness among employers can be a barrier

Attitudes among employers can also be a barrier to deploying AI and advanced technologies to support neurodivergent workers and/or learners making the transition to work. According to Hiren Shukla and Heather Tartaglia, both of EY, many organisations underestimate the productivity gains and innovation potential associated with inclusive AI use, and a lack of concrete examples and negative first experiences further dampen adoption. Nicole Lonican, of the Cuimsiú Inclusive Pathways to Employment Programme, noted that despite generous resources provided by the Irish Government (reimbursement of 100% of the cost of assistive technology and 90% of accessibility awareness training), Irish employers were sometimes unaware of the supports available. In France, as noted by Redwane Bennani, CEO of Talents Handicap, employers often lack knowledge of the regulatory framework, of working with people with disabilities, and of the supports available.

Challenges remain in applying AI to the diversity of work and learning environments in VET

Applying AI and advanced technologies in VET to support neurodivergent learners and workers remains difficult, as many of the roles that VET prepares learners for rooted in manual or practical activities, where one works on their feet, and often in noisy, shared and unpredictable settings (like kitchens, workshops or hospitals). The application of AI to these work environments and to VET has been much slower than in office environments and academic learning due to the diversity of work environments and associated technical challenges, such as portability and edge limits and the lack of rich, industry- and region-specific contextual data to train AI models.

Portable solutions are useful where VET prepares learners for roles where they will work on their feet. Yet there seems to be a knowledge gap on how AI and advanced technologies can be leveraged in these roles and tasks. Assistive technologies have traditionally been oriented around desk-based work and learning and the process of integrating AI and augmented reality into cognitive assistance tools hosted on portable devices (e.g. phones, tablets, glasses or watches) has been slow. Robert McLaren gave an imaginary example of a hospital porter loading an industrial washing machine and receiving customised task reminders and tips on their smart glasses prompted by the porter’s location or their usual behaviours. He explained that many cognitive assistance and coaching tools are in development for deployment in situations such as this, but that they tend to be based on QR codes and NFC (Near Field Communication) rather than leveraging the latest advances in AI or augmented reality to personalise support to the user and environment.

“There's a bit of a bias whereby the industry and government programmes skew towards desk-based roles so people with learning disabilities/neurodiversity in those roles are more likely to get access to assistive technology than people whose roles or their vocational training doesn't involve sitting at a computer very much […] People haven't really thought enough about how technology can work within those roles” – Robert McLaren, Director of Policy, Policy Connect

Part of the challenge is technical. Yonah Welker (Public Technologist and Visiting Lecturer) described how recent advances in visual and 3D foundation models could have significant potential to support a VET learner or worker with disability and/or neurodivergences, but technical challenges (e.g. energy and autonomy requirements) can prevent models from being hosted on portable devices. He suggested that advances in adaptive scaling (to adjust computing resources in real time according to demand), offline inference (to run models in low-connectivity environments) and edge optimisation (to adapt models to device constraints) could allow AI to be deployed more reliably and efficiently in resource‑constrained, real-time environments.

Another technical challenge is a lack of appropriate training environments for AI. Before the latest advances in computer vision, speech recognition and robotics can be usefully deployed in complex work environments (e.g. a hospital) the models must be trained on large amounts of contextual data. While useful datasets already exist showing people navigating everyday activities in real-world settings,2 Yonah Welker described the current suite of video datasets as good enough for use in the Metaverse or for entertainment purposes, but lacking the richness needed for AI to be successfully applied in contexts such as vocational education, where tools will need to be tailored to very specific skills and to the terminology of a particular industry, country and/or region. Clayton Lewis noted that advances in machine learning had reduced the need for contextual training data. In the past, he explored the idea of developing a “job coach” virtual assistant that could, for instance, help a trainee gardener navigate a physical space and handle any safety risks. He concluded that the project was not feasible when the experts asked for 10 000 labelled examples of potential situations that could arise. Today, tools such as ChatGPT can identify, without any specific contextual training, that a picture of a lit match held up to a wasps’ nest depicts a risky situation. While Clayton Lewis acknowledged that rich contextual data would still be needed to support complex applications such as a “job coach” virtual assistant, this suggests a multifaceted solution to this challenge.

