Eating disorders are complex conditions characterized by a disturbance in eating behavior that interferes with an individual’s physical and psychosocial functioning [1]. They are a global concern associated with a high fiscal and health burden. Due to their complexity, managing medical and psychiatric risks is an essential part of treatment [2]. Early intervention is considered best practice and is critical for improving eating disorder prognosis [3-5]. There is a well-documented risk of symptom exacerbation as an eating disorder progresses [6-9]. Moreover, research has indicated that more time spent on treatment waitlists is linked to a greater likelihood of treatment discontinuation [10,11]. Thus, interim interventions are urgently required to support people seeking treatment.

Single-session interventions (SSIs) have been proposed as a strategy to provide short-term support for those on waitlists for mental health treatment [12,13]. SSIs typically involve a one-off interaction with a health professional, structured to support individuals in making intentional, positive changes to their mental health [14]. SSIs have helped improve diverse mental health–related outcomes. In a meta-analysis of 50 randomized controlled trials (RCTs), SSIs for youth mental health problems (anxiety, depression, conduct problems, and general distress) demonstrated a significant beneficial effect compared to control conditions, reflecting a small-to-medium overall effect with positive impacts lasting up to approximately 13 weeks or 3 months [15]. For SSIs specifically focused on supporting people experiencing eating disorders, the literature is still emerging; however, SSIs are considered a promising strategy [13,16,17]. For example, preliminary research using a psychoeducation-focused, psychologist-delivered SSI for adults on waitlists for outpatient eating disorder treatment has demonstrated promising results, including reductions in eating disorder symptoms and clinical impairment [18]. While SSIs are traditionally conducted face-to-face, online SSIs have gained traction due to their accessibility, cost-effectiveness, and convenience, demonstrating levels of efficacy similar to face-to-face SSIs [14,15]. For example, a single-session consultation delivered via telehealth by a clinician for individuals on waitlists for outpatient psychotherapy was found to be acceptable, feasible, and effective in the short term, similar to the in-person offering [19]. In addition, this telehealth consultation elicited pre- to postintervention improvements in hopelessness and readiness for change and, at a 2-week follow-up, improvements in anxiety symptoms [19].

Conversational artificial intelligence agents or “chatbots” reflect a unique opportunity for online or digital service provision, and a growing body of research has indicated that chatbots have been positively received in the broader mental health care sector [20-22]. Chatbots offer several benefits, including wider accessibility, instantaneous responses, and low or no cost to users, and can act as a stepping stone to more individualized services [23,24]. Additional barriers to accessing traditional services can include mental health stigma or the lack of services available at the time, where individuals may feel uncomfortable or unable to seek the support they need [23,24]. A recent systematic review and meta-analysis of 18 RCTs evaluated the efficacy of psychotherapy-based chatbots, primarily using cognitive behavioral therapy (CBT) techniques, in alleviating depression and anxiety symptoms among adults. Findings indicated that interaction with the chatbots led to significant reductions in symptoms of both anxiety and depression, with small effect sizes observed [25]. However, it should be noted that all mental health chatbots included in the analysis delivered support over multiple weeks rather than an SSI [25].

Regarding eating disorders, only 4 chatbots that focused on eating disorders and body image have been reported in the literature to our knowledge: Tessa [26], Topity [27], Alex [28], and KIT [23] (subsequently named JEM). Tessa, which was designed to deliver an 8-session CBT-based program, successfully reduced young women’s concerns about weight and shape in an RCT setting and has shown potential to reduce eating disorder onset [29]. However, there were serious concerns about the safety of this chatbot when it was modified for wider public use [30]. More specifically, in June 2023, Tessa chatbot deviated from its preprogrammed answers to provide dieting and weight loss advice, which can be particularly harmful in eating disorder settings. The incident received a great deal of publicity and commentary, and Tessa chatbot has not been implemented in any public setting since to our knowledge [31]. Before the June 2023 incident, Tessa was adapted into a “single session mini-course” that prompted significant improvements in users’ body image and motivation to change body image in a non-RCT setting [32]. Topity, which was created to deliver body image microinterventions, elicited small, notable improvements in state- and trait-based outcomes for body image and associated well-being constructs in an RCT involving adolescents [27]. Components of Alex, designed to target an individual’s motivation to engage with treatment, have been investigated in an RCT; however, the optimized version of the chatbot is still to be tested [33]. KIT/JEM was co-designed to provide body image– and eating disorder-focused psychoeducation and microinterventions to adolescents and young adults and their loved ones, leading to improvements in users’ body image and mood in naturalistic settings in multiple countries [34,35]; however, an RCT has not been conducted.

