Attention deficit hyperactivity disorder (ADHD) is the most common neurodevelopmental disorder in England and affects nearly 7% of adults globally [1]. It is characterized by impairment across inattention, impulsivity, and hyperactivity. Characterizations of ADHD suggest there are three subtypes—inattentive, hyperactive, and combined subtype, with heterogeneous traits across and within individuals [2].

young people with ADHD are likely to face poorer outcomes in life functioning across domains including academia, antisocial behavior, driving, nonmedicinal drug use or addictive behavior, obesity, occupation, service use, self-esteem, and social functioning outcomes. When compared to people without ADHD, those with ADHD face poorer outcomes whether their condition is treated or untreated [3].

ADHD was previously thought to be a disorder limited to childhood, but more recent understanding suggests that in around 3% of the population, ADHD with a childhood onset is maintained into adulthood, equating to nearly 140 million people worldwide. When symptoms or traits are considered, regardless of childhood diagnosis, this number could include an additional 366 million affected adults [4].

There is ongoing discourse regarding the potential overdiagnosis of ADHD in the United Kingdom; however, evidence suggests that, though a small number of cases may incorrectly be attributed to ADHD, diagnosis rates remain “substantially below the population prevalence in the United Kingdom, providing no evidence at present that ADHD is overdiagnosed at a population level” [5].

ADHD health care provision is failing patients [6]. This issue is exacerbated for people with ADHD in transition between childhood and adulthood (ie, aged 18-24 years) [7].

There are long waiting lists for ADHD diagnosis, with minimal support for those on the waiting list [8]. Not only do young people have better outcomes when they receive treatment, but being on a long waiting list can have negative outcomes. These individuals are more likely to use more health care resources than others [9], and being on a waiting list can cause financial stress due to loss of employment, social isolation, lack of support, and reduced quality of life [10].

Even after diagnosis, pharmacological and psychosocial support can be lacking and is often highly variable depending on the services available in the area where the young person lives [11].

Systemic barriers that face neurodivergent young people when accessing the health care system further decrease the likelihood of accessing support [12]. Compounding this, ADHD is highly heritable, and it is estimated that 44% of people with ADHD also have at least one parent with ADHD, meaning that barriers persist for young people who need parental support to access the health care system [13].

There is potential for carefully curated bespoke digital health technologies to help address unmet need for young people with ADHD and reduce gaps in service provision. Digital technologies represent a financially viable solution, with some methods having demonstrated cost-effectiveness [14]. They also offer the potential to help support young people with ADHD but lack a rigorous evidence base [15]. Mobile phone apps have been shown to create behavior change and engagement by harnessing behavior change techniques [16,17]. Other digital health interventions (DHIs) have been effective for ADHD [18]. Furthermore, the National Health Service (NHS) is encouraging increasing digitalization as an adjunct to usual care, with the July 2025 “Fit for the Future” report recommending a shift from “analogue to digital” care [19].

Nevertheless, current digital information can be unreliable, untrustworthy, and not evidence-based, and yet consumed frequently by young people with ADHD [20-23]. Therefore, there is a need for ADHD-specific DHIs that are evidence-based.

Large language models (LLMs) are increasingly being used in DHIs and have been used for a variety of purposes within mental health support [24,25]. They have shown promise to support young people with ADHD. Chatbots can break down tasks into manageable chunks, reducing overwhelm [26,27]. They can offer on-demand and consistent support [28,29], emotionally neutral feedback that avoids judgment [30,31], and engagement and motivation [32,33]. Users often find chatbots more fun and less intimidating than traditional tools [29,34]. There is also the option for personalization in terms of the user’s pace, language, and preferences [35-37], as well as cultural, linguistic, and neurodiverse needs [35,36]. All the above are features that could make chatbot interventions appropriate for people with ADHD, who may face challenges with being easily distracted, forgetful, and finishing tasks.

Therefore, we suggest that an evidence-based chatbot could improve communication of health care information for individuals with ADHD who may have differing needs. A well-designed chatbot could be accessible to people from different backgrounds and provide a trustworthy alternative to the high volumes of digital misinformation and unreliable resources that exist online [20]. Chatbots have previously been used to provide support for mental health conditions (these are advanced in the eating disorder literature especially) [38], but this work is novel in being the first that the authors know of that aims to act as an adjunct to ADHD care, and particularly with respect to young people in this transitional period between services.

