Are Robots Ready to Deliver Autism Inclusion?: A Critical Review

Following recommendations from feminist methodologies [54], we position ourselves to contextualize our study design, methodologies, data collection process, results, and analysis. The research team is neurodiverse and includes women and men-identifying researchers who have a strong interest in improving inclusivity in Human-Centered Computing studies.

The authors of this paper believe in community-oriented design practices such as participatory design [55] and design justice [60] to address power imbalances in traditional HCI research practices between participants and researchers. We understand that certain methodologies may require adjustments based on the access needs of the participant communities [83].

Due to the international scope of our literature review, we focused on gender and age solely for our intersectional analysis. The first author of this paper has lived in five countries, and recognizes the limitations of the US-centric perceptions of race, religion, and other identities which may not translate to other cultures. The exclusion of other identities is not reflective of their perceived importance by the authors. We acknowledge the difficulties researchers may have with recruiting diverse participants for their study, and want to note that we are not criticizing the individual authors of any of the works critiqued in this paper.

Our corpus selection was carried out in multiple iterations. We obtained full papers, short papers, survey papers, and editorials from the fifty most relevant results using the criteria detailed in Sections 2.2.2 and 2.2.3 from the Google Scholar search engine, the ACM Digital Library, IEEE Xplore, Springer, and Science Robotics. Our search criteria specified the publication years ranging from 2016 to 2022. The distribution of the publication years of our main corpus are detailed in Figure 1. For searches that revealed more than 50 results, we filtered our results based on relevancy and reviewed the first 50 papers.

We selected 142 papers ² for our final analysis. Table 1 summarizes the venues and papers selected for our analysis.

These works were selected through the automated analysis of all the references from our main corpus. We built a web crawler that obtains the Digital Object Identifier (DOI) of each paper as a unique ID. Using these DOIs, we found matching works cited in each paper and obtained a total of 6,611 papers. In order to gain a better understanding of trends in the most frequently referenced works in HRI research for autism, we analyzed the publication years and fields of all of these papers. We selected the 100 works most frequently referenced by our main corpus and conducted a deductive qualitative thematic analysis to identify the model used to define autism in their work.

Our data analysis consisted of two parts: a manual coding of the corpora and an automated analysis of the metadata obtained from the Semantic Scholar API.

Our foundational knowledge and understanding of autism has been shaped by researchers who adapted dehumanizing language through the “Theory of Mind” [8, 53]. They expanded on prior work analyzing whether chimpanzees have a theory of mind and concluded that autistic children lack the “quintessential” human trait [65]. This deficit-based understanding is central to the medical model, which focuses mainly on the causation or cure of autism as a disorder [65]. This dehumanization of autistic end-users is replicated in human-robot interactions. Traditionally, robotics research has pathologized the communication behaviors preferred by autistic people and introduced technologies that encourage them to adopt neurotypical social norms [15, 34, 126, 129]. Figure 2 illustrates how 15 of the studies in our corpus used robots that looked like animals to play a “mentor” role focused on providing social skills training to autistic end-users. By placing animalistic robots in this role, such studies inadvertently perpetuate the findings of the dehumanizing foundational work in autism research that concluded certain animals have stronger social skills and are thus “more human” than autistic children [8, 53]. Additionally, the majority of the 76 studies utilizing anthropomorphic and humanoid robots had the robots adapt a mentor role with autistic users, which perpetuates the belief that autistic people are deficient in their humanity and robots can help them move toward “humanness” [129].

A deficit-based understanding of autism has been pervasive even in computer science and robotics research [22, 34, 63, 68, 78, 98, 129, 132] even though prior work has suggested it contributes to the ostracism and discrimination autistic individuals face in our society [5, 65, 97]. The ostracism of autistic people in studies that use this definition can be quite explicit. For example, one paper studied the use of robots to:

Such language is rooted in ableism as historically the word “abnormal” has been used to describe things that are unhealthy or unnatural  [27, 61, 88]. Yet, an automated analysis uncovered words such as ‘typically develop*’ and ‘abnormal’ which explicitly posit non-autistic people as the norm and their autistic peers as a deviation from the norm were prevalent in 27 papers, while deficit-based language appeared in 11 papers as shown in Appendix 1a ⁶. This led us to uncover the prevalence of the medical model that applies a deficit-based understanding of autism and promotes diagnosis and treatment through therapy and skills training [66], as shown in Figures 3 and 4.