All these chatbots offer support for individuals with body image concerns and eating disorder pathology. However, these 4 conversational agents were not necessarily originally designed with a specific SSI focus. Moreover, they all address prehelp seeking or connecting users with help services rather than bridging a critical gap by providing timely assistance to treatment seekers who are on waitlists for services.

To overcome these limitations, we have developed, to our knowledge, a world-first chatbot, called ED ESSI (Eating Disorder Electronic Single-Session Intervention), a rule-based agent designed to deliver an SSI to older adolescents and adults on waitlists for eating disorder treatment. The co-design and co-development process of this chatbot was extensive, involving people who had recovered from an eating disorder and health professionals in the field of eating disorders [36]. This 30-minute SSI followed an enhanced-CBT (CBT-E) approach, similar to in-person SSI offerings for eating disorders [18], and focused on psychoeducation and assessment for all types of eating disorders. The broad goals of ED ESSI were to promote improvements in eating disorder and other psychological well-being symptoms while people were waiting for treatment as well as increase motivation for engaging with in-person treatment once it became available. In the co-design process, although there were initial concerns from recovered participants and health professionals regarding the chatbot’s ability to provide appropriate empathy and validation as well as monitor medical and psychiatric risk, strategies were co-designed and implemented [36]. Our co-design participants deemed a rule-based approach, where the chatbot could only deliver preprogrammed responses, as the safest strategy. The co-design process also involved reports that seeking support for eating disorder care was highly challenging, given the stigma surrounding eating disorders in particular, so offering a nonhuman alternative that still had an interactive chat interface with a range of activities was seen as optimally engaging [36]. The evaluation of the final prototype version of ED ESSI was positive from people with a lived experience of an eating disorder and health professionals [36].

Although preliminary feasibility and acceptability for ED ESSI have been demonstrated [36], we are yet to determine the chatbot’s effectiveness in an RCT and, in addition, in real-world clinical waitlist settings. Hence, we aimed to conduct the first RCT to determine the effectiveness of ED ESSI for people aged ≥16 years on the waitlist for eating disorder treatment at outpatient psychology clinics around Australia. It was hypothesized that participants who engaged with the chatbot SSI would experience greater reductions in their eating disorder pathology (primary outcome) after 1 and 3 months compared to a control group that received similar web-based psychoeducation information but not delivered by a chatbot in the form of an SSI. Furthermore, it was predicted that the SSI chatbot use would elicit greater improvements in broader facets of well-being (secondary outcomes), specifically psychosocial impairment due to eating disorder symptoms, mood (depression, anxiety, and stress), and motivation for treatment after 1 and 3 months compared to the control group.

This study was a 2-arm, parallel-assignment, and nonblinded RCT with an intervention group and a control group. Screening at least 70 participants (allowing for 15% dropout [37]) would allow for the detection of small to moderate Cohen f effect size changes (Cohen f=0.17) in eating disorder pathology (primary outcome) between groups, in accordance with findings from Fursland et al [18], with a power of 0.80 and a significance level of P=.05 [38]. Simple 1:1 randomization was performed using a Python (Python Software Foundation) script. Allocation was not concealed, allowing investigators to assign participants to either group based on the randomization result.

