AI to Support Neurodivergent Learners in Vocational Education and Training

4. Risks in using AI and other advanced technologies to support neurodivergent VET learners

While AI and other advanced technologies offer new opportunities to support neurodivergent learners in VET, their use also raises a range of ethical, pedagogical and societal risks. Tools that collect and process learner data raise privacy risks while AI can replicate and perpetuate existing biases. Discrimination against neurodivergent individuals is a serious harm associated with these risks. Educators highlight the propensity for generative AI to foster misinformation and are concerned that overreliance on technology more generally could ultimately hinder the development of skills that VET learners will need to transition to work. Meanwhile, teachers’ concerns about AI-assisted cheating risks constraining the legitimate use of assistive technologies that were previously sanctioned. The socio‑emotional risks associated with generative AI chatbots and robots are not yet fully understood, not least for neurodivergent learners. Importantly, the use of AI and other advanced technologies could inadvertently lead to exclusion and widen the gaps that already exist. Many of these risks are not specific to neurodivergent learners in VET – similar risks have been outlined in previous OECD work focussing on equity and inclusion in education (Varsik and Vosberg, 2024[43]), on the use of AI to support labour market inclusion of people with disability (Touzet, 2023[46]), and on the use of AI in training (Verhagen, 2021[49]) – but some of the risks and harms can be heightened for this group.

Tools that collect learner data raise risks related to privacy

While data collection enables adaptive learning, it also raises risks related to data privacy, which were often top of mind for interviewees. Although the harms associated with data protection and privacy violations are not limited to neurodivergent learners, the risks and harms could be amplified for them due to the sensitivity and nature of some of the data collected. In VET, additional privacy risks arise when data are shared between schools, training centres, and employers during apprenticeships or work-based learning, where responsibilities for data protection are often unclear. OECD research (Touzet, 2023[46]) notes that “the current privacy protections don’t work if you are highly unique”. At the same time, some interviewees were concerned that an overly prescriptive approach to these risks could leave neurodivergent VET learners deprived of genuinely useful tools.

Some of the main data privacy risks highlighted by stakeholders include:

• A lack of transparency around what personal data is collected and stored by generative AI tools. One interviewee noted that this was identified as a high risk associated with Microsoft 365 Copilot in a 2024 Data Protection Impact Assessment (DPIA) conducted by SURF (Privacy Company, 2024[50]), a co‑operative of Dutch education and research institutions dedicated to ICT innovation. The co‑operative first advised education and research institutions against using the tool entirely but updated this in 2025 to advise exercising caution when using the tool, after having engaged with Microsoft to resolve some of the risks.

• Data being collected for one purpose and used for another. Some interviewees expressed concern about the potential misuse of data collected through tools designed to support neurodivergent VET learners. Janus Asko (EyeJustRead) described his concern that competitive pressures within the edtech market could eventually push small companies to sell their model and/or their user data to larger AI developers to improve virtual assistants and contrary to original intentions.1 Colm McNamee put the spotlight on the hidden costs of “free” services, while also warning about large tech providers and IP misuse; he added that the European Union is taking a comparatively progressive, “soft” regulatory approach in this space.

“I do have huge concerns about some of the large tech providers […] as someone said, if something is free, you’re the product” – Colm McNamee, Cuimsiú Employability Mentor, FIT

• The sensitive nature of data relating to disability status, neurotype or physical and emotional state of the learner. Health-related data is considered within GDPR to be “sensitive” and thus subject to specific processing conditions. Nasser Siabi and Marius Frank warned about this data being stored permanently and leading later to discrimination in education, employment or health insurance outcomes.

• Monitoring and surveillance associated with the use of AI-enabled wearables and AR tools. Robert McLaren explained that these tools would capture – whether intentionally or not – a very detailed picture of how a person works and/or studies, where they are at all times, and what they are doing. As a result, a neurodivergent worker or learner could be subject to a much greater degree of surveillance than their neurotypical peers, which could create disparities in how performance or learning outcomes are evaluated.