“In school, an AI tutor can help you understand history. In VET, AI has to help you safely wire a circuit, measure medication or navigate a shop floor, often with one hand or through a haptic interface” – Yonah Welker, Public Technologist and Visiting Lecturer

Affordability is a barrier to access and a driver of inequality

Affordability of AI and advanced technologies can mean that useful tools remain inaccessible to public institutions and to VET learners with low and limited resources, or remain undeveloped beyond the pilot stage. Veronika Kaska, Deputy Director of the Astangu Vocational Rehabilitation Centre, described how financial limitations had prevented her centre from implementing a VR tool that she thought had the potential to help learners with a wide range of disabilities in pre‑vocational courses to practise and strengthen everyday skills. Francesc Sistach saw limited budgets as the reason why there is not yet any comprehensive, off-the‑shelf workplace coach for neurodivergent users, despite some university labs and startups developing promising pilot projects and prototypes. Although many generative AI tools currently have a free version, paid models could become more common in the future, limiting access for schools, employers and individuals that cannot afford to pay. Nathaniel Cook pointed out that access to technology is already a challenge for many people with disability.

Access to AI and advanced technologies for institutions, employers and individuals is often funded by government. Yet, a number of interviewees observed that neurodivergent learners in VET were not priorities for public funding. For instance, Carlos Pereira, CEO of Livox, described how his assistive technology devices are often bought by cities, but only if there is money left in the budget, since people with disability are often not considered a priority. While funding for all education is limited, Veronika Kaska remarked that governments often prioritise scientific research and higher education over VET. Robert McLaren pointed out that, in the United Kingdom, the Disabled Students Allowance, which can cover the costs of assistive technology, is only available to higher education students, meaning that most VET students cannot access it. They will instead need to rely on Education, Health and Care plans (which in many cases do not cover assistive technology) or on college resources (which would be higher in SEN institutions targeted towards learners with learning disabilities).

“People with disabilities, they are not a priority. Although they are the biggest minority on Earth – there are 1 billion people with disabilities on the planet – they are always left behind” – Carlos Pereira, CEO, Livox

Many interviewees described current funding levels as insufficient. Elisabetta Bertola (Irisbond) described the situation as high-cost devices meeting limited public subsidies. She remarked that public funding varied significantly across EU Member States, with higher funding available in Germany and France than in Spain, Italy and Portugal. In Chile, public funding has been provided, but at an insufficient level to scale or sustain broader impact, according to Ricardo Rosas, Professor at the School of Psychology at the Catholic University of Chile. Many assistive technologies are costly due to high development costs combined with a lack of competition, according to Yonah Welker. Hardware for many simulation tools is expensive, requiring a substantial upfront investment, according to Jan Schlueter, whose company has created SANDI (Simulator for Advanced Neurodiverse Driving Instruction).

Part of the challenge is convincing decision makers of the long-term return on investment. Current evidence on assistive technology’s impact tends to be fragmented, small-scale and qualitative in nature, according to Robert McLaren. Without more robust impact evaluation, he says that it is difficult to justify government funding and large‑scale integration of AI tools in VET. Susan Scott-Parker explained that the long-term benefits of early intervention could be difficult to measure and evaluate in financial terms. However, even where substantial financial returns are established,3 it can be difficult to translate this into action when key officials responsible for education and training do not articulate the rationale for investing in assistive technology and in related teacher training (according to Susan Scott-Parker), and when there are competing priorities and a lack of political will (according to Francesc Sistach).

In some regions and systems, the process for individuals and employers to apply for access to assistive technologies is a further barrier. Elisabetta Bertola (Irisbond) explained that application processes in Spain and in other EU member countries were complex and lengthy, with the result that only 10 to 20% of those in need get the necessary support. In the United Kingdom, delays accessing the Access to Work grant (which can cover the cost of assistive software) reduce its effectiveness, according to Geena Vabulas. Delays mean that an individual might not be able to access the funds until six months after a job offer, which might be too late for the individual to demonstrate their performance and suitability to the work and might be entirely incompatible with short-term work experience. Where access to supports depends on disclosing a condition or providing a formal diagnosis, this represents a further barrier to assistive technologies, as discussed in Chapter 1.

Funding challenges naturally create inequalities in access. Yonah Welker cited a WHO study showing that in low- and middle‑income countries only 10% of those in need can obtain assistive technologies in contrast to 90% in high-income countries (Stawiska, 2024[48]). Nwanneka Udeka described how the availability of AAC and executive‑functioning devices varies greatly by region, school funding and insurance coverage.

