AI takes on the practice and measurement that scale; people keep the encounters that matter. MyNoema builds the measurement layer for counselor training — wherever a claim of "this works" has to hold up, it should pass through this layer.
A real session with a simulated client — and the utterance-level behavioral coding and four layers of structured feedback the system returns.
The full 90-second demo, hosted on YouTube. Open in a new tab ↗
The demand needs no argument — the breakdown does. Three structural failures in the pipeline, one of them independently measured.
Louie et al., CHI 2026 — a randomized controlled trial (independent third-party study). The practice-hour gap and supervision shortage are structural constraints documented across counselor-training systems.
Whether the dialogue feels human is available to everyone.
Whether the practice works, and who proves it — that is our entire engineering.
Pilots don’t learn takeoffs and landings on a real plane, and counselors shouldn’t learn the basics on real clients. A simulator is worth only as much as two things: a controllable, reproducible environment and accurate instruments — and that’s where all of our engineering goes.
AI simulated-client training system
Trainees run sessions with behaviorally consistent, reproducible simulated clients; the system scores every utterance against an internationally established behavioral coding scheme (MITI) and generates four layers of structured feedback — engineering the ingredients RCTs have shown to work (practice + structured feedback).
CBT-based self-help emotional-support dialogue · for research
A self-help emotional-support dialogue system driven by structured CBT sessions, used to study whether structured, safety-first dialogue can be more helpful than a general-purpose chatbot. Not therapy, not a public product — consent-gated and limited to invited participants.
Our boundaries · what we deliberately don’t do
No substitute for real client hours
Simulated practice is a training bench before real hours, and a multiplier on them.
No substitute for supervision
Low-level correction goes to the coding engine; scarce supervision is reserved for conceptualization and the therapeutic relationship.
No public-facing commercial companionship product
The CBT study above is consent-gated, non-public, and for research only.
Scores are not used for high-stakes decisions
Until transfer-validity evidence is established, system scores serve as formative feedback only.
Whether the dialogue feels human comes from the underlying model and is available to everyone; whether the practice works, and who proves it — that’s where our engineering lands. Every claim can be checked by hands-on trial or peer review.
Behavioral-code annotation and state-transition analysis across 5 public datasets, yielding 4,785 state-transition pairs to calibrate the simulated-client state machine — clinical pacing derived from data, not prompt guesswork.
A double-human-coded gold-standard set, inter-rater reliability (IRR) management, and a 40-case regression suite — scoring behavior is not allowed to regress across releases. Only when measurement is trustworthy does feedback earn the name.
Our interpretability research found empathy representations inside LLMs are decodable but not reliably steerable. So every efficacy claim lands on trainees’ measured outcomes, not the model’s generation quality — a conclusion from our own research.
A behavioral-coding distribution comparison between AI simulated clients and real counseling corpora (a study type still absent from the literature); plus MITI gold-standard reliability and trainee pre/post studies, on a 12-month roadmap.
Related paper
Detection ≠ Reliable Control: Empathy Directions in LLMs Are Decodable but Not a Reliable Steering Lever
BlackboxNLP @ EMNLP 2026 (submitted)
MyNoema was founded by a team with backgrounds in psychology and computer science at UC Berkeley and human-computer interaction at USC. What this field lacks isn’t a bigger model, but infrastructure — controllable training, trustworthy measurement, and verifiable claims.
The system is live; institutions can trial it hands-on, not just watch a demo.
simpatient.mynoema.org →The scoring process is transparent end to end: utterance-level records, expert-annotated reference answers, and 40 safety tests.
The research is submitted to peer review, and the architectural choices are documented.
We’re looking for: university programs with a clinical-counseling track (co-developed courses · joint research · pilot deployments), counselor-training institutions, and academic partners.