Compliance note
Deepform and the EU AI Act
Last updated: 27 September 2026
The EU AI Act bans AI that infers emotions from people at work. It does not ban AI that measures physical states such as fatigue. Deepform only ever does the second thing — and that is a product decision we made before it was a legal one.
The short version. Article 5(1)(f) prohibits placing on the market or using AI systems that infer emotions of a person in the workplace. It sits in the Act’s highest penalty tier: up to €35 million or 7% of worldwide annual turnover, whichever is higher. Recital 18 states that physical states such as pain or fatigue fall outside the definition of emotion recognition, giving the example of systems detecting fatigue in pilots or drivers to prevent accidents.
What the prohibition actually covers
The Act defines an emotion recognition system by its purpose: identifying or inferring emotions or intentions of natural persons on the basis of their biometric data. Recital 18 gives a non-exhaustive list of what counts as an emotion — happiness, sadness, anger, surprise, disgust, embarrassment, excitement, shame, contempt, satisfaction, amusement.
The prohibition applies in workplaces and educational institutions, with a narrow exception for medical or safety reasons. It has applied since 2 February 2025. The 2026 Digital Omnibus on AI (Regulation (EU) 2026/1744) deferred the high-risk regime, but left the Article 5 prohibitions on their original schedule — so this is in force now, not a future deadline.
What Deepform measures, in full
Every signal below is a physical or behavioural measurement with a physiological definition. None of them is an emotion, a mood, an attitude, or a judgement about how someone feels about their work.
- Blink rate
- Blinks per minute. Drops sharply during screen work; a documented driver of dry-eye symptoms.
- Blink completeness
- Whether the lids fully met. Incomplete blinks leave the ocular surface exposed.
- PERCLOS
- Proportion of time the eyes were ≥80% closed. The standard drowsiness measure used in driver-monitoring research.
- Posture
- Head and trunk angle from pose landmarks. A load on the cervical spine, measured in degrees.
- Screen distance
- Estimated eye-to-screen distance in centimetres.
- Break behaviour
- Whether a break was taken, and whether the measurements recovered afterwards.
What Deepform will never measure
This is a written product constraint, not a roadmap gap. It is recorded in our engineering non-negotiables and enforced in the feature schema — there is no field for any of it:
- Emotions or mood of any kind — including any of the states listed in Recital 18.
- Engagement, enthusiasm, motivation, attitude, or sentiment toward work.
- Productivity, output, keystrokes, application use, or time on task.
- Any inference about intent.
Why the architecture matters as much as the metric list
- Processing is on-device. Camera frames are analysed in the browser and discarded. No image or video frame is transmitted or stored, so there is no repository of workplace imagery to be requested, breached, or repurposed.
- Only numbers leave the device, and only if the person opts in.
- Employers never see an individual's scores. Team views show aggregates only, and suppress any group below our k-anonymity floor of five people.
- The individual sees their own numbers first. The product is built for the person being measured, not for someone watching them.
What this page is, and is not
It is a plain statement of what our system does, set against the text of the provisions, so that a buyer’s legal team, DPO, or works council can assess it quickly. It is not legal advice, and not a certification. We are not your lawyers, and no vendor can tell you that your particular deployment is compliant — that depends on how you deploy it and on your own obligations under the Act, the GDPR, and national employment law. What we can do is be precise about our side of it, and we would rather you verify this than take our word for it.
If something here is wrong or out of date, tell us and we will correct it. That offer is part of the point.
Sources
- Regulation (EU) 2024/1689 (AI Act), Article 5(1)(f) — prohibition on emotion recognition in the workplace and educational institutions.
- Regulation (EU) 2024/1689, Recital 18 — definition of emotion recognition; physical states such as pain and fatigue excluded.
- Regulation (EU) 2024/1689, Article 99 — penalties for infringement of Article 5: up to €35,000,000 or 7% of total worldwide annual turnover.
- Regulation (EU) 2026/1744 (Digital Omnibus on AI) — defers the high-risk regime; Article 5 prohibitions unchanged and still in force from 2 February 2025.
Deepform is a wellness and ergonomics product. It is not a medical device and does not diagnose, treat, or prevent any condition.