Infrastructure and connectivity still constrain use of AI and advanced technologies in VET, particularly in training centres and workplaces in rural or under-resourced regions. Marius Frank explained that even in developed countries, such as England, some students and some schools do not have good broadband connectivity, meaning that the potential of AI cannot be realised.

“A critical consideration across every country is connectivity and infrastructure, because otherwise everything we're talking about becomes impossible” – Marius Frank, Education Director, Microlink PC

Many promising assistive technologies fail to make it to market

The process of bringing an assistive technology from lab to market is lengthy and filled with obstacles. Many technologies that show promise in initial research stages will never reach the VET system because of difficulties accessing venture capital and government funding, difficulties scaling and a lack of co‑ordination between actors in this space. Many of the same challenges were discussed in an OECD report (Touzet, 2023[46]) exploring the use of AI to support people with disability in the labour market, which concluded that many promising AI solutions fail due to difficulties establishing a sustainable business model.

Yonah Welker described the typical path to commercialisation, based on his own experience working with start-ups developing assistive technologies. Start-ups will bootstrap at first, supplementing where they can with university grants and competition awards, to develop and to ensure compliance with the relevant standards and frameworks (e.g. AI, medical, disability rights). This process can be lengthy; he has seen this process last 10 years for one European start-up developing a tool to support learners with ASD. Attention from venture capital funds only arrives at a later stage and the venture capital process for assistive technologies can be lengthier than for other products, requiring sustained effort and investment and multiple prototypes. Pavan Konanur, co-founder of Hoja AI, spoke about how the funding journey for the education platform has been challenging. At the beginning, they bootstrapped with their own savings to get to testing/validation stage, before attracting additional funding from family and friends. They are still on the path to commercialisation, talking to investors, investigating grants and participating in some accelerators which provide free LLM tokens.

One barrier to accessing venture capital funding can be the perceived small size of the market (i.e. neurodivergent learners in VET and/or in general education) and associated commercial potential. It is not enough to demonstrate that a new tool addresses a real need, according to Motti Sigel of MassChallenge; venture capitalists must see business traction in the form of a defined market with paying customers. Yonah Welker agreed that the small size of the market was often a reason for venture capital firms not to invest in assistive technologies but he also felt that this was a distorted view of the market, which ignores benefits for other segments of the population. For instance, tools developed to support young neurodivergent people in VET could also benefit a huge number of older people with cognitive decline. However, Clayton Lewis was generally sceptical of the profitability of technologies targeted specifically at people with disability and/or neurodivergences and thus sceptical that tools that meet their needs would emerge from the market without additional public and philanthropic funding.

“My guess would be the main barrier to these assistive technologies would not be the technology, it would be ‘who's the payer?’” – Motti Sigel, Managing Director, MassChallenge Israel

Public and NGO funding can be necessary to align development with the public interest. Yet a few interviewees identified flaws in these systems. Francesc Sistach, whose organisation Specialisterne helps people with ASD and other neurodivergences to find employment, described how government grants were insufficient for organisations such as his to develop and implement AI. Speaking about a recently awarded EUR 90 000 grant, he said that this money would be useful for improving the existing e‑learning system, but that developing any advanced AI tool, such as a virtual coach, would require 10 times as much. Government and NGO funding for innovation are often more accessible to non-profit organisations than for-profit organisations, according to Brad Tombling, Chief Operating Officer at Bud Systems, a training management platform for apprenticeships and skills delivery. In the past, they have benefitted from the United Kingdom’s R&D tax credits but these become less generous and less viable sources of funding as the company grows.

“That help is very good, but with €90,000 we are not going to develop a training system with artificial intelligence” – Francesc Sistach, CEO, Specialisterne Global

For many tools aimed at the VET and broader education sector, scaling relies on securing a government contract. Yet many startups underestimate how difficult this is to achieve, according to Motti Sigel. He explained that procurement in education is slow and risk-averse, particularly in countries with centralised systems. In his view, NGOs and local intermediaries were often more agile partners. Another way for startups to scale is to integrate their product into an existing platform, according to Cristina Anamaria Costescu, but this may involve concessions.

