Unlock Life with a Chat(GPT): Integrating Conversational AI with Large Language Models into Everyday Lives of Autistic Individuals

Autistic individuals often draw on insights from their supportive networks to develop self-help life strategies ranging from everyday chores to social activities. However, human resources may not always be immediately available. Recently emerging conversational agents (CAs) that leverage large language models (LLMs) have the potential to serve as powerful information-seeking tools, facilitating autistic individuals to tackle daily concerns independently. This study explored the opportunities and challenges of LLM-driven CAs in empowering autistic individuals through focus group interviews and workshops (N=14). We found that autistic individuals expected LLM-driven CAs to offer a non-judgmental space, encouraging them to approach day-to-day issues proactively. However, they raised issues regarding critically digesting the CA responses and disclosing their autistic characteristics. Based on these findings, we propose approaches that place autistic individuals at the center of shaping the meaning and role of LLM-driven CAs in their lives, while preserving their unique needs and characteristics.


INTRODUCTION
Conversational agents (CAs) that leverage large language models (LLMs), exemplifed by ChatGPT, are gaining attention as powerful information-seeking tools for autistic individuals 1 [48].By simply typing a query to LLM-driven CAs, autistic individuals can easily access vast information anytime and anywhere whenever they face unexpected daily challenges [58,83].Moreover, it can deliver personalized responses that accommodate specifc traits and intentions of autistic users through its advanced context analysis [21,95].With such capabilities, LLM-driven CAs hold promise for autistic individuals to navigate everyday life independently, ranging from everyday tasks and self-care to social interactions [64,75].
However, coping with day-to-day activities through conversations with LLM might have negative impacts on autistic individuals.Assessing the appropriateness of LLM's advice on social communication would be challenging for autistic users who often struggle to cope with implicit social norms [4].While some studies suggest that LLM-driven CAs can facilitate mental wellbeing in autistic populations [48], their lack of sensitivity to subtle emotional cues and risky signals might lead to harmful outcomes [26,52].Furthermore, previous research has reported that LLMs, in conversations with users with disabilities, portrayed them as people needing help or wanting to be "fxed" [33].LLMs' responses with biased perspectives might hurt autistic users or lead them to make decisions that do not respect themselves.
Given the undeniable wave of LLM-driven CAs rapidly integrating into our lives globally, it is not feasible to simply limit their use for autistic individuals because of the expected drawbacks.Therefore, proactive eforts are needed to augment the benefts and address the disadvantages of this technology to empower the autistic population.However, in the early stages of adoption, the in-depth voices of autistic individuals regarding the impact of incorporating LLM-driven CAs in their daily lives remain largely unexplored.
With this background, we aimed to explore the perceptions of autistic individuals toward LLM-driven CAs and characterize their expectations and challenges in using them as they navigate their everyday lives.We conducted a series of focus group interviews and workshops in which 14 autistic individuals ideated the application of LLM-driven CAs in daily contexts by interacting with ChatGPT.Our fndings indicate that LLM-driven CAs can serve as supportive partners for autistic individuals in independently addressing dayto-day activities and advocating for themselves.Simultaneously, we captured the challenges associated with self-directed technology adoption, over-reliance, and disclosure of autistic identity.Drawing on our fndings, we propose forward-looking recommendations to facilitate the critical engagement of autistic users with LLM-driven CAs with a sense of agency.
The contributions of this work include the following: 1) an empirical study to understand the opportunities that LLM-driven CAs can ofer in facilitating the independent lives of the autistic population from their perspective; 2) the identifcation of challenges arising in adopting LLM-driven CAs as empowerment tools for autistic users; and 3) the proposal of recommendations to foster an inclusive LLM experience for the neurodiverse population.

RELATED WORK 2.1 From CMC to LLM: Leveraging digital technology to empower autistic individuals
Autistic individuals may sometimes face challenges when it comes to organizing, initiating, and generalizing information and actions [16], which can potentially impact their sense of self-determination [49].As a result, many individuals have turned to trusted caregivers, such as parents or family members, for valuable guidance in navigating the numerous daily challenges they face [79].Nonetheless, such human resources may not always be readily accessible for immediate consultation, and signifcant dependence can impose a substantial burden on them [64].A limited support network also restricts autistic individuals from broadening their social relationships or accessing a wide range of viewpoints [12].The advent of the Internet has fostered online support networks beyond the physical environment for autistic individuals.A trustworthy, responsible circle of family, friends, and acquaintances (e.g., neighbors) supported by a commercial social networking service encourages greater social interactions while decreasing the reliance on seeking advice inordinately [46,47].In an ethnographic study focused on an online Minecraft community for autistic children and their families, researchers explored the idea that autistic children could establish and engage in social interaction within a virtual environment, potentially leading to improved ofine social engagements [86].Autistic people have found strangers online are a viable source of support.Computer-mediated communication (CMC) provides autistic individuals with an opportunity to initiate supportive relationships online by reducing social pressure, as explained by Burke et al. [15].Online relationships, however, have been hindered by fear of victimization and misunderstandings caused by unfamiliar CMC-specifc norms [15].Researchers have investigated how crowdsourcing can meet the needs of autistic people by providing them with diverse advice on everyday problems [45].This study found that crowd workers provided quick and immediate responses to autistic requesters, broadening their perspectives.For the efective integration of anonymous online resources, systemic measures are required to ensure the credibility and safety of responses from those outside the autistic community.
Artifcial Intelligence (AI) and Natural Language Processing (NLP) have opened new pathways for autistic individuals to address their daily challenges [32].AI-based CAs have been developed for this population to support or practice a wide range of daily tasks, including establishing a home oral care routine [80], handling school bullying [54], and managing depression [41] and anxiety [78].More recently, studies have explored the potential of leveraging AI for open-domain dialogue, empowering autistic individuals to engage proactively in conversations about their everyday concerns.Cha et al. demonstrated that an AI-based VCA can serve as a conversational partner for autistic adolescents, meeting their various daily needs, including sharing interests, managing daily routines, and practicing communication skills [17].Another co-design study revealed that autistic college students are interested in the multifaceted use of AI-driven CAs, encompassing aspects such as academics, health, and social interactions [17].Despite the identifed needs, AI-driven CAs up to this point have fallen short of addressing the unique and diverse needs of the autistic people, largely owing to their constraints of pre-defned conversational fows [66].
With the ability to engage in open-ended conversations similar to human interactions, newly developed LLMs are now proving to be useful tools for autistic individuals in the search for information.Combined with large-scale pre-training, users can easily access a wealth of organized information through the LLM [58,83].LLM also provides relevant and coherent responses by considering the surrounding context of a conversation or query [5,14,74,87,111].In particular, in the form of a chatbot, LLMs use an instructiontuning process to learn what users want, aligning the model with their specifc needs [14,106].Owing to their superior accessibility and adaptability [114], LLM-driven CAs hold potential as a new medium to empower autistic individuals to tackle everyday issues.While there are anecdotal pieces of evidence-such as articles [23,48], Reddit communities [84], and blogs [92]-documenting the experiences and opinions of autistic individuals using LLM-driven CAs, there is still a scarcity of in-depth research examining the tangible benefts these tools could bring to the lives of autistic people.

