Abstract
I wear many hats professionally. My day job is working with The Center for Health Design as Vice President for Research, but my other hats include being the Vice Chair of the 2022 and 2026 Facility Guidelines Institute (FGI) Health Guidelines Revisions Committee, a board member (2019–2023) and past president of the American Institute of Architects Academy of Architecture for Health (AIA AAH), and now, the latest addition to the talented team of editors at Health Environments Research & Design Journal. In these varied roles, I have the opportunity to attend many industry events (conferences, workshops, webinars, and meetings) as a speaker, moderator, emcee, and participant. At events I’ve attended in the past year, a common question addresses the use of Artificial Intelligence (AI). How many people are using it? How many people are excited about it? How many people are terrified? The preponderance of answers usually lays somewhere in the middle—we’re still trying to figure out the best use in practice, and we are equally excited and terrified.
Making Research Actionable
At The Center for Health Design, our research team has done much work over the years to make research more accessible and actionable. The Knowledge Repository (funded by our partners at AIA AAH, FGI, and The American Society for Health Care Engineering (ASHE)) includes more than 6,000 research references related to healthcare design across the continuum of care. Over time, we have created different forms of research summaries—issue briefs on topics such as noise or infection control, key point summaries that provide a layman’s translation of individual papers entered into the Knowledge Repository database, and slidecasts, where our research team creates a 5-min narrated PowerPoint of a research study. Can AI help?
The AI (R)Evolution in Healthcare
With AI shifting from a majority academic pursuit to entry into the larger public realm in late 2022 with the launch of ChatGPT (Chat Generative Pre-trained Transformer by OpenAI), we are entering a new era. There has been a multitude of papers and press about the use of AI in healthcare and the diagnostic accuracy compared to physicians. As the technology quickly advances, a physician friend described it to me as “freakishly scary.” It might not be there yet, but being able to analyze more data than humanly possible has advantages.
In architecture, firms are experimenting with AI for everything ranging from concept creation and text-to-image visualization to building information modeling (BIM) integration and project management. Despite only 5% of firms (overall) actively using AI in daily operations, nearly 70% of firms are intrigued by or investigating AI integration into their practice (AIA Economics & Market Research Group, 2023; Wroten, 2023). This is expected to grow dramatically in the next few years, with the most anticipated application being related to design options (AIA Economics & Market Research Group, 2023). These architectural AI explorations have been described to me as “like watching kids in a candy shop.”
Creativity, Plagiarism, and AIgiarism
Overarching concerns with AI for many creative fields include how generative AI models are trained. The use of proprietary work has resulted in some high-profile arguments and legal cases surrounding derivatives and style imitation in AI-generated music (e.g., Drake and the Weeknd, “Heart on My Sleeve,” 2023) and artwork (e.g., Andersen v. Stability AI et al.). Such questionable outputs were dubbed AIgiarism (AI-assisted plagiarism) in late 2022 (Dragidella, 2022). We will have to see how this evolves with architectural design—both the use of someone else’s designs to train AI and also whether a firm’s AI-generated content is protected by copyright.
AI for Research Translation
Where does this leave us with translating research for healthcare design? Slidecasts were an outcome of our first Research Matters presentation in 2017. A full room eagerly listened to the rapid-fire format: 10 studies in 50 min. People came to the stage to share their sheer joy in being able to learn about research studies in such a condensed format. Questions were asked about how long it takes to create those 5-min summaries. Short answer—a long time. Unfortunately, brevity doesn’t happen quickly. Think of the quote often attributed to Mark Twain (and others in varying formats), “I didn’t have time to write you a short letter, so I wrote you a long one.”
Recently, our research team has been asking questions about how AI might be used. Can we shorten the lengthy process of creating brevity? What does generative AI mean for summarizing research papers and topics? This editorial is a work in progress (not a systematic approach), but I will share my brief foray (so far) into AI along with my concerns and hopes.