“If a founder came in and they say ‘Listen, I think this is amazing. I want the Minister of Education to sign a contract with me’. I would tell him that that's laughable. That is never going to happen” – Motti Sigel, Managing Director, MassChallenge Israel

Good communication and co‑ordination between market actors across commercial, non-profit, research and open-source environments could help promising technologies thrive. Yet these actors are not well connected, according to David Banes. The consequence is that many useful research findings remain behind paywalls, limiting knowledge transfer and innovation. However, Yonah Welker suggested that there have been improvements in the funding ecosystem in the last five to seven years, with better co‑ordination between accelerator programmes, community and governmental programmes and different foundations. For instance, venture verticals have emerged wherein big pharma, medical or educational companies partner with existing startup accelerators to incubate cohorts of assistive technologies.

Many tools do not meet the real needs of VET institutions, teachers or learners

Many tools that do make it to market do not align with the real needs of VET institutions, teachers and learners, according to interviewees. This criticism was applied to mainstream tools – in which accessibility is a by-product rather than a first intent – as well as to assistive technologies specifically aimed at learners with disability and/or neurodivergences. Susan Scott-Parker explained that the needs of learners with disability and/or neurodivergences are often overlooked in a rapidly evolving market.

“There has to be a point where we follow the needs of the user, not the excitement of the developer” – Susan Scott-Parker, Founder of Business Disability International

Interviewees explained that mainstream tools are generally designed for the “standard user” and therefore often fail to cater for diverse cognitive and communication needs. Elisabetta Bertola described how various messaging, social platforms and productivity suites assume neurotypical users and offer only a “one-size-fits-all” interface. A narrow focus on profitability means that features like simplified language, lower reading levels or support for executive function are often missing, according to Nathaniel Cook.

“A lot of developers are not thinking of use cases for anybody with a disability, including those who are neurodiverse. They also wouldn't think of people who are elderly, who experience memory loss and so on. So we do have this problem that the data and the use cases that they're building are very, very mainstream and often exclude the outliers and particularly people with disabilities” – David Banes, David Banes Access and Inclusion Services

A number of interviewees spoke about the misalignment between LLMs and the needs of VET institutions, teachers and learners. One source of misalignment highlighted by Helen Nicholson-Benn is that these tools were not designed originally for education. While she welcomed OpenAI’s announcement of a 5‑year project whereby they will work with 400 000 teachers in the US, she wished that LLM developers had engaged with the education sector on issues of safety and accessibility at an earlier stage. Efforts to retroactively fit a tool to the needs of a sector and to diverse user needs are welcome but not ideal. Even though generative AI chatbots such as ChatGPT, Gemini and Perplexity are generally considered intuitive and accessible, they often produce outputs which can be overly wordy for neurodivergent learners, unless specifically prompted not to, explained a Curriculum and Learning Manager at Tech Kids Unlimited. Nathaniel Cook mentioned that the Special Olympics had built prompt libraries for people with disability and/or neurodivergences to use to instruct generative AI chatbots to simplify or otherwise adapt language, a workaround that would not be necessary if the relevant configurations were already available. Veronika Kaska described as common the experience where misalignment between tools and needs requires additional (financial and other) resources to adapt them.

Another source of misalignment is that AI appears to serve certain languages and contexts better than others. Interviewees noted that generative AI chatbots and screen readers perform much better in well-resourced languages such as English than in other languages, so not all learners benefit equally. This is a function of the amount of text in each language available online to train the AI, which can not only drive differences in quality but can also imbue the model with conventions, styles, biases and narratives associated with those languages. According to Yonah Welker, ChatGPT is driven by 400 billion texts in English, 1 000 times as much data as the 400 million texts in Hungarian, with the result that accuracy on topics relevant for the Hungarian context will be much lower. The context in which the technology is developed can also shape how the tool works: assistive technology tools developed to meet the needs of neurodivergent learners in the United States community colleges may not be suited to the VET context in Germany, for example.

“Let’s say you try to introduce some assistive technology in Slovenia or Romania… in some smaller languages, the quality and accuracy will be smaller” – Yonah Welker, Public Technologist and Visiting Lecturer

Although assistive technologies are designed to serve people with disability and/or neurodivergences, they too can fail to meet user needs. One issue (discussed in Chapter 4) is that as LLMs are increasingly embedded in assistive technologies, ableist biases present in the data used to train the models could seep in if users’ needs are not centred. Similar issues were identified in an OECD study (Touzet, 2023[46]) exploring the use of AI to support people with disability in the labour market, in which lack of engagement with the end-user in the development of assistive technologies was the most commonly cited barrier to adoption.