Inclusive LLM experiences for autistic individuals
Propelled by research examining AI bias related to gender and race [9,34,42,63,72], recent years have seen growing attention to AI's inclusivity toward identity-based minority groups, such as people with disabilities [38-40, 51, 96, 99, 101, 109] and LGBTQ communities [27,82,103].The increasing interest in AI inclusivity aligns with the growing scrutiny of LLMs, particularly regarding interactions with minority users.Owing to the extensive humangenerated text on which LLMs are trained, there is an inherent risk of perpetuating conversations that are unintended or unanticipated by chatbot designers [5].One of the most active areas of research is the harmful dialogue produced by LLMs, which often originates from biased data or inadequate data representation [7,10,18,30,35,37,107,116].Several studies have discovered that language models such as GPT-2 and BERT can produce biased or ofensive texts, particularly afecting historically disadvantaged groups [7,28].A focus group study by Gadiraju et al. [33] explored interactions between people with disabilities and LLM-driven CAs, specifcally discussing disability-related topics.The study found that LLMdriven CAs frequently portray people with disabilities as either sorrowful beings in need of assistance from able-bodied individuals or as marginalized fgures with compromised autonomy.Just as people with disabilities feel that LLMs violate their autonomy and normalize their exclusion, there is a potential risk that autistic individuals could either be harmed by discriminatory perspectives in LLM-based dialogues or may not recognize such biases.Another critical issue surrounding inclusivity is the potential of LLMs to exert subtle psychological and social efects on users, even without explicit toxicity.Notably, LLMs can produce hallucinations that sound credible but are factually incorrect [30], posing particular risks in scenarios, such as seeking medical information without expert consultation [89].For autistic users who fnd it challenging to navigate subtle social norms, diferentiating between these hallucinations and reliable information can be particularly challenging, leading to misunderstandings and further social conficts.Previous research has also highlighted the detrimental efects of overreliance on LLMs in educational and workplace settings [25,36], suggesting that it can impede users' problem-solving abilities [55].This habitual acceptance of LLM responses could make it difcult for autistic users to critically engage with the information provided, thereby compromising their ability to make informed decisions [102].
Despite the aforementioned potential risks that LLM could pose in the daily lives of autistic individuals, there has been limited research on the unique challenges and experiences that this group might encounter.Numerous assistive technologies have addressed this issue by providing specialized platforms or safeguards specifically designed for autistic users, such as exclusive online dating services [2,43] and moderated online communities for young autistic adults [1].However, in the feld of LLM, the advent of universally accessible and increasingly popular services such as ChatGPT has blurred these boundaries.Many autistic adults are already using these general platforms, thereby making the previous "separate but safe" approach less relevant.In light of this, rather than developing exclusive platforms or safety AI flters for autistic users, we aimed to explore the unique opportunities and challenges they experience on mainstream LLM platforms.

STUDY DESIGN
This study aims to explore the opportunities and challenges of using LLM-driven CAs in the everyday lives of autistic individuals.To this end, we conducted focus group interviews to examine the participants' perceptions of LLM-driven CAs.We then held individual workshops in which participants devised their own strategies for using LLM-driven CAs to address their everyday needs and concerns.In this study, we selected ChatGPT as the representative LLM-driven CA platform due to its high accessibility and current popularity.

Participants
Participants were recruited based on the following criteria: 1) diagnosis of Autism Spectrum Disorder, 2) ability to communicate and comprehend the interview questions provided in the study, 3) willingness to use LLM-driven CAs, and 4) ability to use digital devices such as computers or smartphones.In the fyers, we articulated that the study aimed to explore ways of integrating ChatGPT into the daily lives of autistic individuals.Additionally, we provided details about the participation period, the procedure of the study, and the activities involved.These fyers were distributed to potential autistic participants who met our criteria through the Center for Autism Education and Support in Korea, where one of the authors is afliated.As a result, a total of 14 participants responded that they were interested in the study and provided written informed consent.
The participants were aged between 18 and 37 years (mean=23, SD=5.94), and included 11 males and 3 females.The participants' gender balance mirrors the diagnosis distribution reported in previous literature, showing a male-to-female ratio between 4-7.38 and 1 [6,108].Six participants, including three majoring in engineering, had previous experience using ChatGPT.Among them, four participants identifed that they often used it more than three times a week, while two participants identifed that they occasionally used it approximately once a week.None of the participants used the paid version (model GPT-4).The remaining seven participants mentioned that they had heard of ChatGPT, but had no experience using it.All participants were Korean and were compensated with 100,000 KRW (equivalent to $80 USD) for participating in both the focus group interviews and workshops.