However, there are important concerns around safety after problems have arisen with other chatbots. There are examples of generative AI (GenAI) tools enabling users to generate harmful eating disorder content [39]. Indeed, a chatbot launched in 2023 by the National Eating Disorder Association in the United States was taken down shortly after launch as it provided harmful advice to users about losing weight and dieting [40]. This tool was developed by academics, who claim that the safeguards they set up should not have allowed this to happen, and that edits had been made to the chatbot function before publication. Further demonstrating that chatbots for mental health are in their infancy, this was a rule-based chatbot, which, unlike GenAI models, was developed for precision. Rule-based AI is transparent and inflexible—it always gives the same answer to the same question. This shows that significant improvement is required before rule-based chatbots, let alone GenAI, are safe for use as patient-facing medical devices.

Preliminary evidence suggests that the feasibility and acceptability of DHIs for young people with ADHD are strongest when co-developed with experts with lived experience (LE) [41]. This underlines the need for a development, design, and implementation process that uses co-development methods. We will outline the processes that we followed in a linked paper. Despite rapidly evolving challenges and solutions in relation to AI-based health care, it remains essential that early chatbot development work involves people with LE, digital experts, and health care professionals (HCPs). Co-production is focused on increasing understanding about the priorities, needs and preferences of young people with ADHD for a chatbot that will help them manage their condition safely, to access appropriate health care, and to thrive.

Consultation with young people has shown that they favor digital apps as a source of health care information, and that this should contain psychoeducation (information about the condition) and information about accessing health care services in the United Kingdom [42,43].

Health care providers also favor an online solution, as this allows information to be kept up-to-date more easily and the ability to signpost to one trusted source [11,44]. Clinicians also wish to improve the sharing of information with young people [45], and support digital interventions as an adjunct to primary care if “reliable and well-curated” [44]. Digital resources can be updated frequently when new information comes to light, can be accessed quickly by the patient, and can give an LE perspective that most general practitioners (GPs) cannot. Digital interventions could also reduce administrative tasks for consultants, as common queries and requests for information could be fielded through these interventions.

The provision of information about how to access health care is recommended by the National Institute for Health and Care Excellence (NICE) and the NHS; nevertheless, accessing the health care system in the United Kingdom, whether public or private, can be confusing for young people and supporters. Supplying a web portal for community-based ADHD care saw parents rate improved ADHD symptoms and improved ADHD care quality, showing that digital interventions can improve access to health care [46]. The chatbot may fill this gap by improving communication through the synthesis of information and the ability to tailor communication. This has the potential to be prescribed through a social prescribing model for health care access equity.

This paper describes early co-production work, conducted between January and September 2025, and reports on the preferences of young people with ADHD and their supporters. Preferences relate to an AI chatbot that is trained on evidence-based materials specifically to give psychoeducational, behavioral, and health care information for young people aged 16-25 years with ADHD. The information reported in this article reflects conversations with collaborators with LE of ADHD and reflects proof of concept only; this is development work, and extensive future research would be required to ensure the chatbot was safe and fit for purpose.

An interdisciplinary team was established as part of the broader SmartADHD project [47]. This included project leads from the University of Exeter and the Digital Futures Lab, an LE expert, and a human-computer interaction (HCI) researcher from the University of Bristol. LE and HCP research advisory groups (RAGs) were set up in line with UK Standards for Public Involvement, with linked working groups (WGs) whose members could contribute flexibly to research activities [48]. For details, see terms of reference and role descriptors (see Multimedia Appendix 1 and Multimedia Appendix 2 for examples). A mission statement was co-developed, and training resources (eg, on the basic design principles of HCI) were provided for researchers and lived-experience colleagues and shared via YouTube and the study website [49]. Meetings were held to discuss research, software engineering, and ADHD specific processes and experiences, to help build a shared language across disciplines and experience and, through an iterative process, produce a combined direction of travel and map a perceived ideal outcome.

The wider research studies with which this co-design exercise is associated adopt the meta-paradigm of critical realism. This states that an objective reality exists independently of our knowledge of it, and that while there are real and actual truths about the nature of the world, it is impossible to completely understand this reality [50]. Critical realism is an appropriate and pragmatic approach commonly used in applied health care research and supports the need for theory-driven, person-focused, and context-aware intervention development [51].