The study was prospectively registered on the Australian New Zealand Clinical Trials Registry (ACTRN12623000680662) [39] and approved by the Monash University Human Research Ethics Committee (ID 38277). Participants provided written informed consent to be enrolled in the study via a web-based survey through Qualtrics (Qualtrics International Inc). Participants were provided with the ethics-approved study information by the recruiting clinics immediately upon joining the clinic’s waitlist, and those aged 16 to 17 years were considered sufficiently mature to provide consent for themselves under our ethics approval. Participants were assigned a unique code to include in each of the 4 time point surveys to deidentify their responses. All participants were reimbursed after the final study time point with an Aus $50 (US $32.36) voucher.

Participants were recruited from outpatient psychology clinics or practices around Australia from September 2023 to September 2024. Participants had been referred by their general practitioners or primary care physicians before seeking care from these clinics or practices. All clinics involved charged fees for their treatment services, which is very common for eating disorder treatment in Australia [40], but no fees were charged to participants for study involvement. Participant inclusion criteria were (1) being aged ≥16 years and (2) being registered on an outpatient psychology clinic waitlist to commence eating disorder treatment following a referral from a general practitioner or primary care physician. Exclusion criteria were (1) the lack of access to a working email address or electronic device and (2) high medical and/or psychiatric risk, as assessed and documented by a general practitioner or primary care physician as a standard part of referral to an outpatient psychology clinic for eating disorder treatment. There was 1 deviation in the inclusion or exclusion criteria for the preregistered protocol [39], namely, participants with any eating disorder diagnosis were eligible.

After the provision of informed consent and study enrollment (refer to Ethical Considerations section), the research team members performed the randomization, and participants were informed of the group to which they had been allocated—the chatbot intervention or the control group. In an attempt to reduce the effect of expectation bias, the research team intentionally masked our predictions that the chatbot would elicit greater clinical outcome improvements compared to the control. The research team members remained unblinded, in addition to the participants. The uptake of the chatbot intervention or control information was monitored via email, with the participants confirming completion via email to the research team. Furthermore, we were able to confirm the participant’s use of the chatbot through their log-in history in the chatbot web portal.

SPSS (version 30; IBM Corp) [48] was used for quantitative statistical analysis. Descriptive statistics were used to analyze demographic and clinical characteristics. To determine whether there were differences between groups at baseline (or other time points as required), independent 2-tailed t tests were used for continuous data and Fisher exact test for categorical data. To examine the effects of the chatbot versus control intervention across the 3 or 4 time points on the primary and secondary outcomes, mixed ANOVA with the intervention group as the between factor and time as the within factor were conducted with partial eta squared as the overall effect size. Significant group×time interaction effects were further analyzed using paired sample 2-tailed t tests, with a conservative Bonferroni correction applied to mitigate against the increased risk of a type I error for the multiple pairwise tests [49] and Cohen d used as the effect size [50]. Any missing data were handled with listwise deletion. The open-ended qualitative text responses addressing chatbot feedback and reasons for not commencing treatment were analyzed using an abbreviated content analysis approach [51], given the brevity of participant responses (30 words maximum).

Of the 72 people screened for eligibility, 63 (88%) were admitted and 60 (83%) ultimately completed the study (refer to Figure 2 for an overview of the participant flow). Demographic characteristics at T1 or baseline are provided in Table 1, where there were no significant differences between the chatbot and control groups for any of the characteristics measured (all P>.05).

The descriptive statistics for the chatbot and control groups over the time points are outlined in Tables 2 and 3 (refer to Multimedia Appendix 1 for all EDE-Q and CIA subscale descriptive statistics). There were no significant differences between the chatbot and control groups for any of the measure outcomes at T1 or baseline (all P>.05). At baseline, the participant scores were (on average and as expected) well above the EDE-Q global score required to meet eligibility for specialized eating disorder treatment plans in Australia. According to the mean CIA global score, the participants, on average, met the threshold for clinically significant impairments in their lives owing to eating disorder symptoms. The mean depression, anxiety, and stress symptoms were in the moderate, severe, and moderate ranges, respectively. Furthermore, in terms of motivation, the participants rated it highly important that they change, and scores were moderate for confidence in their ability to change and comfort toward treatment at baseline.

