Built on evidence, not vibes

Reeves, Curran, Gleeson and Hanna (2022) pooled 11 controlled studies with 508 participants that treated public-speaking anxiety either with virtual-reality exposure or with in-vivo exposure. Both worked, and they worked to the same degree: the standardized effect versus control was d = −1.39 for simulated exposure and d = −1.41 for real exposure, a difference with no statistical meaning. The finding matters for Kari because it settles the obvious objection to practicing with an AI, that a simulation cannot substitute for the real thing. For the fear of speaking, it can. Kari's roleplays are spoken exposure with a character who reacts in real time, so every session counts as a legitimate exposure trial, not a rehearsal for one. Kari does not claim clinical treatment; it uses the mechanism the study validated, and it measures the behaviors that change because of it.

Beidel and colleagues (2014) randomized 106 adults with social anxiety disorder to exposure therapy alone, to exposure combined with social skills training (Social Effectiveness Therapy), or to a wait list. Exposure alone reduced anxiety. Only the combination improved what independent judges could actually see: rated social skill improved with d = 0.87 over exposure alone, avoidance with d = 0.61 and depression with d = 0.79. In plain terms, facing the situation makes you less afraid; it does not by itself make you better at the conversation. This is the reason Kari never ships roleplays without lessons. A lesson teaches one concrete technique and makes you say it out loud; the roleplay then puts that technique under pressure with a character who reacts. Drop either leg and the trial says you lose the skill gain.

Niebuhr, Tegtmeier and Schweisfurth (2019) put 72 participants through the PASCAL training system, where they practiced speaking with automated feedback on prosody: pitch variation, pace, emphasis and pauses. After four hours of feedback-driven practice their prosodic charisma score rose from 40 to 71 on the system's scale (p < 0.001), a large effect for a very short intervention. The same line of research shows that acoustic features alone predict 66–75% of how charismatic a speaker is judged to be, before a listener processes a single word. Kari is built on this. It listens through a live voice API rather than reading a transcript, measures energy variation, sentence endings and pauses, and reports them as the Tone trait. The practical claim is narrow and well supported: the vocal part of charisma responds to feedback within hours, and it is the part most apps never hear.

Social skills training has a simple core loop: watch the behavior modeled, try it yourself, get corrected, try again. Turner and colleagues (2018) pooled 27 randomized controlled trials (N = 1,437) and found that this loop produces reliable gains (Hedges' g ≈ 0.3) even in psychosis, the clinical population where change is hardest, and that the gains were still there at follow-up. The effect is moderate, which is what you would expect from an intervention that trains behavior rather than mood. Kari's lesson engine is that loop compressed into a few minutes: Kari models the technique, you attempt it out loud, she corrects one thing while quoting your literal words, and you attempt it again with the difficulty raised. The design choice the evidence forces is one correction per attempt. Feedback that lists everything you did wrong is not what the trials tested, so Kari does not do it.

Antonakis, Fenley and Liechti (2011) ran two experiments to test whether charisma can be taught. Managers were trained in a fixed set of verbal and nonverbal tactics: stories and anecdotes, metaphors, contrasts (“it is not X, it is Y”), three-part lists, rhetorical questions, expressions of moral conviction, and animated voice and gesture. Afterwards their own teams, who did not know about the training, rated them as significantly more charismatic and more influential than before and than untrained controls. These “charismatic leadership tactics” are the only charisma techniques with a demonstrated causal effect, and every one of them can be counted in a transcript. That is why Kari's lessons drill exactly these moves, and why the report can tell you whether you used a contrast or a three-part list rather than whether you seemed confident.

Huang, Yeomans, Brooks, Minson and Gino (2017) studied live conversations, including speed dates, and found that people who ask more questions are liked more, and that follow-up questions, the ones that dig into what the other person just said, carry most of the effect. Speed daters who asked more follow-up questions were more likely to get a second date. Questions that switch topic do not produce the same result; they read as interrogation rather than listening. Kari treats this as a hard metric. Every session report counts your follow-up questions and callbacks, and the Listening trait is built largely from them. It also counts long silences and energy variation, because the same body of research shows these are what listeners actually use to judge you. The number is not a vibe. It is a count of behaviors you can change next time.