With this in mind, the person-based approach (PBA) was used to guide digital intervention development, as this is an established method for combining evidence, theory, and person-based research [52]. As part of a larger body of SmartADHD work, a review and synthesis of relevant literature was conducted by the research team (AP, TE, and EB), and a summary was created of relevant primary and systematic review evidence. A theory of change model was drafted, based on the behavior change wheel and intervention planning table populated, covering target behaviors, for uptake, engagement, and knowledge use (Multimedia Appendices 3 and 4) [53]. Guiding principles for intervention design (covering user characteristics, design objectives, and key features) were drafted and collaboratively reviewed over multiple iterations. Gaps in knowledge were identified in relation to user needs (EN) and noted for discussion during workshops. Details of PBA documents will be provided in a linked SmartADHD publication.

A diverse LE working group of 13 people was formed to co-develop the Chatbot prototype, with work supported and guided by SmartADHD RAG members and the wider study team. The working group, which included 9 young people with ADHD and 4 supporters (parents), was made up of a mixture of individuals with previous involvement in the Science of ADHD and Neurodevelopment (SAND) collaboration, including members of the “Mapping ADHD services in primary Care” (MAP) study RAG, and new members [54]. Members were identified via previous involvement, study networks, and word of mouth, and invited to express interest in involvement via a Microsoft Form sent as a link [55]. Members were provided with terms of reference, links to SmartADHD resources, an introductory meeting, and a schedule of planned meetings. We engaged with colleagues through a range of different methods to help keep all team members well-informed, including frequent reminders. Meeting invites were sent via email, Microsoft Teams, WhatsApp (Meta), and text, depending on member preference [56]. Attendance and demographic information for these groups can be found in Tables 1 and 2. Some members attended several meetings; some did not attend any.

Five workshops were held with LE collaborators on chatbot co-development, covering guiding principles, tailoring for underserved groups, language and content, and chatbot identity. A further workshop was held with HCP collaborators. These are built on the guiding principles and theory of change documents from the PBA. Sessions were run flexibly via Microsoft Teams, often including presentation slides, with notes made on the slides by the meeting facilitator. Minutes were made and shared with attendees afterward. Meetings were recorded to add detail to notes after each workshop. Feedback from the workshops was summarized and translated by the research team, following content analysis methods, into a specification document to inform prototype development. The workshops and the resulting feedback are outlined in Table 3.

Following the workshops, a technical specification was developed to guide chatbot initial development, and then refinement, with consideration of 1. content from workshops, 2. information collected from the person-based approach, and 3. Basic Design Principles of HCI. Discussions were held with the software engineer (DH) and wider team, prioritizing feature development and considering technical and resource constraints. A final prototype was produced and trialed by research and HCI colleagues (EB and EN). Technical information about the chatbot can be seen in Textbox 1.

Following established PBA methods, think-aloud interviews were carried out with LE collaborators to user-test the first prototype of the chatbot and gather feedback on user experience [58]. A set of feedback prompts (see Feedback Points in Multimedia Appendix 8) was developed by EB, trialed with an HCI expert (EN) in a pilot interview, and reviewed by the team. Then, working group members were invited by email to book a 1-hour online meeting with researcher EB, with a link to a Microsoft bookings page. Reminders were sent via email or WhatsApp, in line with individual preferences. EB is a female research assistant who holds a BSc and had over 3 years’ experience in mental health research at the time of interviews. EB had an existing relationship with all members of the working group as she was involved in coordinating the group. Collaborators had been involved in conversations around the chatbot from the conceptualization of the project, some more directly than others. Collaborators knew about EB’s background and role on the project as the coordinator of the public involvement, research, and digital teams. For the purpose of this initial user testing, users were made aware that the chatbot was a prototype. They were only given access to this during the session and used it in the presence of a researcher. The chat history was saved within a password-protected login, which could only be accessed by the developers and research team.

Six think-aloud interviews were conducted over Microsoft Teams, with 3 young people with ADHD (collaborators 2, 3, and 6) and 3 supporters (parents) of people with ADHD (collaborators 1, 4, and 5). Three collaborators were from South West England, with the remaining collaborators from the North West, North East, and East of England. Four were female, and two were male. Collaborators brought a diversity of experience in relation to previous familiarity with Chatbots and with this project. One supporter had not used a Chatbot before, while the rest were experienced users. Some attendees were active members of the working group and had been involved in prototype development, while 2 members - though involved in the SmartADHD project (either in the WG or RAG) - had not participated in chatbot-specific workshops, so the chatbot was completely new to them. Though they may have seen a screenshot of the chatbot previously if involved in workshops, no collaborators had directly interacted with the SmartADHD chatbot before.