• Privacy risks for individuals other than the worker or learner. Geena Vabulas gave an example of smart glasses being used to support a learner working in a care home, collecting sensitive information on the patients being cared for. While not raised by any interviewee as a major concern, interviewees did note that even more benign-seeming tools for recording and transcribing lectures can have implications for lecturers’ and teachers’ privacy (as well as their intellectual property).

While privacy risks were widely acknowledged, some interviewees noted the challenge in striking the right balance; protecting privacy rights of neurodivergent learners without depriving them of useful tools. Yonah Welker described some initial worry in the assistive technology and ASD communities that useful emotion-recognition assistive technologies would be caught up in a ban under the EU AI Act on the use of AI systems to infer emotions in workplaces and education institutions. An exemption was ultimately included in the EU AI Act to clarify this issue and to explicitly permit the deployment of emotion recognition for medical or safety reasons.

David Banes suggested that the risk-benefit ratio could be different for people with and without disability or neurodivergences. If AI and other advanced technologies live up to their potential in enabling neurodivergent people to participate fully in VET and in the labour market, then people with neurodivergences could be more willing to accept privacy or security risks in exchange for the advantages. This has implications for policymakers: if they legislate on the basis of the risk-benefit ratio for the neurotypical population, then they could prevent neurodivergent people from accessing tools that are genuinely helpful to them. If legislation is looser, this leaves neurodivergent people facing greater risk than neurotypical people. David Banes spoke of the importance of neurodivergent people having a voice in these discussions.

Some interviewees also noted that consent can become a practical challenge when using potentially helpful tools. Geena Vabulas described cases where uncertainty about whether a learner has capacity to consent to data processing, or whether a parent, carer or provider may lawfully consent on their behalf, leads staff to avoid using otherwise useful technology. She highlighted the value of tools that do not collect personal data and can be used without creating an account or ticking a consent box (for example, Goblin Tools), because they reduce administrative and legal friction.

AI can replicate and perpetuate societal biases

Many interviewees spoke about the risk of bias associated with AI systems trained on historical data and thus primed to replicate and perpetuate societal biases, including assumptions about what counts as a “normal” body or mind. These ableist biases not only afflict AI-generated content but can also affect a learner’s transition to the labour market when AI is used in recruitment tools. Similar risks were highlighted in a recent OECD report (Touzet, 2023[46]) on the use of AI to support people with disability in the labour market, which identified two main sources of bias: that people with disability and/or neurodivergences are excluded from datasets used to train AI and that training data can reflect embedded ableist biases that are being scaled and magnified through AI.

The first issue is the lack of diversity in the underlying data. Neil Milliken explained that since AI systems work by identifying patterns, they are well suited for classifying information that has been observed many times in the historical data. In this paradigm, people with disability and/or neurodivergences will be considered “statistically insignificant” and thus either excluded from classification or classified with much less precision than people without disability and/or neurodivergences.

The second issue is ableism within the underlying data. According to Nathaniel Cook, current generative AI systems reflect embedded bias, including ableist assumptions, due to being trained on internet data. He observed that these systems tend to offer individuals with intellectual and developmental disabilities help with life skills rather than business skills, thereby showing an imbedded assumption that these individuals are less capable. Rohan Slaughter expressed concern that the same ableist attitudes could potentially seep into AAC tools if LLMs are imbedded into these tools without care or without consulting users.

“I'm willing to bet that every single LLM or AI that we utilise today was trained on the internet. And if you've ever used the internet, it is not a place full of acceptance and tolerance for people with intellectual and developmental disabilities” – Nathaniel Cook, Chief of Information and Technology, Special Olympics