Ethical considerations
The study was approved by the Institutional Review Board of a University.In our research, ChatGPT, served as a representative example of LLM-driven CAs, employing a multi-faceted approach that includes fne-tuning, human reviewer guidelines, and user feedback to mitigate the generation of potentially sensitive or harmful content [76].Nevertheless, we recognize that this investigation may present ethical complexities.These include the potential for LLMs to produce harmful or biased responses [19,28], as well as the limited extant research concerning the psychological and social implications of utilizing LLM-driven CAs within the autistic community.To address these ethical considerations, a psychiatrist on our team-serving as the ffth author-leveraged over two decades of experience with young autistic adults to assist in the establishment of the study's procedural guidelines and risk response protocols.This protocol addresses three major issues: 1) targeted recruitment with a autism specialist, 2) tutorial on ethical issues, and 3) oneon-one guidance during ChatGPT use, which we will elaborate on below.
First, during the recruiting phase, we employed a targeted approach to disseminate information about our study.This was achieved through a collaboration with an autism specialist, serving as the fourth author, who holds a doctoral degree in autism education and possesses over a decade of experience in counseling autistic children.This expert extended invitation to participate in the study to individuals within her professional network who did not exhibit substantial emotional or psychological difculties.Secondly, in both the focus group and the workshop settings, we provided participants with a comprehensive tutorial concerning the ethical deployment of LLM-driven CAs.This tutorial was informed by the recent research and guidelines on ethical usage of ChatGPT [105,115].The instructional material encompassed a range of pertinent topics including bias, privacy and security, abuse, authorship, hallucinations, undue trust in the technology, and self-referential (AI-based) monitoring mechanisms.To enhance participants' comprehension of these complex issues, we provided real-world examples illustrating instances where these ethical concerns have manifested and led to adverse outcomes.Lastly, we orchestrated an environment wherein participants could interact with ChatGPT while receiving individualized support from a member of the research team.Our study protocol mandated the continual monitoring of dialogues between participants and ChatGPT, along with solicitation of participants' emotional responses and cognitive refections after each topic of conversation.A researcher should observe that a participant had an adverse experience directly relating to the ethical issues proposed in the tutorial [105,115].The protocol required that the researcher clarify the issue at hand for the participant.Throughout this process, our fourth author, an expert in autism, was actively involved in the identifcation of potential risk factors and the recommendation of tailored countermeasures based on the individual characteristics of each participant.If it was ascertained that a participant was encountering signifcant and sustained negative efects, protocols were in place to terminate the study.However, full compensation was provided, along with an ofer of subsequent support from the clinician co-author.

Focus group interview.
In the initial session of our study, we conducted interviews to capture participants' prior experiences and perceptions of LLM-driven CAs.We created a group environment in which participants could exchange views with their peers and researchers, thus promoting their autonomy [90] and broadening their perspectives [8].Of the 14 participants, seven were divided into three small groups, each consisting of 2-3 members.The remaining seven participants, long-standing members of a self-help group, were interviewed together, as they felt at ease sharing their thoughts in this setting.We implemented a worksheet method to allow the participants ample opportunities to organize and articulate their thoughts.Participants frst wrote down their responses to the given questions and then shared and discussed them with others.
Given that LLM-driven CAs are rooted in AI, we frst sought to understand the participants' overall AI literacy.We began by asking participants about the everyday AI products they knew and expressing their understanding of AI through drawings or writing.We then provided an overview of basic concepts and primary capabilities of AI to ensure a shared understanding among participants.To avoid biasing how participants perceive AI, we structured our materials based on AI education guidelines for non-expert students, which are "fve big ideas" for teaching AI proposed by Touretzky et al. [97] and AI literacy competencies for learners suggested by Long and Magerko [70].Our explanation covered topics such as "how AI works", "what AI can do", "the positive and negative impacts of AI applications", and "the importance of critical thinking", all of which were presented with relevant examples.Based on our explanation, the participants refected on their prior experiences with AI-embedded products and discussed their relevance to their daily lives.We then introduced the most popular service of LLM-driven CAs, ChatGPT, along with the basic concept of LLMs.We detailed the diverse applications of LLM-driven CAs with examples and highlighted ethical concerns such as hallucinations, data security, and emotional detachment.This explanatory content was developed based on the "Quick start guide for ChatGPT and artifcial intelligence in higher education" published by UNESCO [88].The session concluded with discussions on their impressions of Chat-GPT and its potential positive and negative impacts on their lives.The face-to-face group interview sessions lasted approximately 45 to 83 minutes.For reference, the focus group interview questionnaire can be found in the Appendix A.1.