No repeat interviews were carried out. Informal field notes were made by EB throughout interviews. The think-aloud interviews were recorded with permission from collaborators and transcribed automatically; these were not sent to collaborators for comment as it was not essential for the transcripts to be completely accurate, and recordings could be consulted when the researchers might have been unsure. We (EN, EB, and AP) adopted a deductive analysis approach to categorize feedback, to provide specific information to inform future design iterations. In particular, EB deductively identified feedback using Lumivero’s NVivo 14 [59], in relation to features that worked well, features that did not work well, future developments, and suggestions on every segment of text. These were categorized into the following predefined topics relating to components of the chatbot:

This study describes patient and public involvement and engagement (PPIE) co-production research to inform early co-development of a chatbot prototype for people with ADHD. Public involvement is defined by The National Institute for Health and Care Research (NIHR) as research being carried out “with” or “by” members of the public, and public engagement is described as the “myriad ways in which the activity and benefits of…research can be shared with the public” [60]. In the United Kingdom, “ethical approval is not needed for PPIE activities”, which are considered essential as they facilitate more impactful research outcomes [60,61]. Further clarification can be found in guidance provided by the co-production collective:

“Existing research ethics processes and policies focus on the relationship between researcher and participant and are not concerned with the consideration of ethical issues in public engagement and Patient and Public Involvement and Engagement” [62].

In the methods described above, team members were consulted as expert and equal contributors to the design process and provided iterative design feedback via think-aloud interviews, rather than participating as research subjects and providing data. This work meets UK Health Research Authority best practice principles for public involvement by engaging experts with LE in early stages of intervention co-development [63], and laying strong foundations for jointly defining the ‘rationale, scope, design, and conduct’ of future planned research to develop and evaluate a Chatbot for young people with ADHD.

This PPIE and co-production research was conducted in line with international ethical guidelines [48]. It followed UK Research and Innovation key principles for co-production in research [64], and the established framework of UK Standards for Public Involvement in Research [48]. This included, for example, working together, using clear communication, providing inclusive opportunities, establishing equitable partnerships, and supporting learning.

LE colleagues formed a key part of the interdisciplinary research team and contributed as equal partners. As described above, terms of reference and role descriptors were provided for all RAG and WG members. LE colleagues received payment in recognition of their time, in the form of online shopping vouchers, while HCPs were provided with certificates of continuing professional development. Collaborators were provided with opportunities to learn more about research, receive feedback on the impact of their involvement, and were supported to contribute to outputs, including this publication.

Findings from the 5 workshops, which brought together members of the working groups to co-design the chatbot, are summarized in Table 3 (see above), with details provided in the Feedback From Initial Co-Design Workshops section. Feedback from think-aloud meetings with 6 collaborators, where they were invited to trial the co-developed early chatbot prototype is provided in the Think-Aloud Interviews—Feedback on the Initial Prototype section.

Feedback was collated from the transcripts and notes of think-aloud meetings with 6 collaborators, which captured their real-time experiences and views as they interacted with the early chatbot prototype. Users provided rich feedback in relation to strengths, weaknesses and opportunities for development.

From these think-aloud meetings, five topics were identified by EB and TE: the chat box and typing interface, the text-to-speech function, the conversational flow and feel, the avatar, and the message content. Users were generally very positive about the content of the chatbot and liked the design of the interface (aside from some technical problems) and the text-to-speech, but there was some specific feedback on how the voice could be improved. There were mixed views on the avatar. The flow of the conversation was reported as being generally good, but users did not like it when the chatbot session “crashed” and erased the chat history. Crucially, though they liked the content of the messages, collaborators were unable to identify where the information had come from when asked. A summary of the feedback from the think-aloud interviews, with illustrative quotes, can be found in Table 4, with further details provided below. A screenshot of the chatbot interface can be seen in Multimedia Appendix 7.

Initial user feedback on the early prototype of the evidence-based chatbot suggests that this is an avenue that is feasible for future development. Experts by LE (supporters and young people) commented on aspects that they liked, disliked, and future improvements that could be added to the chatbot, particularly about the overall experience, the purpose, the design, the customization settings, and the integration into a future app. Seven key recommendations for future development can be seen in Textbox 2.