Many interviewees expressed concern about AI being used in recruitment because of the potential for discrimination and harm when using a biased system for decisions that directly affect an individual’s livelihood and ability to transition to the labour market. Many of the same concerns would also logically apply to tools that match VET learners to educational opportunities. For example, bias in guidance, admission or tracking systems used in VET could reinforce the historic under-representation of neurodivergent learners in higher-level or prestigious vocational tracks. At the screening stage of the recruitment process, AI-enabled recruitment tools will often favour a profile that matches those already in the job, according to Michael Fembek (CEO, Zero Project, Essl Foundation), while anything that diverges from the norm raises a red flag. Thorkil Sonne explained that this rigidity will naturally screen out many neurodivergent VET learners, including for example, learners with non-traditional educational or career paths, or those job applicants with ASD who prefer not to boast about their achievements. At the interview stage, AI-enabled video interview platforms introduce further risk of bias if they are designed to value neurotypical behaviours. For example, Neil Milliken and David Banes both spoke of their concern that a difficulty maintaining eye contact with the camera could be misinterpreted and ultimately lead to exclusion.

Generative AI can foster misinformation

Interviewees highlighted the danger of neurodivergent VET learners relying on generative AI that often provide incorrect, inconsistent or compromised information. Generative AI is trained using large amounts of online data, which can include biased, incomplete and unverified information.2 Moreover, generative AI has a widely acknowledged tendency to “hallucinate” and to pander to the user’s demand rather than responding “I don’t know”. Interviewees were concerned that AI-generated misinformation could compromise VET learners’ ability to learn and actively mislead and deceive them. Some interviewees suggested that neurodivergence could make learners more susceptible to the risk of misinformation. Many interviewees agreed on the need to review generative AI outputs for accuracy. -

A related issue is the capacity for new AI models to generate images that can pass for real photos and video footage. Freya Bevan described deepfakes as a worldwide problem to which young students could be particularly susceptible.

“Teenagers are very susceptible to what they see, so sometimes it's just a case of seeing something and believing it, which is the case with many things, you know, like social media and all these trends on TikTok” – Freya Bevan, Digital Learning Coach with AI focus, Gloucestershire College

David Banes was concerned about the potential for commercial influence in algorithmic recommendations, particularly for free‑to‑use tools. He explained that users – especially those with cognitive disabilities – may find it difficult to recognise when an AI response is influenced by commercial sponsorships (e.g. recommending a specific product) and may accept answers at face value, thinking that the information they receive is independent and objective.

Overreliance on technology could hinder learning and skill development in VET

One of the main concerns for interviewees was that overreliance on AI and other advanced technologies would prevent neurodivergent learners from developing the skills that VET was supposed to equip them with, leaving them underprepared for the transition to work. Overreliance could also negatively affect learners’ educational outcomes if learners are not permitted to use the same tools in tests and assessments as they use in their usual learning environment(s), for instance if the national or regional body responsible for assessment comes to a different decision to the VET institution or employer. Interviewees were concerned that overreliance on technologies, and on AI in particular, would come at the expense of critical thinking skills as well as basic skills such as writing and spelling.3 While this is a wider issue affecting all learners, the risks of overreliance on AI could be higher for individuals with learning disabilities, data processing issues or language challenges precisely because it is so powerful as an assistant for these individuals, according to Christopher Patnoe.

This is particularly dangerous, a few interviewees remarked, because critical thinking skills are the precise skills that any learner needs to use AI effectively and safely. Pierre Dillenbourg spoke of the importance of training apprentices so that they can work autonomously. For instance, if an apprentice carpenter always uses a tool to calculate the forces on beams of a house, there is a risk that they do not learn to apply their own assessment and reasoning, steps necessary to verify that the tool has proposed the correct solution.

Where VET learners already struggle with social interactions due to neurodivergences, they may welcome the chance to interact with a machine rather than a human. Yet these interactions may hold them back from developing social skills. Aida Nazari from LuxAI, a social robotics company, spoke of concerns that the use of robots among learners with ASD could ultimately hinder their ability to interact with humans, not only because time interacting with robots is time spent not interacting with humans, but also because interactions with robots could be so comfortable and appealing that learners with ASD could lose the willingness to interact with humans. This concern is why LuxAI is built around a triangular relationship involving a robot, child and an adult teacher or parent.