Summarizing Individual Papers
I used two AI platforms to inquire about three specific papers using the title (in quotes) and the primary author. One paper was open access, one published behind a paywall but made available through PubMed Central (PMC) through federal funding requirements, and one available behind a paywall. ChatGPT acknowledges it does not have access to research papers behind the paywall of publisher websites. In asking ChatGPT about specific titles and authors, the response starts with phrases such as “I’m sorry, but I don’t have access to specific articles or their contents, and I couldn’t find any information related to an article titled….” In Round 2 of pressing for information from ChatGPT using the title without quotes, ChatGPT offered the following: I apologize for any inconvenience, but I still do not have access to specific articles, and it appears that you are referring to a hypothetical article title rather than a real one. If you are looking for information on interventions to improve children’s access to mental health care, I can provide a general overview or discuss strategies and approaches commonly used to address this issue.
Google Bard provided summaries that appeared to be largely generated from the abstract, with some additional information for each of the three papers. The first paper (open access) included a link to the paper at the end of the summary. The second summary (available through PMC) did not cite a source. The third (behind the paywall) provided a link to a different article in a foreign language. (More on sources later in this editorial.)
There was an issue with the additional information; it was not always entirely correct. I say entirely, because it required a significant level of comparison to figure out whether the provided summary was accurate or different enough that it was misleading or no longer correct. While the summaries included information found in the paper, there were instances where the content provided was not the author’s work. Additionally, any important nuance of detail was lost.
General Overviews
As ChatGPT offered to provide a more generalized summary, I entered a query to provide an overview of design strategies and approaches commonly used to address adolescent mental health. The results looked convincing with a bulleted list of 13 considerations. The following is a portion of what was generated by ChatGPT: Design strategies and approaches used to address adolescent mental health in various settings, including schools, healthcare facilities, and community spaces, aim to create environments that promote well-being, reduce stigma, and provide supportive resources. Here are some common design strategies and approaches:
Safe and Welcoming Spaces: Ensure physical safety with secure entrances and exits but also make the environment feel warm and welcoming. Use calming color schemes and natural lighting to create a comfortable atmosphere. Provide clear wayfinding to reduce stress and anxiety.
Privacy and Dignity: Offer private counseling and therapy rooms to protect the privacy and dignity of adolescents receiving care. Provide private sleeping areas in inpatient settings to encourage personal space.
Therapeutic and Sensory Design: Incorporate sensory-friendly elements such as quiet rooms for relaxation and sensory gardens for calming outdoor spaces. Use materials and textures that are soothing and minimize sensory triggers. Integrate art and nature-based elements to promote emotional expression and healing.
… These design strategies and approaches aim to create a supportive, therapeutic, and inclusive environment that helps adolescents access and engage with mental health services effectively, reducing stigma and promoting overall well-being. The specific strategies employed may vary based on the context and the particular needs of the adolescents being served.
Hallucinations, Confabulations, or Hallucitations?
The lack of transparency about the sources of information ties into the next concern—there are no citations to understand where the information is sourced. Are any author’s works being plagiarized by virtue of not being properly cited? Is it correct? ChatGPT’s first response when I asked for citations for a similar search about design for behavioral health follows: I apologize for any confusion, but as of my last training data in September 2021, I don’t have direct access to real-time or external databases, including academic citation databases. Therefore, I can’t provide specific citations to recent research or publications on design for behavioral health.
