A first group of ASD studies targeted nonverbal communication skills. In joint-attention skills training, virtual characters [17,22,24-26] or a robot [27] would nonverbally cue targets by pointing or gazing at them, after which children were instructed to pay attention to the targets [17,22,24-27]. Imitation skills were taught by asking children to repeat nonverbal gestures made by robots [28-31]. Other applications focused on tactile interaction by letting children play with robots equipped with tactile sensors [31-34], turn-taking behavior through playing games with a virtual character [17] or robot [30], and facial and emotion recognition by reconstructing faces of dynamic photographs [22].

Applications focusing on verbal skills used ECAs to teach children communication skills such as general conversation [35,36] and antibullying strategies [37], and stimulated cooperation in multiplayer games [21,38,39]. Lastly, a job interview training application allowed users to practice with a virtual job interviewer [23]. It is in this group of ASD studies that we found the only three applications focusing on adults [23,35,39].

Four of the educational aid applications involved virtual tutors, targeting vocabulary [40], daily-life skills [41], general educational needs of children with ASDs [42], and comprehension of idioms [43]. One study used a humanoid robot to teach children about body postures and aid them in their sense of body consciousness [44].

From a technological perspective, the focus in many of these studies was on letting ECAs communicate with users through human communication modalities, most notably speech (n=18), facial and gaze expressions (n=15), and hand and body gestures (n=14). A notable exception was touch (n=5), which could only be used by human users.

Few studies employed user models to personalize subsequent interactions. Static user models were mostly used to call the user by name [24,25,27,42], and only two studies employed dynamic user models to personalize subsequent interactions. In one study [26], emotion recognition was used to structure the narrative of a game, and in another [23] the interactions were structured according to the user’s performance and level of rapport with the virtual job interviewer. Only two studies allowed users to enter into a dialog with the ECA, both using menu-based dialogs [23,36].

Although the most commonly studied outcomes were behavioral (n=17), and most of the studies (n=21) were conducted with clinical samples, the sample sizes were generally very small (mean 14.6 [SD 14.2]), and in many cases the behavioral outcomes were short term and restricted to experimental settings (eg, a researcher’s observation of child-robot interaction during an experiment [31]). Most studies were still in the development (n=13) and piloting (n=11) phase.

Two ASD studies had moved beyond the development and piloting phase. In the first study [23], the Web platform virtual reality (VR) job interview training was evaluated in a randomized controlled trial (N=26). Users were subjected to an interview with a job interviewer, and could ask a coach for feedback regarding their performance. Users could interact with the interviewer through a menu-based system, but they were also given the option of speaking their choices out loud to help them practice their verbal skills. The interviewer could be configured to have different personalities (from friendly and easy-going to mean and asking illegal questions), and based her responses on the user’s answers and a user model that kept track of the level of rapport between the two. The interviewer used speech output and conveyed dynamic emotional states through facial and gaze expressions. There was a significant improvement in the user’s researcher-scored interview skills in a role-played interview, as well as self-confidence, compared with a control group that received no intervention. Although this result is promising, generalizability to real-world settings and the application’s effectiveness compared with conventional job interview skills training programs remain unclear.

The second study used a computer program to improve joint attention, and emotion and face recognition in children with autism [22]. It involved exercises with dynamic realistic photographs of human faces, “coming to life” after successful completion. This intervention also included an animal avatar coach embodied by a realistic photograph that provided additional motivational reinforcement. Following a randomized controlled study (N=49), the intervention was found to be more effective in comparison to a control condition in which children used drawing software. Both children with high- and low-functioning autism improved in terms of emotion recognition and observed social skills, while children with high-functioning autism also improved in facial recognition. Similar to the other evaluation study, it remains unclear how the intervention would compare with conventional interventions targeting similar social skills.

The CBT-based applications targeted symptoms of depression in general, with a virtual coach guiding people with depression through a Web-based intervention [45], a photograph of a clinician embodying weekly feedback in a Web-based intervention [46], and a fantasy character guiding users through a serious game [19]. In [47], requirements for the virtual agent used in [45] were determined through a focus group study.

A second group of applications explored elements of counseling with a virtual agent, more specifically the elicitation of self-disclosure, that is, getting people to talk about their problems, by a virtual counselor in an open-ended dialogue [20], the elicitation of self-disclosure as well as the provision of relevant information by a Web-based virtual counselor among active soldiers, war veterans, and their families [48], and diagnosis by a virtual therapist guiding users through a Web-based version of the Beck Depression Inventory questionnaire [49].

