The research

The validation table

Full citations are in the bibliography.

What the research supports

That per-user baselining is the right approach. Frontiers in Psychology (January 2026) describes calibration — capturing a neutral expression and adjusting subsequent readings against it — as current best practice, and specifically notes its relevance for passive viewing scenarios. That is exactly the e-learning case, where the learner is watching rather than interacting with a person.

That behavioural signals are strongly predictive. The 2024 ScienceDirect finding that cyberloafing outranks self-regulation and satisfaction as a disengagement predictor is the single most consequential result in this table, for reasons covered below.

That fusing signals beats any one of them. The OUCI/DNTB study's 91.5% → 94.6% improvement establishes that multi-modal approaches outperform single-modal ones. It also, read carefully, establishes how small the facial contribution is.

That disengagement is predictable in advance. Boote, Agarwal and Mostow's 73.3% accuracy with 40% of the activity remaining demonstrates that departure is the end of a process, not a sudden event — and that the process leaves observable traces.

That content-level correlation is a validated approach. The 2021 PMC MOOC framework maps engagement to content positions in real time, which is the same architecture Emotuit was built around seven years earlier.

What it does not

Three honest caveats.

None of this validates our implementation. The papers support the design decisions. They say nothing about whether our code achieves comparable performance. That is an empirical question about our product, and only deployment data answers it.

The basic-emotion model is contested. The Ekman framework underpinning FACS-based emotion classification has been seriously challenged — most prominently by a review led by Lisa Feldman Barrett concluding that the mapping between facial configurations and emotional states is weaker and more context-dependent than the model implies. We use the framework because it remains the field standard, not because the question is settled. This is covered properly in the limits of facial emotion recognition.

Accuracy figures are agreement with a proxy. Every published accuracy number measures agreement with a labelled dataset, where labels come from posed expressions or human annotators. Neither is the person's actual internal state, which is unobservable. This is a structural limitation of the entire field, not a criticism of any particular study.

The finding that changed our product

Read alongside the OUCI/DNTB fusion result, that finding is uncomfortable for a company built on facial analysis. It says the most predictive signal available requires no camera, no biometric processing and about four lines of JavaScript — and that the facial layer contributes roughly three percentage points on top.

When EU AI Act Article 5(1)(f) prohibited biometric emotion inference in workplace and education contexts from February 2025, we had a choice between selling outside our primary market or rebuilding around the behavioural layer. The research made that an easier decision than it might have been: we were being pushed toward something the evidence already favoured.

That is the origin of Signals, the behavioural-only configuration.

How to read vendor research claims

Including ours. Four questions worth asking of any citation in this category:

Was accuracy measured on posed or spontaneous expressions? Posed datasets — actors instructed to display anger — produce much higher figures than field data. This distinction is rarely volunteered and it is usually the largest single factor in a headline number.

What is the comparison condition? "94.6% accurate" means little without knowing what the alternative achieves. In the fusion study the alternative is 91.5%, which reframes the claim substantially.

Is the population representative of yours? Many studies use university students in controlled conditions. A distributed corporate workforce on varied hardware in varied lighting is a different setting.

Does the paper support the product, or just the category? A study showing that facial engagement detection is possible does not show that a specific vendor achieves it. This is the gap most research-led marketing quietly relies on, and it applies to this page too.

Every source cited anywhere on this site is listed in the bibliography, with links, so you can check any of it directly rather than taking our summary on trust.

Frequently asked questions