## The original question
In 2014 Ross Jones was taking an online course and reading an article about facial recognition, and put the two together.
In a classroom, a teacher can read the room. They can see who is lost, who has checked out, and which explanation landed. The moment that course moves online, that entire channel disappears. What replaces it is a completion percentage — a number that tells you someone reached the end, and nothing whatsoever about whether they were present on the way.
Emotuit was founded to close that gap. The original system used facial expression recognition to score engagement per student, with a per-person baseline so that "engaged" meant something individual rather than a population average. It shipped as an engagement analytics app for the Canvas LMS, and it was covered by KCPW radio, BeehiveStartups and the Reimagine Education awards, and profiled on MIT Solve.
## What changed, and what didn't
Two things changed.
**The technology got better.** What needed a server in 2014 now runs client-side in the browser through TensorFlow.js and ONNX Runtime Web. Face detection and classification happen on the learner's own device, and no images need to leave it.
**The law arrived.** In February 2025, Article 5(1)(f) of the EU AI Act prohibited AI systems that infer emotions from biometric data in workplaces and education institutions — which is to say, in exactly the two contexts this company was built to serve.
We could have ignored that and sold outside the EU. Instead we went back to the research and found something we should probably have taken more seriously a decade ago: **the behavioural signals were doing most of the work all along.** A 2024 study ranked tab-switching as the single strongest predictor of disengagement in online courses, ahead of self-regulation and satisfaction measures. Adding facial expression to behavioural data improves classification accuracy from about 91.5% to 94.6% — real, but marginal, and it is the increment carrying all of the regulatory weight.
So we built [Signals](/signals): the same content-level engagement measurement, derived entirely from behavioural telemetry, with no camera and no emotion inference. It is lawful where facial analysis is not, and it is a better fit for how most organisations actually want to treat their people.
**What didn't change is the question.** Which part of this content loses people, and why. That has been the same problem for twelve years.
## What we believe
**Content-level, not person-level.** The useful output is "slide 14 loses 60% of learners", not a dossier on an individual. Aggregate reporting is better product design and it happens to resolve most of the privacy question at the same time.
**Completion rates are worse than useless.** They are actively misleading. Compliance training routinely records 90%+ completion while producing little measurable behaviour change. A metric that reads perfectly while the underlying thing fails is not a weak metric — it is a broken one.
**Show the research.** Every architectural decision in this product is backed by published work, and we publish the limitations too — accuracy ceilings, cultural variation in expression, the gap between posed and spontaneous expression. Vendors who only publish their strengths should be read accordingly.
**Measurement should not require surveillance.** If you can get the answer without a camera, you should.
## In the press
- **KCPW** — *The Bottom Line: Ed-Tech Series* — radio interview with Ross Jones
- **BeehiveStartups** — *How Emotuit Uses Facial Recognition To Enhance Online Learning*
- **Reimagine Education Awards** — *The Next Generation of Educational Analytics*
- **MIT Solve** — solution profile
- **Instructure Edu App Center** — Emotuit Engagement Analytics for Canvas
## Who you deal with
Emotuit is founder-led. If you get in touch, you are talking to Ross — the person who built it — rather than a sales development rep working from a script. That means shorter conversations, straighter answers about what the product does and does not do, and no pretending a limitation is a feature.