Essentialism, in the context of Autism, is the idea that people with autism have common characteristics that are "inherent, innate and unchanging" [101]. Historically, the extreme male brain theory has identified autism as an extreme presentation of a “normal” male profile [9]. This theory has been criticized for its gender essentialism and pervasive influence [73]. For example, the belief that autism is a “masculine” disorder has led to the underrepresentation of women and girls in autism research; due to misconceptions among researchers and the general population, people are more likely to recognize autism in males [1, 128]. Foundational research also infantilizes autistic adults by characterizing them as “children” [17], and viewing autistic bodies as being “frozen” in childhood [16]. This infantilization has led to widespread misbeliefs that autistic people are dependent and lack autonomy [16]. Even in HCI and HRI, prior work on autism research has mainly focused on children [22, 31, 46, 62, 96, 114, 115]. Additionally, the concept that an autistic person’s “mental age” is different from their physical age is rooted in ableism and specifically pervasive in autism research [39].

Many foundational works in autism research have applied social deficit theories; instead of studying the unique social behaviors preferred by autistic individuals, established research in developmental psychology highlights autistic people’s “deficit” of neurotypical social norms such as the ability to “mind read” and make “appropriate” eye contact [8, 53, 65]. This phenomenon, known as Neuronormativity, promotes neurotypical behaviors as the “norm”, thereby introducing a power imbalance in social interactions between individuals of different neurotypes [59].

As a solution to this "problem", researchers have attempted to use robots to train autistic people how to imitate these neurotypical behaviors. Systematic literature reviews and editorial works in HRI research have called for the usage of robots to provide therapy and skills training to autistic end-users [12, 22, 90, 102, 105]. This also includes the use of anthropomorphic and humanoid robots which are designed to be more human-like [91]. This is concerning as there is research suggesting such robotic systems may not be useful at all [12] and their designs directly contradict the known preferences of autistic children [38, 95]. There are additional concerns regarding the context of foundational research in autism which humanized a non-human entity over autistic children [8], as it makes assumptions about autistic people’s skills deficits and promotes the belief that non-human entities may be more socially skilled than autistic people [129].

Additionally, autistic people have historically faced a lack of representation in both the research teams and the study participants. Critical Autism Studies (CAS) encourages broadening the participation of autistic scholars in autism research [134]. According to CAS, a lack of autistic authorship may impact the accuracy of such work and may fail to address power dynamics in the medical model understanding of autism [134]. While participatory design encouraged community involvement in the design process, researchers have highlighted such approaches may fail to focus on under-served communities if they do not take historical context into consideration, neglect community access and perceptions of the materials and activities, and cause unintentional harm [55]. Many disabled scholars have discussed the harms of excluding under-served disabled populations from the design process [136] and proposed alternative approaches [139]. Thus, in our work, we analyzed existing power imbalances in the studies on three levels: 1) autistic and non-autistic people in general, 2) the researcher and participants in the studies, 3) and the resulting human-robot interaction.

As clinical collaborations are pervasive in HRI research for autism, with 90.5% of the studies applying a medical model approach in their work, it is important to critically investigate them. Although prior work uncovered that autistic children prefer engaging with non-anthropomorphic robots [38, 95], the majority of HRI studies did not utilize these robots, as shown in Figure 2. Some researchers have justified this clear gap between autistic users’ needs and the current research practices by arguing human-like systems may have a greater potential for generalizing skills [38].This is concerning as it shows the extent to which autistic people are excluded from research that focuses on them. Thus, more work is needed to determine autistic people’s preferences for and the effectiveness of anthropomorphic robots [38, 95]. However, it is important to note here that prior work has shown the clinical use of robots may be ineffective or even counterproductive [12, 38]. In fact, some prior work demonstrated a decrease in social-communication skills in autistic participants who were communicating with robots [38]. These results suggest that more qualitative and quantitative research is needed to determine how these interactions impact autistic people over time, and the perspective of autistic people about the clinical use of robots.

In our corpus, approximately 90% of HRI research also excluded autistic people as key stakeholders, with nearly 20% of the studies not taking their perspective(s) into consideration at all throughout the research process, despite them being the target end-users of such systems. In order to avoid such misrepresentations and adhere to the more inclusive design practices [55, 118], it is imperative for research focusing on autistic people to center them and consider perspectives beyond the medical model. This may highlight a broader pattern of exclusion of neurodivergent users in design research, as prior work has shown people with ADHD experience similar exclusion contradicting the principals of both Interaction Design and User-Centered Design [118], while the prevalence of the medical model has also caused HCI video games to fail to support the self-determination of neurodivergent populations [116]. Self-determination is an important objective for autistic people in particular, as their historical infantilization has caused misconceptions about their lack of autonomy to be pervasive in our society [16, 119], with even HRI research explicitly making such assumptions in published work [77]. We encourage researchers who employ robots to provide "support" and other forms of assistance to autistic people to consider the ethical aspects of that work.