Workshop.
After the group interviews, we conducted a workshop to capture the participants' expectations and challenges in integrating LLM-driven CAs into their daily lives.To prompt participants to envision the diverse application of CAs in various life aspects, we encouraged them to frst refect on their day-to-day needs and then consider the potential use of CAs to address these needs.The workshop was conducted individually with each participant to ensure that the researchers provided one-on-one support.This individualized approach not only enabled participants to share personal stories but also allowed researchers to closely follow their interactions with the CAs, probing their reasons and emotions [22].
We provided participants with a worksheet in the form of a daily planner to stimulate the use of LLM-driven CAs for various purposes in their lives (Figure 2).To recall their daily routines, we asked them to document their plans for the upcoming week in the weekly planning section.Based on this, participants were asked to choose a day they were eager or worried about and detail that day's events in a circular daily planning section.Afterward, they noted any anticipated challenges or signifcant issues next to each planned activity for the day.Moreover, we showcased categories and examples of questions posted by autistic individuals in online communities [45] (e.g., grooming, health habits, work and study, social activity, interest, etc.) to draw out everyday concerns and interests unrelated to their documented schedule.
Subsequently, the participants explored how to leverage LLMdriven CAs to address the daily issues and challenges they raised.We frst introduced various types of conversations with CAs through example prompts, including simple question-answer, simulations, small talk, and creative work.Next, we provided each participant with an individual ChatGPT account (using the GPT-4 model) and encouraged them to explore its features freely.After exploration, the participants brainstormed possible conversation topics with ChatGPT related to their previously mentioned daily issues and jotted them down on post-it notes.They then freely interacted with ChatGPT on the listed topics about 30 minutes and added, removed, or adjusted topics as needed.Throughout the process, the researcher did not directly engage with the participants' interactions and ideation with ChatGPT, to prevent our perspectives from leading their behaviors and thoughts.Instead, we served as moderators, assisting with technical issues related to ChatGPT and reminding them of unaddressed daily issues from their worksheets.We also stimulated participants' thought by continuously questioning their motivations for discussing the given topics with ChatGPT and their impressions of the CA's responses.
At the end of the session, we asked participants about their overall workshop experience.Our questions delved into what they had learned or realized regarding LLM-driven CAs through the workshop, any changes in their perspectives on integrating LLM-driven CAs into their daily lives, and any perceived challenges or potential risks associated with using this technology.The debriefng interview questionnaire can be found in Appendix A.2.The workshop lasted between 62 and 84 minutes.

Data Analysis
All interviews and workshops were audio-recorded and transcribed.The dataset for analysis included transcriptions, participants' worksheets, post-its listing desired conversation topics with ChatGPT, and ChatGPT usage logs from the workshop.We frst sorted the 99 desired conversation topics with ChatGPT and documented the post-its to understand their expectations of LLM-driven CAs.Using nine categories from a previous study that analyzed questions posted by autistic individuals in online communities [45], the frst and second authors collaboratively categorized the participants' conversation topics.Throughout this process, the two categories were merged, and four new categories emerged.The fnal categories were:1) daily planning, 2) physical or mental health habits, 3) grooming, 4) household chores, 5) fnancial management, 6) work and study, 7) social relationships and communication, 8) romantic relationships, 9) interest and leisure, 10) emotional support, 11) self-inquiry, and 12) small talk (Table 2).
We analyzed the qualitative data using a thematic coding approach [13].All raw data were digitized using ATLAS.ti[3], and the frst author initiated the analysis with the open-coding process.The second and third authors then critically reviewed the codes created by the frst author and corresponding raw data, which led them to develop their individual codebooks.Following this, all three authors jointly conducted a comprehensive review of these codebooks and provided detailed written feedback.The team then gathered to reconcile diferent views, resulting in a list of 109 initial codes.To further validate our coding, we conducted two additional rounds of review and cross-checking.This process led to the addition of four new codes, bringing our total to 113.These codes were then categorized into broader themes using the afnity diagram method [44], with a primary focus on: 1) perceptions and attitudes toward AI and LLM-driven CAs, 2) interaction patterns with LLM-driven CAs and their reasons, and 3) potential challenges with CA interactions.Consequently, we structured our results into ten primary themes and 23 sub-themes.

FINDINGS
In the following section, we delineate our fndings into three distinct themes: (1) our participants' perceptions of AI and how they relate to the use of LLM-driven CAs; (2) our participants' expectations of LLM-driven CAs in everyday contexts; and (3) the unforeseen challenges they encounter when interacting with LLMs.

What is AI? What role does it play in your
daily life?
4.1.1AI is said to be omnipotent.Of the 14 participants, eight (P2, P3, P7, P9, P10, P12, P13, and P14) were aware of AI, but felt it was distant from their daily lives.They could only associate AI with recommendation systems on YouTube or Netfix and had limited frst-hand experience with such AI-embedded products.Most of these participants, without a technical background, learned about AI through media sources such as news outlets and YouTube.In particular, the recent outpouring of media content spotlighting the life-changing ChatGPT has led them to perceive it as an extraordinary entity capable of achieving anything.Participants viewed ChatGPT as "smarter than humans (P3)" and believed it could "address every issue in my life (P7), " showing strong optimism about using this technology.
I have been so hooked on the news these days, and what catches my eye every single time is ChatGPT.I have read tons of articles saying how groundbreaking and exciting it is.I even heard an episode where they were calling it a "world-changing technology."There have even been debates about whether using it is cheating or not.So, to get to use ChatGPT in this study?I am super pumped.(P9) However, the elevated expectations surrounding LLM's capabilities served as a barrier for some participants (P3, P7, P10, and P12) in using ChatGPT.They expressed low confdence in learning and adopting ChatGPT, attributing this to their own lack of educational availability.This led to concerns like, "Can someone like me, who knows nothing about this all-powerful AI, actually use it?(P10)", thus lowering their confdence in accessing ChatGPT.P12 even reached out to us along with their parents the day before the workshop, asking, "I have autism and developmental disability.Can I use ChatGPT?" 4.1.2I'm proficient with AI, but only in the area I first encountered.The other six participants (P1, P4, P5, P6, P8, and P11) who had been using ChatGPT self-identifed themselves as skilled AI users and viewed AI as an integral part of their everyday lives.Most of them were either college graduates or undergraduate students, and four of them were majoring in data science and AI; therefore, they understood in depth how AI works and what it can do.They spontaneously discussed privacy and overreliance concerns that they encountered in various AI applications, such as job interviews and smartphone facial recognition.In addition, some participants established criteria for responsibly harnessing ChatGPT after refecting on their usage.
I have been hesitant to fully adopt ChatGPT because if everyone uses AI, the outcomes will all become homogenized.(P8) My writing needed more than a little help, so I tried having ChatGPT edit it for me.The edited version definitely felt well-written and excellent.But I couldn't bring myself to submit that version.I fgured I should just use it as a reference, especially given the ongoing discussions about copyright issues and misinformation.(P6) While the participants were already using ChatGPT with high levels of AI literacy, their use was mostly limited to specifc areas.Their experiences with the technology during the initial phase heavily infuenced how they interacted with the technology, as well as their attitude toward ChatGPT.Most participants in this group were introduced to ChatGPT through university classes and reported using CAs primarily for academic purposes.During the workshop, in which participants were encouraged to ideate using ChatGPT in diverse aspects of their lives, they were surprised that they had not considered using the technology outside academic contexts.This shows that our autistic participants had few opportunities to proactively explore ways to integrate ChatGPT into their lives.
[After sharing personal concerns with ChatGPT] I never imagined I could have such a meaningful conversation with ChatGPT.I mean, I knew it was possible, but why didn't it cross my mind?I guess I was too focused on using it for school, especially since my professor hyped it up as a great study tool.(P4)

