Guide · Published November 14, 2026 · Reviewed November 14, 2026 · Ingenious AI team · Not legal advice
Emotion Recognition and Biometric Categorisation: The Article 50(3) Disclosure Duty
The short answer: Article 50(3) of the EU AI Act places a disclosure duty on deployers: businesses running emotion-recognition or biometric-categorisation systems must inform the people exposed to them. The duty has applied since August 2, 2026. Its practical difficulty is discovery, not drafting: emotion inference hides inside interview platforms, sentiment tooling, and engagement analytics that nobody purchased under that name. A deployer who has never audited for these capabilities does not know whether it owes the disclosure — which makes the audit, including its negative findings, the first compliance artefact.
Check your duties and dates — free Build your first notice — free
What are emotion-recognition and biometric-categorisation systems? In the terms of the kit's checker data, they are systems that recognise emotions or categorise people biometrically — inferring states or categories from faces, voices, or other biometric signals. The checker lists the informing duty among Article 50's four, alongside interaction disclosure, content marking, and deepfake disclosure, in EU AI Act Article 50 Transparency Duties.
Discovery: where emotion inference hides
Start the audit where the capability actually lives: video-interview platforms (engagement or sentiment scoring), voice analytics in call tooling, proctoring and assessment software (attention or affect flags), and customer-experience analytics that categorise callers or visitors. Ask vendors the direct question — does the system infer emotion, or categorise people by biometric characteristics, in any mode, including modes we have not enabled? — and record the answer per system.
Modes matter. A capability present but disabled is a different position from one that is active, and both differ from a capability the vendor added in an update. The audit should record, per system: capability present or not, enabled or not, populations exposed, and the date and source of the finding (vendor documentation, contract, or written answer). That record is what converts 'we don't think we use that' into an evidenced position.
Who must be informed: the people exposed
The duty runs to the people exposed to the system — the interviewee whose affect is scored, the caller whose voice is analysed, the candidate whose assessment includes biometric categorisation. Informing is owed by the deployer running the system on them, in the same deployer pattern as every other duty in this series: the business whose process the person is inside of cannot point at the vendor.
Timing follows the Article 50 logic: the person should be informed in connection with the exposure, in a form they actually encounter — at the interaction point, as with the interaction disclosure itself. A line in a supplier-facing policy document the person never sees does not inform the person exposed. Coordinate the 50(3) information with the 50(1) interaction disclosure where both apply: one clear notice at the point of contact can carry both, provided each duty's content is actually present.
Worked example: Northwind Talent Partners
Example — fictional business, real duty structure. Northwind is a fictional staffing firm with EU-touching hiring operations that ran the Article 50(3) audit across its stack:
- Systems audited: 7 systems reviewed with the direct vendor question — interview platform, assessment suite, ATS, voice analytics, proctoring tool, CRM sentiment add-on, and scheduling suite.
- Capability found: the interview platform includes an engagement-inference feature — present, and enabled by default on 2 of the firm's 5 interview workflows, exposing an estimated 310 candidates in the audited quarter.
- Decision taken: the firm disables the feature (its hiring process has no use for affect data), records the configuration change with its date, and adds a contractual update-notice requirement so re-enablement or capability changes surface.
- Negative findings filed: the other 6 systems are recorded as no-capability findings, each with its evidence source (vendor documentation or written answer) and date.
- Register updated: an emotion/biometric capability field is added per tool; the audit joins the Article 50 file and the register's quarterly review.
Negative findings are findings: record the audit either way
Many businesses will audit and find nothing — no emotion recognition, no biometric categorisation in the stack. That result is not the absence of a compliance artefact; it is the artefact. File the audit with its scope, method, per-system findings, and date. If a vendor update later adds the capability, the audit's date is what makes the change visible as a change, and the material-update conversation with the vendor is where it should surface.
Re-run the audit on the same triggers as the tool inventory: new systems, material updates, new use cases (an assessment platform extended from skills testing into interview analysis is a new exposure). The inventory register is the natural home — add a field for emotion/biometric capability per tool, so the question is asked structurally at intake rather than remembered annually.
Hiring stacks: the highest-probability exposure
Employment deserves its own paragraph because it concentrates the risk: interview intelligence, video assessment, and screening add-ons are the consumer market for affect and engagement inference, and job applicants are the population least positioned to question what is run on them. Connecticut's employment track and the US notices do not contain an equivalent of 50(3) in the kit's data — this is an EU duty, owed where the EU frame applies, and it should not be silently generalised to or from the US duties.
An employer with EU-touching hiring who finds an active emotion-recognition feature has three design options: inform the exposed candidates properly, disable the capability and record the configuration, or replace the tool. What the duty excludes is the fourth option — running it uninformed. Record which option was chosen per system, with dates, in the same register.
The record, again: one audit trail for Article 50
Article 50's four duties generate four small records — interaction disclosures shown, marks applied and tested, deepfake and public-text disclosures logged, and the emotion/biometric audit with its per-system findings. Keep them together as the Article 50 file, reviewed on the inventory cadence. The file's reader, a year from now, should be able to answer for each duty: what we run, what we tell people, and what we checked — with dates.
Confirm your EU position through the free checker: its EU result is the same for every decision role precisely because Article 50 attaches to interaction, content, and exposure rather than to decision weight — the structural fact that makes this duty easy to overlook from a decision-centric programme, and worth its own guide.
- Ask every vendor the direct question — emotion inference or biometric categorisation, in any mode — and record the answer per system
- Distinguish capability present, enabled, and actively exposing people; each state has a different action
- If people are exposed, inform them at the point of exposure, as the deployer — coordinated with the interaction disclosure
- If you disable a capability, record the configuration and its date, and contract for update notice so it cannot return silently
- File negative findings as findings, and add a capability field to the tool register so intake asks the question structurally
Sources, as carried in the kit's own data: the duty summaries in app/rules.py and the checker data in app/checker_data.py, verified against primary or authoritative sources on October 28, 2026. Run the free duties checker for your jurisdiction, decision type, and the tool's role, and compare the kit's plans for the record, builder, and review queue. More from Ingenious AI is on the hub.
Related guides
- EU AI Act Article 50 Transparency Duties: What Must Be Disclosed, and Since When
- Colorado AI Act Automated Decision Notices: What to Send
- Adverse-Outcome Explanations in Plain Language
- ADMT Deadline Tracker: Which State Duties Land When
See all ADMT guides for the full set.
Frequently asked questions
What does Article 50(3) require?
Deployers of emotion-recognition or biometric-categorisation systems must inform the people exposed to them. The duty has applied since August 2, 2026.
Who owes the duty — vendor or deployer?
The deployer: the business running the system on the people exposed. The vendor's role is to answer capability questions accurately — which is why the audit asks them in writing.
We don't think we use emotion recognition. What should we do?
Audit and file the negative finding: scope, method, per-system evidence, and date. An evidenced 'we checked and found none' is a compliance artefact; an unexamined assumption is not.
What if a capability is present but disabled?
Record the configuration and its date, and make sure vendor updates cannot re-enable it silently — update notice in the contract, and the capability field in your register.
Is there a US equivalent in the kit's data?
No equivalent appears in the kit's Colorado, California, or Connecticut data. This is an EU Article 50 duty and should not be generalised to or from the US duties.