Built on evidence, not vibes
Every rule in Kari — how characters respond, what the report measures, why lessons pair with roleplays — traces back to a specific finding. These are the load-bearing ones.
Talking to an AI by voice matches real-life exposure.
A meta-analysis of 11 controlled studies (N=508) found simulated exposure for public-speaking fear practically identical to in-vivo exposure (d=−1.39 vs d=−1.41). Practicing out loud with a voice AI is a legitimate form of exposure.
Meta-analysis · 11 RCTs · N=508Technique plus practice works — exposure alone doesn't build skill.
In a randomized trial (N=106, social anxiety), exposure plus skills training beat exposure alone on judge-rated skill (d=0.87), avoidance (d=0.61) and depression (d=0.79). That's why Kari pairs lessons with roleplays: neither leg is optional.
RCT · social anxiety · N=106Vocal charisma is trainable in hours, not years.
The “Pascal” system (N=72) produced a large improvement in prosodic charisma after 4 hours of feedback-driven practice (score 40 → 71, p<0.001). Acoustic signal alone predicts 66–75% of perceived charisma.
Pascal system · N=72 · p<0.001The model → rehearse → feedback loop is the proven core.
A meta-analysis of 27 randomized trials of social skills training found reliable effects (g=0.33) even in hard clinical populations, maintained at follow-up. That loop — see it, try it, get corrected — is exactly Kari's lesson engine.
Meta-analysis · 27 RCTsCharisma can be taught — and trained people are seen as more influential.
Managers trained in concrete verbal tactics (stories, metaphors, contrasts, three-part lists) were rated significantly more influential by their own teams. These are the only charisma tactics with demonstrated causal effect — and the ones Kari drills.
Antonakis et al. · Univ. of Lausanne · 2011Follow-up questions predict being liked.
Asking questions that dig into what the other person just said predicts likability and even second dates (Huang et al., 2017). Kari counts your follow-up questions, long silences and energy variation as hard evidence in the report.
Huang et al. · JPSP · 2017Four principles the app refuses to break
Behavior, never “calmness”
End-of-session calm doesn't predict long-term results. Kari's report measures what you did — techniques used, questions asked, filler words — never whether you seemed relaxed.
Expectation violation over repetition
Exposure teaches when reality disconfirms your catastrophic prediction. Kari captures what you fear will happen and contrasts it with what actually happened.
Drop the safety crutches
Minimal answers, constant apologies, deflecting humor — safety behaviors keep fear alive. Kari's characters make crutches visibly costly, and lessons assign “this time without X”.
Variety beats grinding one scene
Varied stimuli and contexts retain more than repeating the same scene until comfortable. Kari always nudges you to switch scenarios and characters.
What we don't claim
Direct studies of AI voice roleplay are early and promising — medical students practicing with voice AI improved confidence across three domains (p<0.001), and 82% of MIT negotiation-course students felt better prepared after bot practice — but most of that is self-report without control groups. So Kari doesn't promise clinical outcomes, and it is not therapy.
What the strong evidence does support: simulated exposure works like real exposure, skills training plus exposure beats either alone, prosodic feedback improves perceived charisma fast, and specific behaviors (follow-up questions, hedging, silences over four seconds, concrete language) reliably shape how people perceive you. Kari measures exactly those behaviors — nothing mystical, nothing it can't count.
See how this becomes the methodRead enough. Say something.
The research says hours of spoken practice with feedback move the needle. Your first session takes five minutes.
Download Kari on the App Store