What do you expect from LLM-driven CAs?
We uncovered what autistic participants broadly expected from ChatGPT in their daily lives by further exploring the topics they ideated by interacting with ChatGPT in workshops, as shown in Table 2.

Go
-to source for daily information.All participants were adults either living independently or aspiring to achieve fnancial and social independence from their parents.Therefore, one of their primary concerns was how to manage day-to-day challenges without the assistance of their family members.The immediate, ondemand nature of ChatGPT contributed to the participants' confdence in handling daily issues.It is noteworthy that the participants noted a lack of online communities in Korea where autistic individuals can fnd information tailored to their particular situation.They welcomed ChatGPT as a newfound channel for discussing issues that they wanted to keep private to those around them.
So I used to go on Jisik-in (a Q&A service (e.g., Quora) by the Korean portal company, Naver) a lot.Many of my autistic friends use it too.But the problem is that Figure 2: Translated versions of participant worksheets.Participants wrote their daily routine in the circular daily planner section and identifed anticipated challenges and issues.They then explored approaches for addressing these issues with LLM-driven CAs and wrote down these on post-it notes.We marked issues related to the daily routines written by participants in the circular planner in black, and issues related to their broader life in blue.The original worksheet is attached in the Appendix B.
it is hard to fnd questions asked by autistic people there.I am usually left with questions and answers that wouldn't really apply to my situation.And when I urgently post a question, I rarely get an immediate reply.It is pretty much useless.(P7) As for me, there are things I don't want to discuss with my parents anymore... like dating or sexual questions.Then I have no one else to ask.When you guys [the researchers] are gone, I would like to ask ChatGPT about that stuf.(P10) Workshop participants inquired about a range of life skills from ChatGPT.Their questions spanned topics such as dealing with cockroaches that suddenly appeared, curbing excessive spending on gaming, matching clothing colors, using a nagging bot to beat laziness, and strategies to quit smoking.Overall, the participants rated the CAs' responses as highly useful.Participants specifcally appreciated the directness of the CA's answers, which "only responded to exactly what I asked (P9)."Several of our autistic participants commented that sorting through huge amounts of information on the internet to determine what is relevant is challenging.By  contrast, ChatGPT provided contextualized information in an easyto-understand format, enabling participants to efectively absorb the information they sought.