In the same vein, interviewees warned that overreliance on generative AI chatbots could prevent neurodivergent VET learners from developing social skills that would be useful for them in the workplace. David Voss explained that generative AI chatbots, unlike humans, do not provide a multi-voice view, do not contradict the user, and rarely say they do not know the answer. He outlined his worry that neurodivergent individuals who relied on generative AI chatbots throughout their VET training could enter the workplace unprepared to approach colleagues with questions, and unprepared for nuanced, ambiguous or challenging answers. He explained that while studies have suggested that the use of avatars in classroom settings can improve participation, there is still a legitimate debate about whether intervening to remove challenges such as this for neurodivergent learners could leave them inadequately prepared for life’s challenges later on. A similar point was raised by Ann Kennedy, a tutor in further education, who encourages her computer science students to use AI for website design but insists that human interactions are crucial to understand a client’s needs and to deliver a website that meets those needs. To impart this important principle, she sets an exercise for students consisting of making a website centred around her niche interests so that the students must think outside their own comfort zones and must consider the client’s needs.

Similar questions surround the use of tools to enhance focus, typically aimed at individuals with ADHD. Alisdair Gurling wondered whether these tools, which he uses himself and finds very helpful, could ultimately reduce an individual’s internal capacity to focus if relied on too heavily.

“The wider concern I think going forward is that some people's internal capacities don't grow and not only do they not grow, they might atrophy, they might actually diminish” – Alisdair Gurling, researcher at Wonderlab, Monash University

Concerns about AI-assisted cheating risks constrain the legitimate use of assistive technologies

Related to the topic of overreliance are concerns about generative AI chatbots undermining academic integrity by facilitating cheating and plagiarism. These concerns were raised by many interviewees, who spoke of the unresolved question within VET and the wider education system about how to fairly assess students when these tools are so readily available. While this has implications for all learners, a handful of interviewees noted the risk that efforts to address AI-assisted cheating and plagiarism could bring suspicion upon neurodivergent learners, particularly those with dyslexia, who depend on assistive technologies for writing.

A few interviewees referred to the tool, Grammarly, to describe the challenge. Grammarly is a longstanding writing assistance tool, which has recently embraced generative AI, meaning that the tool can generate suggested text on top of its original features focussed on spelling and grammar. Helen Nicholson-Benn explained that while many users would consider this a significant enhancement, it also increases the complexity of using Grammarly in education because of widespread concerns about AI-assisted cheating and plagiarism. David Voss pointed to instances where assignments written with the aid of such tools have been flagged by plagiarism detection platforms and the users penalised. Interviewees described uncertainty among users of this and similar tools about whether use was permitted under their institutions’ rules and anxieties that they could be unfairly accused of cheating.

Developments such as this add to a more general uncertainty regarding the use of assistive technologies in assessment. Rohan Slaughter described how, even for the same learner, decisions regarding the use of assistive technologies in assessments can vary between subject and between exam board. In a VET context, it is also possible that an examination board comes to a different decision to the VET institution or employer. Not only does this incur an administrative burden for examination officers and teachers, but difficulties interpreting the rules can discourage learners from seeking support. Rohan Slaughter noted that better guidance for examination boards is emerging on how to make assessments more accessible and on what principles to follow in authorising assistive technologies. Generally, authorisation depends on what is being assessed. For example, a learner with dyslexia may be authorised to use a writing assistance tool in an assessment testing technical knowledge but not in an assessment testing literacy skills.

The socio‑emotional risks associated with generative AI chatbots and robots are not yet fully understood

Greater interaction with AI and other advanced technologies could affect mental health and socio‑emotional development in ways that are not yet fully understood. A recent OECD report on the potential impact of AI on equity and inclusion in education (Varsik and Vosberg, 2024[43]) highlights the potential negative implications of AI for socio‑emotional learning, such as a reduction in human interaction, which can erode students’ sociability, sense of trust and empathy, and lead to loneliness and isolation. In the most alarming cases, recent newspaper reports (OECD[51]) have tied use of generative AI chatbots for psychological support to incidents of violence, suicide and psychosis.4 Interviewees expressed concern about the potential socio‑emotional implications for neurodivergent VET learners.