As predicted, there was a significant group×time effect for EDE-Q global scores (F_2,116=5.98; P=.003; η_p²=0.09). For the chatbot group, there was a significant reduction in eating disorder pathology from baseline to 1 month (P<.001; Cohen d=0.40) and another significant reduction from 1 month to 3 months (P=.018; Cohen d=0.29). At 3 months, the mean score had reduced such that it was only slightly above the eligibility criterion for specialized eating disorder treatment plans in Australia (≥3.0). For the control group, although there was a trend toward a small reduction in scores over the 3 months, these were not significant, unlike for the chatbot group (T1 vs T3: P=1.000; Cohen d=0.07 and T3 vs T4: P=.712; Cohen d=0.14).

In concordance with prediction, there was a significant group×time effect for CIA global scores (F_2,116=4.90; P=.008; η_p²=0.08). The psychosocial impairment scores for the chatbot group significantly reduced from baseline to 1 month (P<.001; Cohen d=0.53) and had another significant decrease from 1 to 3 months (P=.001; Cohen d=0.37). There was no significant change from baseline to 1 month for the control group (P=1.000; Cohen d=0.07) but a significant reduction from 1 month to 3 months (P<.001; Cohen d=0.35). The chatbot group’s CIA global scores at the 3-month postbaseline time point were still significantly lower than the control group as predicted (P=.023; Cohen d=0.61). Both groups remained in the clinically significant impairment range (mean score ≥16.0) at the final time point at 3 months.

Unexpectedly, there was no significant group×time effect for the DASS depression subscale (F_2,116=1.40; P=.251; η_p²=0.02). However, the depression scores decreased for both groups from baseline to 3 months, and upon further investigation, only the chatbot intervention elicited a significant decrease (P=.002; Cohen d=0.35 and control: P=.382; Cohen d=0.16). The mean depression scores for both groups remained in the moderate range at 3 months.

As expected, there was a significant group×time effect for the DASS anxiety subscale (F_2,116=3.61; P=.040; η_p²=0.06). The anxiety scores for the chatbot group significantly reduced from baseline to 1 month (P<.001; Cohen d=0.35) but no significant change from 1 to 3 months (P=.248; Cohen d=0.15). For the control group, although there was a trend toward a small reduction in scores over the 3 months, these were not significant (T1 vs T3: P=1.000; Cohen d=0.05 and T3 vs T4: P=.419; Cohen d=0.12). The mean anxiety score for the chatbot group was reduced to the moderate range by T4, while the control group remained in the severe range.

Although it appeared that the scores for stress decreased for both groups, slightly more so for the chatbot group, there was neither a significant group×time effect (F_2,116=0.56; P=.575; η_p²=0.01) nor were there significant reductions for either group from baseline to 3 months (chatbot—T1 vs T3: P=.420; Cohen d=0.18 and T3 vs T4: P=.590; Cohen d=0.13; control—T1 vs T3: P=1.000; Cohen d=0.01 and T3 vs T4: P=.225; Cohen d=0.23). The mean scores for both groups remained in the moderate range at the final time point.

Participant-rated importance to change was highly ranked for both groups across all 4 time points; however, in contrast to prediction, there was no significant group×time effect (F_3,174=0.29; P=.830; η_p²=0.01) and no significant changes for either group across the time points (all P>.05). Nevertheless, there was a significant group×time effect for confidence to change (F_3,174=3.46; P=.018; η_p²=0.06). There was a significant increase in confidence for the chatbot group from T1 to T2 (immediately postchatbot use; P<.001; Cohen d=0.74) but a significant decrease from T2 to T3 (P<.001; Cohen d=0.48) and no change overall from T1 to T4 (P=.328; Cohen d=0.32), hence staying at a moderate level of confidence. There was no significant change in the confidence for the control group for all time points (all P>.05), remaining at a moderate level. Finally, for comfort toward treatment, while there (similarly) appeared to be an increase for the chatbot group from T1 to T2 and then a decrease, the group×time effect was not significant (F_3,174=2.08; P=.105; η_p²=0.04), and no significant changes were observed for either group across the time points (all P>.05). It appeared that the comfort toward treatment remained at a moderate level for both groups across the 4 time points.