In search of social clues.
As reported in prior research [4], autistic participants expressed that understanding and navigating implicit social rules remains a challenge.Furthermore, they noted that few resources were available for seeking social advice after reaching adulthood.This aspiration was further confrmed by the fact that participants often brainstormed topics related to social activities for ChatGPT.All participants except P2 and P11 inputted prompts related to their social relationships with family, friends, colleagues, and romantic partners.Conversation on this topic among participants largely took two approaches: 1) seeking coping strategies and 2) interpreting others' intentions.
I had to do a group project in college, and the professor said we had to form our own groups.But I could not fgure out how to approach people to form a group, so I ended up doing the entire project by myself, for the entire semester.If I had ChatGPT back then, I would have asked how to approach people to form a group.(P1) I uploaded a selfe with my friend on Instagram today.And my friend's face got cut of in the picture so I said that it seemed like they had been hit by a bomb.After that, I kept wondering if I made a mistake, but I could not ask my friend about it.I have never really resolved these kinds of concerns.I just let them go.I'm not even sure if I'm doing things right... (P4) The most commonly engaged social conversations with Chat-GPT involved exploring strategies to cope with unfamiliar and stressful social contexts, such as "what to do when customers get out of hand (P5)," "how to comfort parents after they've had an argument (P3)," and "how to report tragic news with composure (P9)." They also consulted ChatGPT to determine whether their behavior was appropriate in particular scenarios, such as interacting with friends or controlling their facial expressions on the job.Some participants were interested but worried about romantic relationships, and simulated blind date scenarios with ChatGPT as a form of practice.
Some participants went beyond refecting on their own social behaviors and employed chatGPT to interpret the emotions and thoughts of others.For example, P7, P10, and P12 copied and pasted text messages that they had received from their friends and asked ChatGPT what their intentions might be.P5, who had just launched an online store, provided ChatGPT with information about their target customers and asked how potential customers might perceive their marketing text:  ------------------------------I plan to sell bathroom wipers online to the people mentioned above.How would this person feel about the promotional content I wrote below?What features should I emphasize to increase the likelihood of her making a purchase?(P5's prompt) 4.2.3Common ground, unlimited conversation.Many participants were deeply interested in discussing specifc topics, but they often found it difcult to satisfy this need in everyday conversations.There were a few people around them with shared interests, and they experienced people avoiding conversations when delving into their favorite topics.In this context, the participants wanted Chat-GPT to become an unweary conversation partner, allowing for unlimited discussions on any topic they desired.In our workshop, some participants with specifc interests (P1, P6, P7, P8, and P11) engaged in ChatGPT by debating social and political issues, cowriting song lyrics and poetry, and brainstorming game plots.In addition, P12, who dropped out of college due to bullying, had unfulflled academic curiosity, particularly in mathematics and science.He quizzed ChatGPT about scientifc knowledge and viewed CAs as outlets for satisfying his intellectual needs.These participants valued LLM-driven CAs as allowing them to fully express their interests without social pressure and to conform to the standards or norms of others.
Talking with people, you can't just talk about what interests you.They would get bored of me.But this thing (ChatGPT) answers sincerely no matter how many questions I ask, and I really like that.(P11) 4.2.4Qestioning the Inner Self .Some participants (P3, P9, P10, and P12) shared questions about themselves with ChatGPT.When refecting on their own characteristics and forming a self-concept, they referred to ChatGPT answers as second opinions.Participants often noted that their perceptions of themselves difered from those of others judged them.To gain self-understanding without being swayed by others' opinions, they hoped that ChatGPT, trained on extensive data, would provide more objective perspectives.P6 asked ChatGPT whether he is an introverted person by elaborating on past events with friends.P12 wanted to explore "autism" in depth and how it difered from other developmental disabilities Meanwhile, some participants (P9, P10) expressed concerns about their career potential: P9: I am currently working in a creative role, but my boss told me that someone like me, who has autism, is better suited to traditional roles.It feels as if he is suggesting that I should quit.Am I not a good fit for this job?[...] Also, I want to know what kinds of stress can autistic workers experience in their workplace and what are some good ways to address it?ChatGPT: It is important to remember that Autism Spectrum Disorder (ASD) encompasses a wide range of characteristics and abilities, and it is incorrect to generalize that someone with autism is not suitable for creative tasks.As a result of a conversation with ChatGPT, P9 not only reassessed his work experience but also realized that communicating with his boss was a major stressor.This led him to request for improvements.
Sudden changes in tasks without notice. . .vague instructions. . .Yes, that is it!That's been the issue all along.I really need to show this to my boss!(P9) 4.2.5 Finding comfort in chat.We observed that many participants (P3, P4, P10, P11, and P13) sought emotional support from ChatGPT.Participants emphasized that, unlike their interactions with people, they felt that ChatGPT appeared to make an efort to understand them without harboring biased views.This allowed them to freely express their innermost thoughts and emotions.For example, P13 discussed his concerns about feeling inferior to ChatGPT: P13: I saw some friends using more fancier trading card game gear than I had, and it made my stuff look like trash.I felt worthless and lagged behind.Then, I hear a voice in my head saying, "Hey, don't you wanna beat them?Don't you want to be the perfect gamer?"It feels like I am threatened by myself.I want some comfort.
ChatGPT: I know that it is very tough and disheartening when you compare yourself to those around you, especially when they seem to shine so brightly.However, remember that your worth is not determined by the material things you have or do not have.[...] And always remember, you are on your own unique journey.You are already enough, as you are, and you are constantly growing.
After receiving ChatGPT's response, P13 paused for quite some time and shed tears, saying, "This is the frst time I've received comfort instead of advice to correct myself." Other participants (P4 and P10) mentioned that conversations with ChatGPT could potentially serve as an alternative to counseling in the future, pointing out that professional counseling was helpful but fnancially burdened.
[After writing a prompt that asks "praise me unconditionally"] It (ChatGPT) sounds just like the counselor who comes to the mindfulness sessions I attend.She is great, but I cannot see her all the time, and individual sessions are too expensive.I think it could really help with my self-esteem, like writing a journal before bed.(P10)

What challenges do you face with
LLM-driven CAs?
4.3.1 Excessive reliance on ChatGPT.The logical and persuasive responses of ChatGPT impressed the participants, resulting in a high level of trust.They even felt that ChatGPT was smarter than humans, noting its fawless grammar and spelling.This occasionally prompted participants to regard ChatGPT's responses as an "answer sheet" to which to conform.In particular, they used the ChatGPT recommendations as a benchmark to compare and evaluate their own behavior.
[After asking the question, "Recommend me how to deal with my parents' nagging without getting angry?"]Oh, that is what I was doing.When my parents were nagging me, I would purposely leave home and go somewhere else just to cool of.I was doing something right!(P1) I've only tried the frst recommendation from the AI to improve my reading comprehension, regular reading.I didn't try any of the others, like making fashcards or joining a reading club.Maybe that's why my comprehension stayed poor... (P7) Whenever this occurred, we informed the participants that Chat-GPT's responses may not be absolute and should be used as a reference for their own decision-making.While participants were cautious in accepting the information related to work and study, they reported an unintended reliance on the CA, particularly when discussing "areas of insecurity (P14)" like social interaction.To avoid blind faith in ChatGPT, some participants came up with the idea of gathering multiple CA answers representing diverse perspectives to crosscheck their consistency.I'm not confdent about how well I'm doing in my school life.But this thing [the CA] is really eloquent.So, until you (the researcher) told me, I just believed it without any doubt.There are a lot of variables, and it's hard to make judgments, so I think I need to get multiple answers repeatedly.It's like asking it to "Talk like my friends, talk like my mom." (P8) 4.3.2Disclosing autistic characteristics to ChatGPT.Autistic participants, who perceived themselves as having a unique approach to thinking and communicating, desired to obtain CA responses that meet their specifc needs.Therefore, some participants (P12 and P14) detailed their physical and social backgrounds, particularly their autism-related characteristics, during their prompts.Can you explain the laws of the exponents in a way that is easy for me to understand? (P14's prompt) When participants included information about their autistic features along with their questions, the ChatGPT ofered responses like " : Not all autistic individuals are the same, so different approaches and support may be needed depending on individual needs and abilities."Participants noted that by disclosing their autism, they not only avoided generalized answers but also felt that their identities were respected.
However, the majority of participants stated that they would never disclose their autism traits or disability status to the ChatGPT.These participants had faced discrimination in schools or workplaces for disclosing their autism, leading them to be fearful of even revealing their autistic identity to AI. P3 said, "I worry that ChatGPT could subtly ignore me if I bring up autism.Like it might provide good information to others but not talk to me at all." Participants who were well-versed in how AI works raised concerns about the potential harm arising from data leaks.
Suppose my data gets leaked.It could spread and then all the machines around the world might know I have autism, right?Then when I use other technologies that have AI, they might already know I have a disability and exclude me accordingly.(P4)