Self-management skills were targeted by an application that supported hospitalized patients during their discharge procedure [50], and by a serious game in which people with depression could practice communicating about their health with a virtual doctor [51].

The last study concerned the same job interview training application used in [23], this time targeting people with other psychiatric disabilities, including depression [52].

Looking at the use of human communication modalities, the most technologically advanced developments take place in counseling interventions from studies conducted by the Institute for Creative Technologies, associated with the University of Southern California (USC-ICT). Over the years, they have developed an extensive framework that allows users to communicate with ECAs in a natural manner through verbal and nonverbal behavior. In one study [48], users could communicate by textual natural language. Using speech and synchronized nonverbal behaviors, the ECA was able to take initiative in the conversation and probe for information related to depression and PTSD. An even more advanced approach in terms of user input was taken in another study [20], in which users’ speech and nonverbal behavior was taken as input, and used to engage them in open-ended dialogues aimed at self-disclosure about psychological problems. The ECA was endowed with a set of fixed utterances and interview questions, applied back-channel behaviors (eg, saying “uhuh” and nodding while listening) and empathic responses to build rapport with the user, and used continuation prompts (eg, a new question) to keep the conversation going.

Another, technologically less advanced, approach that emphasized longer-term user modeling was taken by the Relational Agents Group of Northeastern University. During the past decade, this research group has developed a framework for so-called relational agents that apply a variety of techniques (eg, daily small-talk, empathic displays, referencing to previous interactions) to develop a long-lasting relationship with the user through menu-based dialogs over multiple interactions [50].

A last set of studies focused on modeling users’ emotional state in real time based on their interaction with the application itself, for example, based on their answers to depression questionnaires [45,49].

The outcome studied the most (n=8) was user satisfaction. Although the studies focusing on more advanced technologies such as natural interaction and empathy modeling were still in the development (n=5) and piloting (n=2) phase, n=3 studies moved beyond that.

The sole study around an implementation question [19] concerned a focus group study (N=16) of the acceptability of SPARX, a gamified CBT intervention developed in New Zealand, in which players can, for example, use a staff to shoot physically manifested negative thoughts. ECAs as such are not a predominant theme in the SPARX game, but regular mention of a guide character is made. Players choose their own avatar that provides instructions throughout the game in dialogs with the user. Australian participants indicated that it was important that the guide’s gender could be customized, did not mind its foreign accent, and liked the idea of being able to socialize with it. Being a focus group study, it remains unclear whether these results would hold in an experimental setting.

The first evaluation study (N=134) used the photograph of a clinician as an embodiment to deliver automated motivational support in a computerized acceptance and commitment therapy [46]. Users could not interact with the ECA directly, and personalized support occurred on a predefined schedule through a user model based on the user’s actions in the intervention. Participants receiving automated ECA feedback were found not to be significantly less involved than those receiving real human support. Although this result was very interesting in the sense that ECA support was compared with real human support, the ECA itself made little use of state-of-the-art ECA technology, and therefore gives us little to go on in terms of ECA design. The second evaluation study concerned another randomized controlled study (N=37) using the job interview training application [52] that was also used for ASDs [23]. Again, the users’ researcher-scored interview skills in a role-played interview, as well as users’ self-confidence, improved significantly compared with a control group that received no intervention.

The development and piloting studies provided us with some initial evidence that practicing health communication with virtual health care providers in a serious game can be efficacious [51], that ECAs in a CBT-based intervention should have a coaching role, be configurable, adaptable, trustworthy, guiding rather than directive, and capable of empathic expressions without reflecting negative ones back to the user [47], that ECAs endowed with empathy are more highly valued than those without it [45,49], that people do not experience less rapport when interacting with an ECA than when interacting with a human [20], that people appreciate the anonymous nature of interacting with an ECA [48], and that people with depression experience a stronger working alliance with a virtual nurse guiding a hospital check-out procedure than do the nondepressed [50].

The counseling studies experimented with various techniques to elicit self-disclosure in counseling sessions with a virtual agent. While we already discussed one study [20] for depression, two other studies focused solely on eliciting personal information from people with anxiety in the context of finding a new roommate [53] and counseling [54]. In the CBT-based applications, virtual animals helped children to conquer performance anxiety in a serious game [55], and a virtual character evoked anxiety in a VR environment [56].