The purpose of these questions is to help researchers avoid the most common harmful stereotypes and historical misrepresentations uncovered in our analysis. HRI researchers who focus on autism may consider this non-exhaustive list of questions to assess the inclusiveness of their work:

The clinical use of robots for autistic users mainly focuses on providing ’treatment’ in the form of therapy or skills training, as shown in Figure 4. Even if we assume robots were effective in achieving these objectives, critically examining the broader impact they have on our society uncovers several ethical concerns. First, these systems promote the idea that autistic people are "deficient" in their humanness, and that robots can teach them how to be more human-like [13, 66, 129]. This echoes foundational work that has questioned the humanity of autistic people, and proposed non-human entities such as animals may be more human than them [65]. We see these views being inadvertently replicated in the assigned human and robot roles in HRI studies, as shown in Figure 2, as there is a notable correlation between the humanness of robots and their likehood of playing a mentor role. Second, such work may also promote age and gender-based essentialism, as our work uncovered adults and gender minorities are either underrepresented or misrepresented in the majority of these studies. For example, one study attempting to study the intersection of age and neurodivergence included only neurotypical adults, and compared them to autistic children and neurotypical children [52]. Such study designs contribute to the widescale erasure of autistic adults in our society [16], and fail to take their unique needs into consideration [19]. This widens our knowledge gap in understanding the role of intersectional identities such as age and gender on users’ needs. As well, it leads to misrepresentations explicitly linking certain traits to autism, which may only be found within a particular group of autistic people, such as aggression which is uncommon in autistic gender minorities [48, 108]. Finally, as shown in Figure 5, there is a lack of diversity in the literature examined by HRI research, as the majority of works referenced were published in psychology, computer science, or medicine focused venues. As technology researchers are more likely to cite the works of other computer scientists even in human-centered research, it is imperative to shift to more inclusive research practices. This may lead to a positive domino effect in other areas of technological research.

We also uncovered examples of more inclusive work which may be indicative of an imminent positive shift in HRI research for autism. For example, 5 papers had a representation of gender minorities larger than their actual population [71, 110, 111, 137], out of which one study even took gender into consideration in the study design [84]. Additionally, as shown in Figure 3, 9.1% explored perspectives beyond the medical model in their work, such as robots used as educational companions for autistic children in rural schools, which were well-received by autistic children [21]. These works may be reflective of changing attitudes within the research community, which may not be reflected in the existing research due to systemic and institutional barriers.

Funding. There is a known misalignment between the funding distribution for autism research and community priorities [45]. A 2019 investigation of US federal funding for autism research revealed that the majority of the grants awarded for research focused on the biological aspects, treatment, and interventions for autism and that these funding patterns remained constant  [56]. Other studies have revealed similar funding priorities in nations including Canada, Aotearoa (New Zealand), and Australia, though a shift has been observed in Australia in the past decade [35, 43, 74]. Research suggests that funding can impact what researchers focus on and the methods they use [122], thus potentially increasing the number of opportunities that encourage more inclusive methodologies may improve the way autism research is conducted.

Publication Process. The publication processes of journals and conferences impact the type of research published. In recent years, journals such as "Autism" have prioritized inclusion in their leadership, editorial board, and pool of reviewers to ensure their processes are led by community members [45]. They have also highlighted how reconsidering the focus of the work published may contribute positively to changing the research landscape [45]. In contrast, two editorials published in the years and journals fitting the search criteria for our literature review focused exclusively on medical model papers, and did not include any researchers or editors who apply other approaches to defining autism in their work [57, 135], which may influence their publishing standards.

Autistic Autism Researchers Dwyer et. al published a paper in 2021 that focused on the concerns and insights from roundtables with autistic researchers who focus on autism [41]. They found that autistic autism researchers may face many tensions between the broader research community, institutions, and participants in the research process. Some of these tensions are related to language, as researchers may grapple with the pervasiveness of offensive language used by researchers, and trying not to make autistic participants or their families feel attacked for using such terminology [41]. Tensions may also arise due to differing perspectives and research agendas. Autistic researchers may feel torn between their sense of responsibility to their community and adhering to institutional requirements that may not align with the community needs (especially concerning the use of the medical model) [41]. Additionally, they may face infighting between members of the community with different perspectives [41], which may be difficult to incorporate [92]. Furthermore, researchers, in general, may struggle to meet the needs of diverse autistic participants as they are difficult to generalize and not addressed by best practices [24].

