Virtual humans can appear realistic without always being recognized as artificial. Across studies from 2007 to 2026, measured outcomes depend on more than facial appearance: voice, breathing, motion synchrony, personalization, agency, environment, and task context all matter. The figures below are study findings, not a global market estimate. Scores, percentages, and effects retain each source’s population, scale, geography where reported, and experimental setting.
Table of contents
- Core measurements of virtual human performance
- Photo-realism and detection
- Personalization and motion synchrony
- Social presence, agency, and interaction
- Behavioral realism and engagement design
- Context, trade-offs, and evidence boundaries
Core measurements of virtual human performance
A 2025 study of 20 healthy adults evaluated a virtual human breathing coach in non-immersive, browser-based VR. The coach received a likeability score of 72.75 and a trust score of 73.85 on 0–100 scales. Likeability and trust were positively correlated at R = 0.68 (p = 0.001). These results describe participants’ responses to that breathing-exercise system, rather than a universal trust level for virtual humans.
Overall realism was lower, at 53.8 on the same 0–100 scale, even though voice realism was rated 77.25 out of 100. A regression model explained about 72% of the variance in overall realism (adjusted R² = 0.72). Within that model, breathing realism was a positive predictor of overall realism (β = 0.74, p = 0.003). The result suggests that dynamic behavior can influence perceived realism alongside visual design.
The same study connected perceived realism with practical interaction outcomes. Realism correlated with felt influence at R = 0.54 (p = 0.014) and with ease of following at R = 0.46 (p = 0.043). Participants rated ease of following at 81.25 and relaxation at 66.25 on 0–100 scales. Usability averaged 76.15 (SD 18.45), while interest in continuing to use the coach averaged 60.4 (SD 22.65).
| Measure in the 2025 breathing-coach study | Result | Scale or statistic |
|---|---|---|
| Likeability | 72.75 | 0–100 |
| Trust | 73.85 | 0–100 |
| Overall realism | 53.8 | 0–100 |
| Voice realism | 77.25 | 0–100 |
| Ease of following | 81.25 | 0–100 |
| Usability | 76.15 | 0–100; SD 18.45 |
| Continued-use interest | 60.4 | 0–100; SD 22.65 |
Source: Usability, acceptance, and the role of realism in virtual humans for breathing exercise training.
Photo-realism and detection
A four-experiment avatar-face study conducted in 2025 and published in 2026 found a substantial realism gap between high-resolution face photographs and avatars: t(49) = 18.82, p < 0.001, d = 2.66. High-resolution avatars also appeared more realistic than low-resolution avatars, t(49) = 14.38, p < 0.001, d = 2.03. The interaction between face type and resolution was F(1,49) = 146.34 (p < 0.001).
Recognition was not a simple mirror image of realism. Only 28.0% of participants spontaneously reported seeing avatar faces, but 84.0% noticed them after a direct prompt. In the initial detection task, participants estimated that 53.6% of displayed faces were avatars (SD 19.1%). Classification accuracy was 94.1% when an avatar and a photograph were shown side by side, compared with 87.4% when images were presented individually.
Performance varied substantially by identity. High-resolution avatar classification accuracy ranged from 32.0% to 98.0% across face identities. Among high-resolution avatar images, 66.0% received the maximum realism rating of 7 from at least one participant. Avatar realism correlated negatively with forced-choice identification accuracy at r(38) = −0.91 (p < 0.001): more realistic avatars were harder to identify as avatars in that task.
Trait impressions tracked the corresponding photographs. Avatar and photo attractiveness ratings were 3.76 and 3.71 on a 7-point scale, with r(38) = 0.82 across identities. Trustworthiness ratings were 4.04 for avatars and 4.36 for photographs, with r(38) = 0.63. Eye-region fixations were about 5.0% higher for avatars than for photographs.
Source: Avatars versus the people: Photo-realism, spontaneous detection and trait inferences of digitised faces. The experiments were conducted in 2025 and published in 2026, so these findings should not be presented as a timeless benchmark.
Personalization and motion synchrony
In a 2022 embodiment study, personalized avatars scored 0.92 on appearance similarity, compared with −0.79 for generic avatars. The appearance difference had η² = 0.589 (p < 0.001). Personalization therefore changed how closely the avatar matched participants’ appearance judgments in that experiment.
Motion synchrony had an even stronger relationship with agency. Synchronized personalized avatars averaged 2.38 for agency, compared with −0.79 for asynchronous personalized avatars. Synchronized generic avatars averaged 2.25, compared with −1.75 for asynchronous generic avatars. In personalized-avatar testing, motion synchrony had a large agency effect of η² = 0.799 (p < 0.001).