Interviewees suggested that some neurodivergent learners could be overly trusting in generative AI chatbots, overlooking “hallucinations” and ending up more exposed to harmful content. David Voss and Eleni Damianidou identified learners with ASD as one group that could be overly trusting of AI outputs. However, Eleni Damianidou was careful to note that this risk also applied to interactions with other humans – in particular, that trust can lead students with ASD to fall victim to classmates who mislead and have malicious intentions – and that the risks of AI should be evaluated within this context when compared with the benefits.

“There's a risk that certain types of autistic learners could take what they're given a bit too, not literally, but truthfully. They really believe that it's the correct information because it's been given to them from an AI. Maybe that's because an AI is a little bit more interactive than a Google search” – David Voss, who works on digital learning in higher education

Interviewees also highlighted possible harms for mental health if neurodivergent learners form emotional attachments to humanoid robots or generative AI chatbots, believing they possess genuine emotional understanding. For some neurodivergent (and neurotypical) learners, conversations with chatbots may allow them to temporarily retreat from a stressful world and unload worries and frustrations as they navigate adolescence and the transition to adulthood. The danger for potentially harmful interactions emerges, according to Elisabetta Bertola, because these chatbots have not been designed to protect people in these circumstances and to take into account various vulnerabilities, neurodivergences, disabilities or needs. Additionally, if interactions with digital interfaces displace interactions with other humans, the learner will become further isolated. Another danger, according to Lorenzo Desideri, is that if a humanoid robot suddenly malfunctions or powers down, or if a subscription to a generative AI chatbot lapses, the learner may experience feelings of abandonment or betrayal. As AI advances and as the boundary between human and artificial intelligence blurs, these dynamics become more difficult to navigate.

Exclusion can be an unintended consequence of these technologies

Although interviewees generally spoke about the potential of AI and other advanced technologies to support the inclusion of neurodivergent learners in VET, some interviewees noted that they could undermine this principle in unintended ways, widening the gaps that already exist. For instance, exclusion can result from discrimination based on bias (discussed earlier in this chapter) and can also result where cost, connectivity or language act as a barrier (discussed in Chapter 3). In VET, such unintended exclusion can also stem from unequal access to digital infrastructure across training centres and workplaces, reinforcing regional and sectoral divides that already limit participation of disadvantaged learners. Michael Fembek spoke of the risks of exclusion associated with telepresence robots for those that cannot physically attend school (due to anxiety or other health needs for example). The robot gives the learner a presence in the classroom and the means to ask the teacher questions and to see other classmates. While this is clearly better than being excluded from school entirely, Michael Fembek made the point that telepresence was still inferior to physical presence when it came to feeling that one belongs in the classroom community. The concern is that these tools, notwithstanding their advantages, could slow learners’ reintegration into the classroom and work against the objective of adapting schools to be truly inclusive of all needs.

Among the many other risks highlighted by interviewees were the environmental impact, implications for intellectual property rights (of teachers but also in relation to the content used to train the models), the potential for AI to displace labour (affecting teachers but also the idea that VET learners will be transitioning into a disrupted labour market).

Notes

← 1.  Similar risks were highlighted in a recent OECD report on the potential impact of AI on equity and inclusion (Varsik and Vosberg, 2024[43]), which noted that during COVID‑19, Human Rights Watch found that many education technology products used data practices that compromised children’s rights, collecting detailed personal information, including location, activities, family information and socio‑economic status.

← 2.  A broader point, made by Alisdair Gurling, is that because these tools generate content based on pre‑existing information and often without clear attribution, they could endanger the entire process of knowledge production.

← 3.  Pierre Dillenbourg likened this to how reliance on GPS has diminished students’ navigation skills and how calculators have diminished their ability for mental calculations.

← 4.  For instance, OpenAI has acknowledged ChatGPT’s shortcomings in interacting with people in serious mental and emotional distress and has said it is working to strengthen safeguards (including referring people to professional help) and to block harmful content more effectively (OpenAI, 2025[62]).