Most innovative here are the counseling studies, again conducted by USC-ICT. Although all three ECAs are based on the same framework, those described in [53] and [54] differ from [20] in that a so-called “Wizard of Oz” paradigm was applied to control the ECAs’ verbal behavior, that is, it was controlled by the researchers. Whereas this violated our definition of agency in terms of verbal behavior, the so-called rapport agents’ nonverbal behavior was completely automated. Interpreting the phonetic aspects of a user’s speech input, as well as video recordings of his or her nonverbal behavior, they were able to display appropriate nonverbal behaviors themselves.

The applications we considered were still in the development (n=2) and piloting (n=3) phase, and there was no predominant outcome measure used. Although some studies worked with large sample sizes (N=351 in [20], and N=108 in [53]), none of the studies experimented with clinical samples.

In these studies, most relevant to our purpose were the findings that people with elevated levels of social anxiety may find it easier to disclose personal information to an ECA than to a human [53], that human backstories may be more effective in this respect than (true) computer backstories [54], and that highly anxious people approached a character in a VR environment more slowly, and kept more distance, than less-anxious ones [56].

In the first study [58], a fantasy character acted as an engaging information repository in a virtual healing environment for returning soldiers that stimulated social comradery, healing activities, and personal exploration. The other study involved a focus group of experts in trauma treatment, in which design requirements were gathered for a virtual agent supporting a Web-based exposure therapy-based application [57].

From a technological perspective, the most interesting developments took place in the two studies we already discussed. The guide in [58] was implemented as a virtual character in a private space built in the Second Life virtual worlds platform, but the details about its design remained unclear.

All studies (n=4) were still in the development phase. Even though [58] had an impressive sample size (N=700), the focus was still on usability of the healing environment itself. Some examples of suggested guidelines for the virtual coach from the focus group study (N=10) [57] were that it should acknowledge patients’ feelings, remind them of their goals when they indicate they wish to quit, be factual in complimenting, and never express negative emotion.

Two studies applied the Relational Agent framework described in the section on depression [50], and used an ECA to host a system that provided general lifestyle support with an emphasis on promoting medication adherence for people suffering from schizophrenia over a 1-month period [18,59]. In the other study, people with schizophrenia could practice conversational skills with virtual characters in a VR social situation [60].

The self-management interventions that were based on the Relational Agent framework used a similar set of techniques as in [50] to develop a long-term relationship between a virtual psychiatric nurse and the user. The conversations in [60] followed a branching tree model approach, which allowed users to communicate through a multiple choice menu. A virtual coach could help users in case they ended up in negative situations. All of the studies allowed users to interact with ECAs through menu-based dialogs.

Although the studies that have not already been discussed under the various disorders were all in the piloting phase (n=3), and sample sizes were fairly small (mean 19.8 [SD 11.9]), they all studied clinical populations. Whereas usability (n=3) and satisfaction (n=4) outcomes were studied most often, with positive results, there is some initial evidence that the Relational Agent applications helped people with schizophrenia to adhere to their medication intake [18,59], and that VR social situations evoke similar negative symptoms in people with schizophrenia to what would be expected from real-world situations [60].

In [61], a “makeover host” was used to deliver and highlight personalized content in an intervention called REALU2, targeting healthy lifestyle behavior with an emphasis on smoking cessation. Smoking cessation was also the topic of [62], which investigated the acceptability of a proposed virtual agent for an intervention based on motivational interviewing. Motivational interviewing was used in a brief intervention targeting problematic drinking behavior in [63,64].

The makeover host in [61] delivered personalized messages based on how users interacted with the intervention, but the details of its design remained unclear. The motivational interview intervention described in [63] allowed users to interact with a virtual counselor through a menu-based system, and their facial expressions were recorded to deduce their emotional state. The combination of user input and emotional state allowed the ECA to guide the conversation and respond empathically. In [64], efforts were made to make interaction with the same system more natural by using speech rather than menu-based input.

This set of studies contained one study in the evaluation phase, which had the highest number of participants (N=1317 adults with a history of smoking) out of all the studies considered. In a randomized controlled trial, the ECA intervention was found to be more effective in reducing self-reported smoking than a control condition in which participants used an intervention based on general lifestyle support. Including peer support further boosted the ECA intervention’s effectiveness. Because the ECA’s design was not described in detail, and because the intervention with the ECA was not compared with one without it, it remains unclear how the ECA itself contributed to the results. Some evidence for the importance of empathic behavior by ECAs was provided in a randomized controlled trial on the brief motivational intervention [63]. The ECA using an empathy module performed significantly better than an ECA without it on various outcome measures, but long-term effects regarding substance use remained unclear.