DISCUSSIONS
In our study, we found that autistic individuals expected to use LLM-driven CAs for social advice and care for their cognitive health.However, the excessive reliance of users on the responses provided by CAs and the tension between personalization and privacy protection in the context of LLM-driven CAs have been identifed as concerns.Building upon these fndings, we propose an approach to encourage autistic users to proactively integrate LLM-driven CAs into their daily lives while preserving their unique attributes.

Taking initiative to appropriate LLM-driven CAs
Historically, the majority of research studies on autistic individuals have focused on changing their behaviors and traits to conform to societal norms that typically apply to non-autistic individuals [50,81].A signifcant shift occurred when an autistic sociologist introduced the term "neurodiversity," initiating a movement to view autism as an inherent characteristic rather than a condition [53,57].Essentially, the neurodiversity paradigm seeks to empower autistic individuals with more control over their lives by enhancing their autonomy and self-determination [65].In light of this shift, recent studies on HCI and accessibility have focused on fostering self-advocacy and self-empowerment among autistic users through technology, as opposed to merely accommodating technology [60,85,86,93,104].
Our research fndings confrm that interaction with LLM-driven CAs can put autistic users at the forefront of addressing their dayto-day needs.By reducing the dependence of participants on traditional online and ofine support networks, CAs improved their confdence in managing everyday concerns.Participants also creatively appropriated CAs to meet various everyday needs, such as rehearsing dating scenarios, creating a conversation partner that is knowledgeable in a certain domain, and building self-esteemboosting compliment bots.In some cases, the responses from LLMdriven CAs even motivated autistic individuals to take steps to improve their problem-solving skills in the workplace and advocate for their rights.
Our workshop encouraged autistic individuals to develop scaffolds for integrating LLM-based CAs into their daily routines.Before participating in our study, those who frst encountered LLM-driven CAs for academic purposes found it challenging to envision their potential applications in various aspects of life.Moreover, a lack of knowledge on AI and emotional barriers to LLM-driven CAs inhibited some participants from employing them.However, through the workshop, participants themselves identifed 99 diverse potential applications for LLM-driven CAs.They engaged in activities such as 1) identifying challenges they often face, 2) examining whether LLM-driven CAs could efectively address these challenges based on their personal needs, and 3) developing strategies to align technology use with their goals and values.Thus, autistic individuals were more than passive users learning how the technology works; they played an active role in shaping the use of technology to suit their specifc needs.This approach resonates with the argument by Spiel et al. that autistic people should experience technology in an agentive manner, respecting their unique identity and characteristics [94].Therefore, we call for the necessity of an on-boarding process that gives autistic people the opportunity to recognize their own needs instead of focusing only on their technical abilities.The LLM-driven CA platform could provide moments where users can express their values, explore the capabilities of the technology, and defne its role in a self-guided manner.In addition to eforts at system level, casual discussions in domestic settings [71] or within the autistic community could provide opportunities for autistic individuals to establish and refect their own criteria for the proactive use of LLM-driven CAs.

Toward an inclusive LLM-driven CA experience for neurodiverse population
Companies such as OpenAI, responsible for developing LLM models, have made signifcant strides in advancing conversational AI through accessible platforms, enriching the lives of countless individuals.Nevertheless, given the ongoing concerns about the underrepresentation of minority groups in AI training datasets [56,62,73,73,100,101], it is likely that LLM-driven CAs primarily refect the perspectives of neurotypical users, who constitute the majority in our society.Our fndings reveal that the responses from these LLM-driven CAs can have a signifcant impact on the lives of autistic individuals, from day-to-day decision-making to selfperception.Notably, autistic individuals tended to compare their own behavior with the responses provided by CAs, often adjusting their actions according to the suggestions provided by the CAs.This raises concerns that interactions with LLM-driven CAs could potentially enforce normative behaviors defned by neurotypical society on autistic users, potentially causing them to compromise their own needs [20].In this context, creating an inclusive LLMdriven CA environment where neurodiverse users can preserve their unique characteristics and values should be considered.Furthermore, it is crucial for users to provide textual prompts that refect their desires in LLM-driven CA conversations to receive outputs tailored to their individual needs [113].The participants in our study mentioned that they received more personalized responses when they clarifed their goals, situations, and needs.Particularly, when participants disclosed their autistic features to CAs, the CA responses considered their physical and cognitive attributes, and also emphasized the individuality within the autistic spectrum, stating that " : not all autistic individuals are the same, so diferent approaches and support may be needed depending on individual needs and abilities." However, some autistic users might not fully realize the benefts of disclosing their specifc requirements, or may be reluctant to do so owing to past experiences of discrimination, as observed among some of our participants.Therefore, the system can ofer an accessible environment that encourages users to articulate their specifc needs in a conversational manner.This could involve providing follow-up questions to gather additional information needed for tailored results.
Additionally, facilitating autistic individuals to critically evaluate whether responses from LLM-based CAs align with their own standards and values could serve as another pillar for promoting an inclusive experience.In recent years, signifcant eforts have been made to foster critical thinking skills in individuals to become responsible AI users, including UNESCO's ChatGPT usage guidelines [88] educational curricula [31,61,98] and informal experiences through exhibitions [67][68][69].While some studies aimed to cultivate critical thinking in AI usage among underrepresented groups (e.g., children from lower socioeconomic backgrounds [29] and black girls [91]), there is a noticeable gap in initiatives aimed at empowering neurodiverse populations to become critical AI users.There is an opportunity to develop education and infrastructure support that enables individuals to establish their own guidelines for the ethical use of LLM-driven CAs, similar to existing initiatives within the autistic community that ofer guidelines [112] to enhance digital literacy (e.g., social media usage, email writing).