The structured nature of the workshops may have influenced the categories chosen and the nature of feedback that was identified. However, this pragmatic approach is appropriate for early intervention development and was necessary to gain specific feedback and actionable suggestions. The think-aloud interviews are likely to have elicited any feedback that did not fit into earlier sessions and, although they followed a brief topic guide, were left as open as possible on purpose.

This was a multidisciplinary project which involved experts by LE, digital experts, HCI experts, clinicians and ADHD researchers. We all come with a wide range of experience and expertise, and though the research team were able to understand the needs and wants of the users, this was sometimes difficult for us to communicate and action due to the mismatch in discipline-specific language between us and the digital team [74]. Further information on challenges and opportunities involved in this type of interdisciplinary working will be discussed in a linked SmartADHD publication.

Versions of the chatbot were developed with inbuilt languages (Arabic, Bengali, Polish, and Romanian) which are commonly spoken as a primary or secondary language in England. Emails were circulated to LE collaborators, colleagues, and a public sector organization, asking for their help in finding a young person who spoke any of these languages fluently (as well as English). However, there were no responses from any of these avenues, so we could not test the functionality of the multilingual chatbot feature. These versions should be tested in the future, especially considering the cultural differences that can apply to ADHD. Differences could include cultural understanding of ADHD, including stigma, parental expectations and social attitudes, differences in help-seeking behavior, and barriers or facilitators to accessing support, though there is a paucity of research in this area [75]. Considering that objective queries were simpler for the chatbot to answer than questions very individual to a user, complex discussions around culture and ADHD will be important to test. Throughout the process of developing the chatbot in English, it was important to people with LE that the chatbot used the right language and semantics. Thus, this work to determine word choice and nuanced meaning would be essential to complete in any alternative languages.

Though there was some diversity across the group in terms of gender, ethnicity, and region, it is critical that any future development work prioritizes inclusion of underserved groups. The potential impacts for these groups are significant, and interventions need to address specific barriers to work for these groups. Additionally, all the users had existing background knowledge of ADHD to varying levels and had been involved in research before. None of the users were completely new to learning about ADHD, so they may represent the intended user group to a limited extent. Additionally, users had been involved with this project since conceptualization and were aware that they were using an early prototype that would require extensive development to be deemed fit for use. Because of this, users were also aware that the options for avatar design using Convai were limited, and so our questions and feedback were concentrated on aspects we could address, for example, language. Due to the restricted choices of the platform, time, and resource constraints, we were unable to fully report on user preferences of the avatar in the role of supporting communication.

This study presents early development of a prototype chatbot, as the main aim was to collect opinion-based feedback to inform co-development, not collect generalizable or transferable data. These opinions will be used to inform the direction of development of digital interventions by this team and may also guide future research in the rapidly evolving area of AI-supported digital mental health interventions for young people [76]. As such, claims about effectiveness cannot be made because the current prototype does not represent a finished product. Future testing of the chatbot could include formalized usability testing, safety assessments, and evaluation of feasibility or performance.

A further limitation is that this initial prototype included less functionality than would be present in the final version. For example, this version was trained only on information from the NHS and the NICE guidelines, whereas future versions could be trained on a more extensive range of evidence sources. Additionally, there could be alternative platforms which could be better suited to hosting the chatbot, both from a developer and a user perspective. This work provides proof of concept for future development, but many questions remain about what the eventual product might look like.

All collaborators who took part in think-aloud interviews were offered the opportunity to provide feedback on the findings by acting as a coauthor on this article, and one accepted. Findings from this engagement research are also being communicated via multiple channels to ensure our LE partners remain informed of developments.

This study acts as a proof of concept for an LLM chatbot produced specifically to provide support for young people with ADHD. Young people and parent-carers were positive when user-testing a prototype of the chatbot, particularly about the content. Future iterations should aim to state where specific information is from, test its functionality in other languages and with underserved groups in ADHD care, consider improvements to the avatar and text-to-speech function, and address some of the technical glitches. In the space of digital interventions for ADHD, a chatbot could offer a new solution to the existing problem of gaps in health care provision by combining tailored communication and evidence-based information. Key recommendations are presented in Textbox 2.