Designing assistive technologies to be inclusive of all needs, following universal design principles, comes with its own tensions and practical challenges. Alisdair Gurling described how trying to serve the needs of learners with ASD and learners with ADHD could drive developers in opposite directions. Ricardo Rosas spoke about his own experience trying to implement universal design principles in a project but ultimately concluding that it was not possible to develop the tools in a way that served both deaf and blind users simultaneously.

“Universal design can be insufficient because sometimes the need of someone with ASD is in tension with someone with dyslexia or ADHD for instance: sensory overwhelm on one side and desire for novelty and engagement on the other” – Alisdair Gurling, researcher at Wonderlab, Monash University

How the technology aligns with existing teaching methods is also important. In explaining why VR and AR tools are not yet widely used in educational settings despite their potential (see Chapter 2). Kevin Gonyop Kim, a Professor for Spatial Computing and 3D Technologies, described a fundamental misalignment between VR headsets and the one‑to-many classroom dynamic, in that headsets physically block sight and communication between the teacher and their students, and between students. While they could facilitate other forms of learning (e.g. one‑to‑one tutoring, distance or independent learning), they seem incompatible with current classroom-based VET. He saw other forms of spatial computing, for instance based on projections or holograms, as more promising technologies for a collaborative classroom dynamic. Other interviewees raised concerns about the limited availability of software suitable for VET. For instance, Brad Tombling, COO of Bud Systems, expressed the view that simple tools like text-to-speech, speech-to-text could have more impact than large Augmented Reality projects, which are likely to be too costly to develop and implement in a VET context.

Tools are not well integrated with each other

According to interviewees, insufficient integration between mainstream and assistive technologies limits their potential to assist neurodivergent learners in VET. This is referred to as interoperability and was also highlighted as a barrier to the adoption of AI to support people with disability in the labour market, in previous OECD research (Touzet, 2023[46]).

Interoperability is a particular challenge when it comes to AAC software such as screen readers and speech recognition, as these tools change the way a person interacts with a device (e.g. a computer or a phone) and all its software. Elisabetta Bertola, AAC Specialist Co‑ordinator at Irisbond, described a stark divide between specialist AAC and mainstream software (e.g. messaging, social platforms and productivity suites). In her view, each sits in a separate niche, designed with a separate audience in mind, evolving according to its own timeline, with little crossover. As well as limiting interoperability between the two sets of solutions, this divide also slows innovation in AAC software and prevents mainstream tools from understanding and catering to diverse needs. Alisdair Gurling described many technologies in this space as “little walled gardens” and provided a vision of what voice assistants could offer if these walls could be broken down: a VET learner with ADHD could ask their voice assistant about the day’s schedule and the assistant would draw on to-do lists and calendars from all relevant systems on the device, not only those affiliated with the brand of the voice assistant.

Limited interoperability also hinders the transition from training to work. Access to assistive technology is often tied to institutions meaning that individuals lose access to useful tools once they leave VET, according to Susan Scott-Parker. There are a few exceptions that try to preserve continuity. For example, according to Nicole Lonican from FIT in Ireland, their Cuimsiú programme provides multi-year licences and mentoring after participants move into jobs. However, interviewees stressed that such practices are not widespread, leaving many learners without their supports at the point of transition. Even where the individual retains access to their preferred tool, the new employer may object to the tool being used on its processes and on its devices. Interoperability poses practical challenges for employers that wish to deploy assistive technology, according to Neil Milliken, as the process of ensuring interoperability at employer level is complex and time‑consuming. For this reason, many employers will prefer to support only a limited selection of AAC tools and may be reluctant to switch to new solutions even if they seem to hold promise.

“One of the challenges that new assistive tech providers have is actually the technology may be a great idea, but if it doesn't integrate with the environment which someone is studying in or working in, it is only partially useful” – Neil Milliken, Global Head of Accessibility and Digital Inclusion at Atos

Notes

← 1.  In the United Kingdom vocational training system, rigid frameworks such as predefined training plans and compliance requirements pose a significant barrier to implementing personalised or adaptive learning pathways at scale, according to Brad Tombling, Chief Operating Officer of Bud Systems.

← 2.  Ego4D is one such dataset, comprising 3 670 hours of video of 923 participants performing everyday activities in real-world setting, allowing AI to understand functional relationships between tools and gestures.

← 3.  Francesc Sistach cited a Danish study that showed that every Danish kroner invested in the labour inclusion of people with autism ends up generating 2.4 kroner in taxes. In the United Kingdom, Neil Milliken estimated that the Access to Work grant had a 20% return on investment.