The mean score for the SUS for the chatbot group was 83.8 (SD 12.5), which equated to an “excellent” or “A” grade category [47]. In total, 40% (12/30) of the participants used the opportunity to provide a very brief comment in the open-ended text option about their experience with the chatbot. These were predominantly positive comments about the “usefulness” and “helpfulness” of the chatbot (10/30, 33%). A small number of participants (3/30, 10%) reported that they would have appreciated the option to go backward in the chatbot conversation to change or clarify some of their earlier answers in the SSI. This backward function was not available in the ED ESSI version used in the RCT.

All participants were on a waitlist for treatment at T1 (baseline) and T2 (immediately after intervention), but participants were also asked whether they had commenced treatment at 1 month (T3) and 3 months (T4). If not, an opportunity to provide a very brief update on their treatment commencement was offered, and if they had started treatment, participants nominated how many treatment sessions they had completed.

At 1 month, 50% (15/30) of the participants in the chatbot group and 47% (14/30) of the participants in the control group had started treatment, so there were no significant differences between the groups according to Fisher exact test (P=1.000; ϕ=0.03). Among the chatbot group’s 15 participants who had not started treatment, 13 (87%) reported that they were still on the waitlist, and 2 (13%) reported that they had “rescheduled their first session” owing to a scheduling conflict. In the control group, 15 (94%) of the 16 participants reported that they were on the waitlist, and 1 (6%) participant reported that they had to reschedule their first session. For the participants who had started treatment, there was no significant difference between the number of treatment sessions attended for the chatbot (mean 2.3, SD 1.4) and control (mean 1.9, SD 0.7) groups (t₂₇=0.81; P=.424; Cohen d=0.30).

At the 3-month postbaseline time point, 28 (93%) of the 30 participants had started treatment in the chatbot group, while only 21 (70%) of the 30 participants in the control group had started treatment, which represented a significant difference (P=.042; ϕ=0.30). The 2 participants in the chatbot group who had not started treatment clarified that they had been offered a first appointment, but “financial constraints” or “money concerns” meant that they had not pursued treatment. In the control group, similarly, all participants had been offered a first appointment, with all (9/9, 100%) participants stating that financial constraints or money concerns had prevented their engagement and 4 (44%) participants additionally mentioning “concerns” about their “confidence in [their] ability to recover.” For the participants engaged in treatment, similar to the 1-month postbaseline time point, there was no significant difference between the number of treatment sessions attended for the chatbot (mean 4.9, SD 2.4) and control (mean 4.3, SD 1.1) groups (t₃₅=1.154; P=.256; Cohen d=0.32).

This study aimed to examine the effectiveness, for the first time, of a novel SSI chatbot, ED ESSI, for older adolescents and adults on the waitlist for outpatient eating disorder treatment. Through an RCT in clinical settings, we demonstrated that chatbot use induced early reductions in eating disorder pathology (primary outcome) and psychosocial impairment due to eating disorder symptoms (secondary outcome) at 1 month and then further reductions after that time to 3 months with small to moderate effects. The control group who received similar information as in the chatbot SSI group did not show these decreases, although there was a small, significant reduction in psychosocial impairment by 3 months in the control group.

For the other (secondary) outcomes, the chatbot-induced early anxiety reductions were detected at 1 month, such that symptoms were classified in the moderate range (reduced from severe), and depression symptoms decreased by 3 months (moderate range); however, there was no change in stress symptoms (moderate range). There were no changes in the mood measures for the control group. Unexpectedly, there were minimal differences between the chatbot and control groups for measures of motivation. There was a moderate to large increase in confidence for the chatbot group immediately after chatbot use, but this was not maintained. Nevertheless, these ratings were in the moderate to high range for both groups at the start of the RCT, so there was less opportunity for improvement across the trial.