This is what I would do without AI, so this has not reduced my time or effort. On the second query to press ChatGPT to provide an answer, six sources were cited (contrary to the previous reply), although most are not specific to behavioral health. I include the citations (without the abbreviated description provided) as readers will most likely recognize some of the authors and publications. Please note that these references are from before my knowledge cutoff date in September 2021, and I recommend checking for more recent research in academic databases for the most up-to-date information: Ulrich, R. S. (1991). Effects of interior design on wellness: Theory and recent scientific research. Journal of Health Care Interior Design, 3, 97–109. Heerwagen, J. (2001). Design, productivity, and quality of the indoor environment. In J. P. C. L. de Groot, H. B. Marquardt, V. V. Burgen, M. Goeden, & S. Loosemore (Eds.), Proceedings of Indoor Air 2002 (Vol. 2, pp. 3–13). Devlin, A. S. (2013). A review of the role of the healthcare environment in the satisfaction of patients and their families. Health Environments Research & Design Journal, 6(3), 95–119. Shepley, M. M., Pasha, S., & Wills, A. (2017). Design for mental and behavioral health. Designing for Health: Healing Environments, 169–194. Hassan, A., & Hassan, T. M. (2019). Designing for psychological well-being: A behavioral architecture approach. IGI Global. Clements-Croome, D. (2006). Creating the productive workplace (2nd ed.). Taylor & Francis.
The results presented above highlight the challenge. I am familiar with Citation 1, and it is a paper published many years ago by The Center for Health Design. Citations 2 and 5 could not be located through a variety of means, and there are no digital object identifiers (DOIs) to aid in location. I could not determine who the authors were for Citation 5, and the work is not listed on the publisher’s website. I know the author for Citation 2, and while she has a lengthy list of publications in her Google Scholar profile, Citation 2 is not among them. Citation 3 is not real as that volume, issue, and page numbers correspond to a study about the relationship between flooring and outcomes by Harris and Detke. Citation 4 is a book that I own. The book does not include the referenced chapter.
In preparing this editorial, it was natural for me to see if others had published on this topic. This same phenomenon has been documented by others in medicine. In one example, a researcher asked ChatGPT to generate 30 papers cited with a minimum of three references. Nearly half of the citations were fabricated, and a similar percentage were actual citations with inaccurate information (Bhattacharyya et al., 2023). Perhaps a more serious concern arises with false information as found by a researcher who was excited to have ChatGPT provide “unexpected suggestions” about the brain regions related to antipsychotic treatment. Unfortunately, upon further inquiry about the sources of information (offered as citations with credible authors and journals), the sources were found to be unrelated or fake (Emsley, 2023). Furthermore, Emsley queried information about known papers to find that ChatGPT incorrectly reported content found in the publications.
Fake information, like what I and other authors have found, is often called AI hallucination, although there has been some debate about the term. Hallucinations are considered false sensory perceptions, while confabulations are fabricated but “logically generated semantic statements” (Hatem et al., 2023). Brender (2023) agrees that while the use of the term hallucination in AI may not be best, the term confabulation could also have implications. Medically, the term confabulation implies aims, goals, and intent, which might suggest AI sentience. Specific to the sourcing of information, my preferred term was shared with me by The Center for Health Design’s Advisory Board member, Mickey McManus, who used the term hallucitation—a creative word to differentiate falsified sources from other types of AI fabrications.
Conclusion
I will continue to explore the capabilities of AI, including generating summaries created by loading a pdf version of a paper into an AI platform (also powered by ChatGPT models). A clear ethical challenge in this context is uploading copyrighted content, which will be used to further “train” the algorithm. The summarized content may also not be correct. There are other platforms that offer different types of research consolidation and organization. Perhaps these explorations will be the topic of a future piece, although to be fully functional, subscriptions are required.
At this point, I would not trust AI to summarize research given some of the identified challenges. Maybe it could be a place to start if you do not know what you’re looking for (e.g., related search terms), but just like the surprised researcher asking about brain regions, the results might not be true. In that sense, you need to know quite a bit before diving into these waters. Everything needs to be checked and verified.
For me, it took more time to check AI than to do the work in the first place. For now, I would rather spend my time reading the author’s own words that include the details, limitations, and figures so that I can draw my own conclusions about strength and relevance. When trying to create brevity, it is important to consider context, and in my limited experience, AI creates content that is too general to evaluate. If we can eventually use AI to reliably “write a shorter letter,” it would be welcomed. I look forward to continued discourse and watching the rapid evolution of AI with hopes that it can be reliably used in an evidence-based design process in the future.