The balance in terms of the ability to communicate through human communication modalities such as gestures, expressions, and speech highly favors ECAs compared with human users (Figure 3, graph f). Nevertheless, work to shift this balance has been conducted in research on social interaction training and counseling interventions. The latter made a lot of use of technologically more advanced user modeling (eg, real-time emotional states), and innovative ways for humans to communicate with computer systems, for example, by interpreting natural language, phonetic aspects from speech input, and recorded nonverbal behavior. While ASD and counseling interventions have a more short-term focus, there has also been some activity in establishing longer-term relationships by using less technical user models, the input of which came from dialog-tree systems and indirect communication with ECAs through the intervention interface itself. Given the development phases of the studies in which these technologies were applied, the more high-tech solutions seem to be most suited for experimental research into the client-therapist relationship or screening for disorders such as depression or PTSD (eg, [20] or [48]). At present, the more low-tech approaches have been evaluated most thoroughly and, therefore, seem most promising for direct application in routine clinical practice.

From the evaluation phase studies we learned that there is reasonable evidence that (1) ECAs can have a positive effect on user engagement and involvement, (2) they can be effective in this sense as an adjunct to already existing interventions for mental disorders, and (3) it is important for them to convey empathy when interacting with users. An important limitation of the evaluation studies we considered is the nature of the control groups. Besides the study by Kelders [46], none compared an ECA intervention to a conventional treatment used in routine practice, or compared the intervention with the ECA to the intervention without it. Additionally, the more rigorous studies either put little emphasis on the role and design of the ECAs, or targeted disorders indirectly (eg, job interviews skills). Although most of the developmental and piloting studies show promising results with respect to usability and user acceptance, we still have little hard evidence that the proposed applications are reasonable alternatives to established treatments, or that ECAs used as an adjunct to existing interventions, bridging the gap between guided and unguided interventions, make them more clinically effective.

To revert to delivering support in Internet interventions, we know that not all people require the same amount of support, considering, for example, that people with low intrinsic motivation benefit more from human support than those who are already motivated or prefer to work on their own [5]. User models on, for example, patient motivation could contribute to an accurate timing of ECA support, for example, by providing support when motivation is low and not disturbing them when motivation is already high. Considering the working mechanisms through which human support may increase the effectiveness of Internet interventions, little work has been conducted in the area of treatment adherence, making this a pertinent target around which to focus our efforts with respect to increasing motivation.

Another important point related to extending the evidence base is illustrated by the Help4Mood project [45] and the study by Kelders [46]. The Help4Mood project attempts to integrate a Web-based CBT-based treatment with a technologically advanced ECA that communicates through speech, and that is endowed with a dynamic emotion model used to convey empathy in real time. In the study by Kelders [46], automated textual feedback in an already existing intervention was embodied by accompanying it with the photograph of a clinician. Help4Mood is technologically more challenging, but requires a long period of development and piloting. Because the ECA plays such a central role in the intervention, it is far from straightforward to simply “add” the ECA as an adjunct to an already existing and well-evaluated intervention to study its effects. Rather, the intervention would have to be built around the ECA framework, and the resulting new intervention would once again have to proceed through the development phases. This is a general issue whenever we consider adopting one of the other ECA frameworks we came across, such as the Relational Agent Group’s Litebody [65], and USC-ICT Virtual Human Toolkit [66], especially if we consider that the input used in Help4Mood is still relatively straightforward to interpret compared with human speech or nonverbal behavior.

If we want to investigate how to improve interventions that have already been set out in the field, a “low-tech” approach similar to [46] can be advocated because it (1) saves development time that can be used to design and set up larger studies, (2) forces us to think about the core attributes that can make the ECA effective, and (3) makes it easier to judge whether it is safe to use the ECA in a clinical setting with real patients. Given the sparse evidence on the clinical effectiveness of ECAs thus far, this approach and its three advantages may be just what we need to study how we can effectively use ECAs in existing Internet interventions: it will be easier to (1) conduct studies that move beyond the piloting phase, (2) identify the core attributes that make ECAs effective in Internet interventions such that ECA design can be more focused and less time-consuming, and (3) conduct experiments with clinical populations such that we can study ECAs’ effects on clinical outcomes.