Community Insights. In a round-table discussion on their experiences with conducting autism research, autistic researchers generated recommendations to improve the challenges and tensions they face [41]. Several of them echoed the sentiments discussed in earlier sections of our paper: broadening the inclusion of autistic people (particularly those with intersectional identities), switching the focus of research to more strengths-based approaches, refraining from assuming autistic people need to be ’fixed’, and recognizing the full diversity of autism (including understudied and marginalized populations) [41]. Autistic researchers also called for reforming academia to make institutions and practices more inclusive, and encouraged their colleagues to treat autistic people as peers and equals, be open and receptive to learning, and have empathy for and validate the diverse experiences of autistic people [41].

While systemic changes such as addressing the misalignment between funding priorities and community needs, and diversifying reviewers and editorial boards to include more community members are important, making changes to traditional research practices may also help HRI research shift toward autism inclusion. These include diversifying research collaborations, foundational works considered, participant demographics, and objectives for research directions to explore perspectives promoting a more neurodiverse understanding of autism.

Diversifying Perspectives. Our work uncovered HRI research for autism continues exploring ways to diagnose or treat autism as shown in Figures 3 and 4 even though prior work has linked this perspective to anti-autistic ableism [5, 65, 97]. Despite the diversity of alternative viewpoints in autism research that emerged in other fields over the past decade, the majority of the HRI research in our corpus continued referencing works published in the 1990s and 2000s in the medical and psychology fields, which mainly apply a deficit-based medical model understanding of autism, as shown in Figures 7, 6, and 5. By including newer research published in other fields such as critical autism studies in their foundational work, HRI researchers who focus on autism may move toward more inclusive research directions that focus on promoting neurodiversity. By reviewing, reflecting on, and incorporating the diverse perspectives in autism research, particularly those of autistic researchers themselves, HRI researchers can move toward more inclusive research directions and practices. For example, future work may focus on resolving the double empathy problem by promoting conversational equality between different neurotypes instead of placing the burden solely on neurodivergent people to adapt to unique communication styles  [87, 97].

Diversifying Collaborations. We found clinical collaborations dominate HRI research for autism in lieu of more community-based research collaborations. The majority of the papers did not include autistic people in the design process (90%, n=109) and pathologized their communication behaviors (93.75%, n=135) as shown in Figures 11 and  3. To address this power imbalance, HRI researchers who conduct this work should consider giving autistic people decision-making power in the design and objectives of the research study as prior work has found this may help lower anti-autistic ableism [17]. As researchers’ identities may influence how comfortable the participants feel around them, it is important to consider more community-oriented approaches to ensure the results obtained are accurate and reflective of the users’ needs [55, 76].

Diversifying Demographics. Our work also uncovered of the 144 papers reviewed, only 55 and 139 reported the participants’ genders and ages respectively. Out of these papers, the majority overrepresented autistic participants that identified as male (78.18%, n=43), or only included participants under the age of 18 (82.64%, n=119) as shown in Figures 9 and 8. Examples of language promoting essentialist gender and age-based stereotypes such as linking a lack of agency [77] and aggression to autism were found even in highly cited works [108]. In order to avoid promoting essentialist stereotypes in their work, HRI researchers who focus on autism should prioritize intersectionality in their study design and recruitment and provide more information on their participants’ demographics to help contextualize their results. Although other marginalized identities such as race and class go beyond the scope of our paper, we encourage researchers to also take them into consideration by familiarizing themselves with other works such as Annamma’s DisCrit [5]. We note here that diversifying demographics and collaborations should also include respecting identity disclosure decisions and recognizing the ethical limitations of diversifying demographics.

Humanizing Participants. Finally, we found that HRI research reproduces dehumanizing foundational work through the robot types and roles in user-interactions. For example, Figure 2 shows how human and animal-like robots continue to be employed in mentor roles to provide autistic end-users with skills training and therapy for selected social "deficits". This reproduces dehumanizing foundational work in autism research [8] while promoting the ableist notion that robots may help autistic people move toward "humanness" [129]. HRI researchers may address this by obtaining input from autistic end-users on the traditional user and robot roles, especially with respect to the robot’s anthropomorphism to avoid inadvertently dehumanizing autistic people. Exploring alternative research directions such as focusing on the double empathy problem [87] will also help promote more equitable user interactions.

Using Inclusive Language & Positionality Statements. Additionally, future work should reconsider language that promotes such stereotypes or anti-autistic ableism in general. We recommend referring to the works of Bottema-Beutel et al. to learn more about language that may be considered ableist against autistic people [18], the works of Stevenson et al. [119] and Krahn & Fenton [73] to learn about essentialist stereotypes, and autism research focusing on intersectionality such as the works of Kanfiszer et al. [64] for examples of more inclusive and intersectional research practices. While positionality statements are largely absent in HRI research, their inclusion can help provide information on the researchers’ various identities and how they may have shaped the data collection and analysis.