Synchrony also affected ownership and presence. Its body-ownership effect was η² = 0.499 (p < 0.001), while avatar appearance had η² = 0.312 (p = 0.004). Synchronized personalized avatars produced a presence mean of 74.21, compared with 55.71 asynchronously. Synchronized generic avatars produced a presence mean of 74.71, compared with 54.92 asynchronously. Motion synchrony explained η² = 0.636 of the variance in presence scores, while avatar appearance had no significant main effect (p = 0.93).
Emotion ratings showed a smaller but measurable difference: synchronized personalized avatars had valence of 6.38 versus 5.63 asynchronously on the study’s 9-point emotion scale (p = 0.004). These results separate two design choices that are often treated as one: making an avatar look personal and making its movement track the user.
Source: Impact of Personalized Avatars and Motion Synchrony on Embodiment and Users’ Subjective Experience, 2022.
Social presence, agency, and interaction
A three-condition avatar study from 2020 found significant differences in human-avatar ownership, F(2,52) = 6.239, p = 0.004, η² = 0.194. Human-avatar self-location also differed across conditions, F(2,52) = 16.253, p < 0.001, η² = 0.385. In a second study, motion synchrony produced an agency effect of η² = 0.885, an ownership effect of η² = 0.792, and a presence effect of η² = 0.760; all were reported at p < 0.001.
Older synthesis research provides a useful scale boundary. A 2007 meta-analysis descriptively synthesized 46 empirical studies and statistically analyzed 25 formal studies. Across those 25 studies, adding a human-like face to an interface explained about 2.5% of outcome variance. A face can contribute to user experience, but the aggregate effect was modest in that evidence base.
A 2015 meta-analysis included 32 studies comparing avatars and agents. It found stronger social responses for perceived human-controlled avatars than for computer-controlled agents. Agency effects were greater in competitive and cooperative tasks than in neutral tasks. Thus, whether a virtual human is perceived as human-controlled, and what users are trying to do, can alter the measured interaction effect.
Sources: Controlling the Sense of Embodiment for Virtual Avatar Applications; A meta-analysis of the impact of the inclusion and realism of human-like faces on user experiences in interfaces; and Avatars versus agents: A meta-analysis quantifying the effect of agency.
Behavioral realism and engagement design
An 8-week longitudinal VR study followed a group of 81 participants. Entitativity, presence, enjoyment, and realism all increased over the eight weeks. Self-avatars increased self-presence and realism but decreased enjoyment compared with uniform avatars. The trade-off matters for avatar design: a representation can feel more self-related while producing a less enjoyable experience in the same study.
A second VR study included 137 participants who cycled through 192 unique virtual environments. Larger visible virtual spaces increased nonverbal synchrony, restorativeness, entitativity, pleasure, arousal, presence, enjoyment, and realism. Across the 192 environments, outdoor settings produced greater restorativeness and enjoyment than indoor settings. These findings concern transformed avatars and environmental context in group interaction, not all virtual-human applications.
Behavioral expectations can also outperform visual detail as a realism cue. In 2021 resized-user VR experiments, 73% of study-1 participants and 70% of study-2 participants selected movie physics as more realistic. Movie physics matched expectations for 91% of study-1 participants but only 63% of study-2 participants. Positive presence was reported by 82% and 75% of participants, respectively. In study 2, a logistic model explained 38% of response variance and achieved 79.5% overall accuracy when predicting physics preference.
Source: The Plausibility Paradox for Resized Users in Virtual Environments.
Context, trade-offs, and evidence boundaries
A 2017 public-space surrogate study observed 7,685 people during human-surrogate interactions. Overall engagement was higher for females than males, while reactivity was higher among younger than older people. Physical presence increased the likelihood of engagement, but body-part movements and gestures did not show the expected engagement benefit. Those observations come from a public-space surrogate setting and should not be treated as direct evidence for browser avatars, VR coaches, or social platforms.
The breathing-coach results also show why individual metrics should not be collapsed into one “realism” score. Realism and influence correlated at R = 0.54, but likeability and trust did not individually correlate with relaxation or ease of following. The coach’s trust score was 73.85, yet continued-use interest was 60.4, indicating that positive judgments on one scale do not automatically translate into future-use intention.
Across the evidence, the strongest measured effects often concern synchrony, embodiment, or task-specific expectations rather than appearance alone. The studies include primary experiments, university research publications, and meta-analyses using different samples, interfaces, scales, and outcomes. Statistical significance and effect sizes are reported as published; they should not be generalized beyond the populations and tasks studied. Percentages and scores are therefore best used as design evidence and comparison points, not as a global prevalence, adoption, or market forecast.