Overall, it appeared that the experience with ED ESSI promoted early gains (not evident in the control group) that were able to be leveraged once the individual left the waitlist and started treatment with a psychologist. The results suggested that those who used the chatbot were significantly more likely to start treatment by 3 months, possibly because of the early improvements they experienced. Importantly, chatbot users were still in the clinical symptom and psychosocial impairment ranges by the 3-month postbaseline time point, although much closer to entering subclinical ranges. Nevertheless, it was not the intention that a 30-minute SSI delivered by a chatbot would alleviate all symptoms; rather, it would provide individuals with a “head start” on treatment and assist with the transition to in-person treatment, which is recommended to be between 20 and 40 sessions with a psychologist for CBT-E [42]. Even though it seems like waitlist lengths for eating disorders peaked during the COVID-19 pandemic [52] and have subsequently reduced, positive and effective early treatment experiences remain crucial to implement, and ED ESSI is a promising, convenient, and accessible strategy.

ED ESSI chatbot is an important addition to the short but emerging list of chatbots focused on eating disorders and/or body image (and mental health chatbots in general) that have been found to induce small to moderate improvements in the mental health of users [25,27,29,32,33]. ED ESSI was able to induce these improvements in a novel way, more specifically, in a clinical sample and in clinical settings rather than community-based studies. We are seemingly still relying on simpler rule-based conversational agents in the field of eating disorders, most likely owing to the high level of medical and psychiatric risk involved with these disorders [53] and the highly publicized harms when eating disorder chatbots have deviated from preprogrammed answers (eg, Tessa chatbot [30]). Potentially, the further use of ED ESSI in clinical settings will help to rebuild trust in patients and health professionals and pave the way for more advanced generative artificial intelligence agents to be introduced more broadly [31]. The positive results for ED ESSI are also another demonstration of a successful co-design process with key stakeholders, particularly those with lived eating disorder experience, leading to efficacious results with actual patients, where ED ESSI chatbot’s usability was rated as excellent [23,28,36]. Moreover, the results represent a further progression in the emerging field of SSIs for eating disorders [13,17,32,54]. Our chatbot SSI was broadly based on a psychologist-delivered SSI for adults seeking outpatient eating disorder treatment where the in-person SSI induced reductions in eating disorder symptoms (Cohen d=0.27), psychosocial impairment (Cohen d=0.29), and depression symptoms (Cohen d=0.21) as well as increased the likelihood of patients entering treatment (70%) [18]. As described in the Results section, this study’s chatbot SSI seemed to yield superior results in terms of effect sizes (Cohen d=0.35-0.53) for the same outcome measures and treatment uptake (28/30, 93%) as well as significant decreases in anxiety symptoms. Furthermore, Fursland et al [18] reported a significant increase in BMI units (0.9) for patients who were underweight (BMI<18.5 kg/m²) after SSI completion, which was not examined in this study for this cohort of patients and should be the subject of future research.

As proposed by Fursland et al [18], we suggest that the engaging psychoeducation and assessment delivered by the chatbot resulted in positive changes to patient attitudes and behaviors, and by delivering this information so soon after joining a waitlist, we were able to capitalize on the motivation for treatment. Indeed, there was an initial moderate to large increase in confidence in the ability to change immediately after using the chatbot, and although this was not maintained at subsequent time points, perhaps this was a particularly crucial time frame to start promoting positive changes in attitudes and behaviors. Furthermore, as found by Messer et al [16], there was no reduction in any of the moderate to high motivation measures for the chatbot (or control) group over the course of the study, so the SSI experience did not have negative impacts. Indeed, patients who are seeking treatment and are willing to be part of a research study where “extra” treatment is offered are likely to be highly motivated.

