Artificial intelligence and assistive technology: risks, rewards, challenges, and opportunities

Assistive technology (AT) has historically been at the forefront of emerging technology development. There are countless examples of technologies which were present in the world of disability long before they became mainstream, from audiobooks to text-to-speech software, among other things, which were initially developed for the disabled population (Ovide, 2021). The same is true for many environmental control systems, word prediction software, and virtual assistants. The opposite is also true – as mainstream technologies become more accessible, and principles of inclusive design more prevalent, these technologies become critical for disabled people, among others, to participate in their communities and control their environments.

The UN Special Rapporteur on the rights of persons with disabilities highlighted Artificial Intelligence (AI) in his remarks to the UN General Assembly (2021), noting “unprecedented and rapid growth in the use by States of artificial intelligence, automated decision-making and machine-learning technologies.” (p. 1), invoking the need to initiate and inform a debate in this area (OHCHR, 2001). However, he also warned of the dangers of its being (sometimes unintentionally) marginalizing people with disability. The Special Rapporteur highlights that once the rights of persons with disabilities are placed at the center of this debate, and risks addressed, then the practical benefits of artificial intelligence might be realized (OHCHR, 2001).

Artificial intelligence can be broadly understood as the capability of a computer system to perform tasks typically requiring human-level intelligence such as speech recognition, visual perception, and object classification. There are different applications of AI that we see presented to the market with perhaps the most notable being generative AI. Generative AI, an AI system that generates text, images, and audio (amongst other outputs) is just one of the various permutations of AI; others include embodied AI (drones, robots, autonomous vehicles), human-in-the-loop AI (medical diagnosis, facial recognition), and transfer learning AI (natural language processing, translation).

Many forms of AT are already digitally “connected” and each of the above forms of AI has potential applicability for AT. AI is involved in the development of assistive robots, autonomous wheelchairs, and guidance systems for people with visual impairments (Kahraman & Turhan, 2021; Kim et al., 2023; Lim et al., 2022). Facial recognition technology has been explored for helping people who have difficulty with social interaction (Adiani et al., 2023). Some AI technologies are in widespread use already. Speech recognition, for example, doesn’t feel like emerging technology to many of us, but AI systems are revolutionizing our traditional speech recognition systems and their application in AT. Smart home systems, now largely powered by AI, are shown to have a positive impact on independent living skills, quality of life, and psychological well-being (Landuran et al., 2022; Seidel et al., 2022).

However, these emerging technologies are not without challenges. While AI may be fast, reliable and able to manage huge amounts of data, this usually involves learning from a data “training set,” from which algorithms (rules for problem solving) can be developed – such as identifying the characteristics of potentially good workers from existing worker profiles and performance records. However, if “training sets” don’t include diverse groups then their opportunities may be progressively diminished, as “the usual suspects” are selected and constitute an ever grouping “training set.”

The data that has been assembled for use in AI training comes from a variety of sources and collections, reliant on humans to achieve the level of fine-tuning and validation necessary to be useful. Consequently, it can be biased – based on subjective understanding of the data set, rather than objective criteria. Additionally, the transparency of the labeling process, the sparsity of data set characterization, unverified labels, and lack of diversity in the data collection process all lends credible challenge to ensuring an unbiased or fair data set is used to enable properly trained AI tools (Daneshjou et al., 2021; Kamikubo et al., 2022; Khan & Hanna, 2022). Evidence suggests a lack of diversity in AI training datasets can result in technologies which do not reflect the needs of a diversity of human experiences (Nakamura, 2019). AT users need to be part of the training sets for AI to ensure that their experiences, talents and requirements are part of the evolving learning of AI systems. Perhaps, and ever more increasingly, this further illustrates the crucial aspect of keeping the human in the loop when it comes to artificial intelligence in the health domain (Bakken, 2023).

In addition, many of the technologies themselves are still in development and have yet to address some significant barriers. Speech recognition systems continue to struggle to understand people with dysarthria (Jaddoh et al., 2022). Smart wheelchairs, although we are seeing components like obstacle detection systems, are still not completely accurately detecting descending stairs – a critical safety feature (Utaminingrum et al., 2022). Despite some long-standing AI applications in AT, the reality remains that AI is emerging technology – and may always be so, in terms of our understanding of how to best apply its potential and understand its impacts.

Considering the health and care workforce from a global perspective, there is a strong and compelling case to be made to reskill the existing workforce in the application of digital products and services. Equally this must include educating the next generations of health and care workforce in the use of digital and artificial intelligence tools as part of a broader framework aimed at supporting its development and sustainability (WHO Regional Office for Europe, 2002).

As often happens with emerging technologies, we are more advanced in the technologies themselves than we are in the regulations, standards, and policies which govern their use. To that end, the European Disability Forum, the umbrella organization for organizations of persons with disabilities in Europe, has highlighted concerns with proposed EU legislation not sufficiently addressing and protecting the rights of persons with disabilities (European Disability Forum, 2023). Among other concerns, they highlight disproportionate impacts on persons with disabilities when AI is used for access to public and private services. They also highlight issues of privacy and data protection for vulnerable people, including those with intellectual or psychosocial disabilities.

While the academic literature in the fields of AT and disability is limited in terms of discussion of ethical issues related to AI (Lillywhite & Wolbring, 2019), they are being addressed and explored outside the field. Organizations like the DAIR Institute (https://www.dair-institute.org) have engaged in research around mitigation of potential harms caused by AI while others, like the Center for Human-Compatible Artificial Intelligence at UC Berkeley, have proposed taxonomies for determining the social-scale risks of AI to engage a broader sociopolitical dialogue around ethics and regulation (Critch & Russell, 2023).

Technology is never neutral or equally available, and yet AT has been shown to be a crucial driver in realizing the UNCRPD (Smith et al., 2022). If the benefits of AI for AT users are tovbe shared equitably then industry, state and private providers and civil society must work deliberately and together toward a Just Digital transition to achieve the Sustainable Development Goals (O’Sullivan et al., 2021). Addressing the challenges with emerging technologies can only be meaningfully accomplished if AT users, and those delivering health and care services, are engaged in the development process. It is imperative that AT users be engaged not only in the development of the technologies themselves but also in conversations about their ethical application, and in establishing the regulations, standards, and policies which govern them.

The Assistive Technology Journal is proud to be at the leading edge of AT research, including the role of AI and other emerging technologies. We encourage researchers to consider contributing to our Product Development and Evaluation section, or to explore relevant issues in our Ethical Considerations, Policy Studies, or Commentary sections. Researchers exploring the application of emerging technologies as AT may also consider a submission to our Research Papers section. The Assistive Technology Journal encourages meaningful involvement of AT users in all forms of AT research.