LIMITATIONS AND FUTURE WORK
There are multiple limitations to consider in our study.First, since all participants were from Korea, the cultural context specifc to Korean individuals with autism may have infuenced our results.Although our sample size aligns with other qualitative studies on technology use in autism [77,110], its relatively limited size could constrain the generalizability of our results.Second, our research only allowed participants to interact with LLM-driven CAs in a controlled setting guided by researchers, aiming to create a safe setting for engagement.However, the workshop's brainstorm-focused nature restricted our examination of on-demand use of the technology in daily scenarios, such as how autistic individuals accept CA's responses in real-life settings, the impacts this technology might have on their lives, and any additional desires or unexpected risks.We recommend that future studies employ empirical approaches, including diary studies, longitudinal studies, and ethnographic observations, to gain a more comprehensive and contextualized understanding of autistic users' real-life experiences.Furthermore, the results of our study, rooted in direct engagement with autistic users to understand their needs, can also serve as a starting point for comprehending their technology use in empirical settings based on their unique perspectives.Particularly, the challenges identifed in our study shed light on critical areas that require more attention in real-world applications.This understanding is essential for researchers to ensure ethical and safe interactions for participants in future studies.Last, our study focused only on a specifc subset of the autistic population that is literate.While the insights gained offer a holistic understanding of the technology experiences of users with autistic features, we might miss capturing opportunities for broader inclusivity, particularly among non-verbal autistic individuals.Prior research has shown the potential for non-verbal autistic children to communicate via chatbots through augmentative and alternative communication (AAC) tools [24].Therefore, future studies need to explore the possibilities of using multimodal approaches with LLM-driven CAs as a new communication channel for this group.

CONCLUSION
This study explored the opportunities and challenges of integrating LLM-driven CAs into the lives of autistic individuals through focus group interviews and workshops.We observed that autistic participants aspire to use LLM-driven CAs in proactive and creative ways to address a wide range of issues in their daily lives.However, autistic individuals often inadvertently adopt the responses from CAs and attempt to adjust their behavior accordingly.We conclude with recommendations aimed at empowering autistic users to take an agentive and critical stance in defning the meaning and role of this technology in their lives.We hope that these fndings and recommendations pave the way for LLM-driven CAs to serve as new mediums through which autistic individuals can better communicate with the world, express their voices, and actively engage as social participants.

Figure 1 :
Figure 1: Focus group participants discussing their prior experiences with AI (far left), and workshop participants brainstorming how to use LLM-driven CAs in addressing their daily issues by interacting with ChatGPT.

P5:
Basic Information: 30s-40s, Female, Raising children, Employed Values: Children should be raised in a clean environment!Goal: Needs a clean space + wants to create a clean space without spending too much time.Taste and Personality: It is the bathroom my kids are using after all. . .Emotion: I need to start taking care of my health. . .
P12: I have autism and work a contracted job for 8 hours a day.I have been here for seven months and cannot transition to a full-time position.However, if I continue to give my best, always greet people cheerfully, and respond respectfully to my team, can I land a better job in the future?For reference, I understood Carnegie's principles of human relations in 3.5 hours.I can use public transportation alone, but I cannot go overseas because of difficulties in English communication.However, I am good at writing and summarizing.(P12's prompt) P14: My dream is to master middle school math, especially the laws of exponents.So I have recently been learning math from friends I met on Discord.But I have autism, and I find it difficult to focus while studying.

Figure 3 :
Figure 3: Original versions of participant worksheets.The translated version is located in Figure 2 of the main text.

Table 1 :
Demographics of study participants.The data are presented in aggregate form to preserve the anonymity of individual participants.

Table 2 :
Conversational topics generated by participants with LLM-driven CAs during the workshop.Participants engaged in conversations with ChatGPT on all listed topics throughout the workshop.

•
Have you experienced or do you anticipate any positive impacts or benefts of AI in your daily life?• Have you experienced or do you anticipate any negative impacts or challenges of AI in your daily life?• Do you have any personal rules or criteria for using AI?If so, could you describe the events or experiences that have shaped these rules or criteria?A.1.2(Afer Introducing ChatGPT) First Impressions and Thoughts about ChatGPT.• What is your (frst) impression of ChatGPT? • Are you interested in using ChatGPT?• Do you think ChatGPT can enrich your everyday life?In what aspects of your daily life do you think it would be most useful?• Do you have any concerns or worries about using ChatGPT?A.2 Workshop Debriefng Interview Questionnaire • Which conversation with ChatGPT today did you fnd most satisfying?• Were there any conversations with ChatGPT today that disappointed or displeased you? • Do you plan to continue using ChatGPT after the workshop?In which specifc areas or aspects of your life do you expect to use this technology?• If you have no intention to continue using ChatGPT, what are the main reasons?• Are there any challenges or risks that concern you about using ChatGPT?• Did participating in the workshop give you new insights about ChatGPT or AI technology?• Have there been any changes in your perceptions and attitudes towards using ChatGPT in your everyday life when compared to before the workshop?
B ORIGINAL WORKSHEET OF PARTICIPANTS