The RCT was implemented in naturalistic settings in outpatient psychology clinics around Australia, and the attrition rate throughout the study was very low, particularly compared to other SSI studies in the same field of research (eg, [16]). The highest level of attrition observed in our study occurred at the screening survey stage rather than at any point after participants had completed the chatbot or control intervention. From our discussions with the clinics involved in the study, it is possible that patients may have mistaken our study screening survey web link for the clinic’s standard intake registration web link, with all information being sent in a single email or correspondence. Thus, individuals may have stopped completing the screening when they realized that it was part of a separate research study rather than their core treatment at the clinic. Another strength of our study and the chatbot’s design [36] was the ability to assess and alert when a user’s psychiatric and/or medical risk was detected. Although not experiencing high medical and/or psychiatric risk was an eligibility criterion for our RCT, this level of risk can readily change while waiting for and during the course of treatment [53]. The chatbot alert system was triggered for 1 participant, and follow-up with the research and clinical team was performed. The participant reported that they triggered the alert “by accident,” but it was, nevertheless, a demonstration of the alert system functioning appropriately in a naturalistic setting. Owing to the previous instances of malfunctioning chatbots in our field of eating disorders (eg, [30]), demonstrating no adverse events is especially important. A further strength of our study was that all participants involved were offered treatment by the 3-month postbaseline time point, and a proportion (29/60, 48%) of participants had started by 1 month. This not only allowed for the investigation of receptiveness to treatment and reasons for not commencing, which were primarily financial in nature, but also revealed loss of confidence in the ability to recover for some control group participants. Financial concerns were most likely reflective of the increase in the cost of living in Australia and globally [55] and potentially emphasize the importance of free or low-cost online eating disorder treatment offerings.

There are also limitations to this research that must be considered. First, we did not specifically examine whether the chatbot or control group engaged with the psychoeducation content through, for example, a knowledge check immediately after intervention. Furthermore, a more optimal control group should be used in future research. For example, the control participants could have been provided with a sham chatbot, which included engagement-related factors similar to ED ESSI chatbot and an active component that differed from ED ESSI chatbot’s psychoeducation and assessment approach, such as self-compassion [56]. Such a strategy will also likely assist ED ESSI with broader dissemination and regulatory approval internationally [56]. Furthermore, the participant sample consisted predominantly of White, younger cisgender women. This is reflective of the demographic characteristics of individuals who most often seek treatment [57]; however, there is great diversity in the people who experience eating disorders [58,59]. It will be important to examine the effectiveness of the chatbot in more diverse demographic samples. Potentially, ED ESSI may be even more appealing to diverse samples owing to high accessibility, convenience, and the fact that the chatbot has been explicitly programmed not to stigmatize the user, as has been unfortunately documented for some health professionals working in the field of eating disorders [60,61]. However, it will also be critical to explore other implementation strategies for ED ESSI in future research, for example, access via a website or social media chat functions in accordance with how we have previously used JEM in multiple countries [34,35]. Such strategies are possibly more likely to reach diverse users, as we found for JEM and should be investigated for ED ESSI. Furthermore, we did not have the statistical power to investigate the effectiveness of the chatbot by, for example, different eating disorder diagnoses or BMI categories. There was some diversity in these areas in the RCT, and future research would usefully focus on deciphering whether chatbot use is more optimal for certain clinical presentations and demographic characteristics. Moreover, it appeared that our participants were moderately to highly motivated at the commencement of the study, which may explain why they were receptive to participating in our research. It will be vital to determine the chatbot’s effectiveness in individuals experiencing lower levels of motivation. Finally, although a 3-month follow-up time is fairly standard in the SSI literature [15], a longer follow-up time would be beneficial in future research. It is possible that chatbot users required less time and fewer treatment sessions with a psychologist and other health professionals in order to reach recovery. However, it is also possible that those who did not use the chatbot “caught up” in terms of treatment gains after the 3-month time point and ended up at a similar phase of recovery in a similar amount of time.

In conclusion, a novel chatbot-delivered SSI for older adolescents and adults on the waitlist for outpatient treatment for any type of eating disorder effectively reduced eating disorder pathology, psychosocial impairment, depression, and anxiety at 1 and 3 months in an RCT. Furthermore, chatbot usability was excellent, and users were more likely to enter in-person treatment when it became available. ED ESSI potentially represents an effective, accessible, convenient, and scalable form of early intervention for people waiting for eating disorder treatment; however, further research is required in more diverse samples and with longer follow-up periods.