Emotion and Intent Inference

This is the one category in this collection where the recommendation is not "proceed carefully". It is: do not build or buy this.

What is claimed

Emotion from facial expression: happy, angry, sad, surprised, and derived scores like satisfaction or frustration.

Attention or engagement from head pose and gaze.

Stress or fatigue from facial and postural cues.

Intent, usually marketed as detecting suspicious behaviour before an act.

Deception, which appears periodically and has no scientific standing at all.

Why the science does not support it

The premise is that discrete emotions map reliably to facial configurations. Substantial reviews of the psychological literature have found that this mapping is far weaker than the products assume.

Expression varies by culture, context and individual, and the same configuration means different things in different situations.

People produce expressions strategically, which is a normal social behaviour and not a malfunction.

Context dominates. The same face reads differently depending on what is happening around it, and a camera analysing a face has no access to that.

Which means an emotion score is not a measurement of an internal state. It is a classification of an appearance, and the two are being conflated.

Why intent detection is worse

There is no ground truth. You cannot label a dataset with what someone was about to do.

So the label becomes a proxy — usually who was subsequently stopped, or who someone judged suspicious, which encodes whatever biases produced those judgements.

The system then reproduces those judgements at scale, with the appearance of objectivity.

"Suspicious behaviour" is not a definable visual class, which is the test applied elsewhere in these notes and which this fails outright.

The legal position

General orientation; specifics require advice for your jurisdiction.

Under the EU AI Act, inferring emotions of a natural person in the workplace or in education based on biometric data is a prohibited practice under Article 5, except for medical or safety purposes, and has been enforceable since February 2025. Penalties for prohibited practices reach the highest tier in the Regulation.

Biometric categorisation to infer sensitive characteristics is separately prohibited.

Emotion recognition outside those contexts is classified high-risk rather than banned, with obligations including a transparency duty toward the people subject to it.

Under data protection law generally, inferring an emotional state from biometric data is processing of special-category data requiring a lawful basis that is difficult to establish in an employment context.

What to say when asked for it

Say the science is contested, and that the products present classifications of appearance as measurements of internal states.

Say the error distribution is uneven, as it is for every appearance-based system, so the misclassification falls on the same groups.

Say it is prohibited in the workplace and educational contexts in the EU, and high-risk elsewhere.

Then answer the underlying question, which usually exists: a manager asking for engagement detection generally wants to know whether people are struggling, and that is answered by asking them.

The adjacent cases that are legitimate

Worth naming, because a blanket refusal is imprecise.

Fatigue detection in a driving or machinery context, for safety, falls within the medical-or-safety carve-out and has a better evidence base than emotion inference, and it still needs the full assessment.

Detecting a person on the ground, which is a posture and a condition rather than an inferred mental state.

Aggregate crowd density and flow, which infers nothing about individuals.

The distinguishing question: is the system detecting an observable condition, or asserting something about what is going on inside someone's head?

The condition-versus-state test

One question that separates the legitimate adjacent cases from the prohibited ones.

Is the system detecting an observable condition, or asserting something about what is going on inside someone's head?

A person on the ground: condition.

A person appearing frustrated: state.

Eyes closed for a sustained period in a driving context: condition, with a safety purpose.

Engagement, satisfaction, suspicion, intent: states, every one.

Apply it to any proposed detection class, and refuse the second category rather than negotiating its accuracy.

Answering the underlying request

Requests for emotion detection usually conceal a real and answerable concern.

"Are customers frustrated?" — measure waiting time, which is objective and actionable.

"Are staff disengaged?" — ask them, in a survey that is aggregate and honest.

"Is someone about to become violent?" — a security and de-escalation question with established practice, none of it visual inference.

"Is the driver tired?" — a safety question with a narrow, evidence-based technical answer under the medical-and-safety carve-out.

Offering the real answer is what makes the refusal a consultation rather than an obstruction.