DEV Community

RuleReddy
RuleReddy

Posted on

EU AI Act Article 50(3): emotion recognition and biometric categorisation notices, for developers

If your product reads faces, voices or body signals and outputs a mood score, an "engagement" level, an age bracket or a style category, the EU AI Act has a paragraph written for you. Article 50(3) has applied since 2 August 2026, and it's the Article 50 duty developers think about least, mostly because it sits next to two of the Act's hardest prohibitions.

This post goes through the text, gives you a decision tree you can encode, and shows the notice, the gating and the logging patterns we'd ship.

What Article 50(3) actually says

From Regulation (EU) 2024/1689, Article 50(3):

"Deployers of an emotion recognition system or a biometric categorisation system shall inform the natural persons exposed thereto of the operation of the system, and shall process the personal data in accordance with Regulations (EU) 2016/679 and (EU) 2018/1725 and Directive (EU) 2016/680, as applicable."

Some things to note:

  • It's a deployer duty. The company using the system has to tell people. If you sell an SDK or API, your customers are the deployers, so your job is to make the notice easy to show and hard to skip.
  • "Exposed thereto" is the trigger. That's a wider group than "users". A shopper walking past an in-store camera is exposed. So is a caller on a voice-analytics line.
  • Timing and form come from Article 50(5): the information must be given "in a clear and distinguishable manner at the latest at the time of the first interaction or exposure" and "conform to the applicable accessibility requirements."
  • There's one carve-out: systems "permitted by law to detect, prevent or investigate criminal offences", with safeguards. Commercial products almost never fall under it.

The Digital Omnibus on AI, Regulation (EU) 2026/1744, did not change Article 50(3). It added a four-month transition only for the Article 50(2) marking duty, so 50(3) applies on the original 2 August 2026 date.

Is your feature even in scope? The definitions

Emotion recognition system (Art. 3(39)): "an AI system for the purpose of identifying or inferring emotions or intentions of natural persons on the basis of their biometric data."

Recital 18 narrows it down in ways that matter for product scoping:

  • Emotions include things like "happiness, sadness, anger, surprise, disgust, embarrassment, excitement, shame, contempt, satisfaction and amusement."
  • Physical states such as pain or fatigue are excluded, for example driver-fatigue detection.
  • "Mere detection of readily apparent expressions" (a smile, a frown, a raised voice) is excluded "unless they are used for identifying or inferring emotions."

So a smile_detected: true flag in a camera app is probably out of scope. Once you map it to customer_satisfaction: 0.82, you're inferring emotion.

Biometric categorisation system (Art. 3(40)): "an AI system for the purpose of assigning natural persons to specific categories on the basis of their biometric data, unless it is ancillary to another commercial service and strictly necessary for objective technical reasons."

Recital 16 lists categories such as "sex, age, hair colour, eye colour, tattoos, behavioural or personality traits", and it names the ancillary exception's classic examples: virtual try-on filters on online marketplaces and face/body filters on social networks. The feature must be impossible to use without the main service, and it can't be a way "to circumvent the applicability of the rules."

Biometric data (Art. 3(34)) is "personal data resulting from specific technical processing relating to the physical, physiological or behavioural characteristics of a natural person, such as facial images or dactyloscopic data." Voice and gait analysis can qualify too, because they are behavioural or physiological characteristics.

Check the prohibitions before you build the notice

Article 50(3) only matters if the use is allowed in the first place. Two Article 5 prohibitions have applied since 2 February 2025 (Art. 113(a)):

  • Art. 5(1)(f): AI systems "to infer emotions of a natural person in the areas of workplace and education institutions, except where the use of the AI system is intended to be put in place or into the market for medical or safety reasons."
  • Art. 5(1)(g): biometric categorisation systems that categorise individuals "to deduce or infer their race, political opinions, trade union membership, religious or philosophical beliefs, sex life or sexual orientation" (with narrow exceptions for labelling lawfully acquired datasets and for law enforcement).

A notice doesn't make a prohibited use legal. Fines for Article 5 go up to €35M or 7% of worldwide annual turnover (Art. 99(3)). Fines for Article 50 go up to €15M or 3% (Art. 99(4)(g)). For SMEs and start-ups, the lower of the two figures applies (Art. 99(6)).

On top of that, Annex III point 1(b) and (c) lists biometric categorisation by sensitive attributes and emotion recognition as high-risk. After the Omnibus, the high-risk requirements for Annex III systems apply from 2 December 2027 (amended Art. 113(c)(i)). Article 50(3) applies now either way.

Encode the decision tree

We like turning this into a config check that runs in CI and at startup, so nobody can turn on an emotion model for an HR customer by flipping a flag.

# ai-features.yaml
features:
  - id: call-sentiment
    model: voice-affect-v3
    infers: emotion            # emotion | biometric_category | expression_only | none
    input: biometric_voice
    contexts: [customer_support]
    notice_id: notice-voice-affect-2026-08
  - id: tryon-face-shape
    model: face-shape-v1
    infers: biometric_category
    input: biometric_face
    ancillary_to: virtual-try-on   # Art. 3(40) / recital 16 exception, document why
    contexts: [ecommerce]
Enter fullscreen mode Exit fullscreen mode
// scripts/check-ai-act-50-3.ts
import { readFileSync } from "node:fs";
import { parse } from "yaml";

const PROHIBITED_EMOTION_CONTEXTS = ["workplace", "education"]; // Art. 5(1)(f)
const PROHIBITED_INFERENCES = [                                    // Art. 5(1)(g)
  "race", "political_opinion", "trade_union", "religion_belief",
  "sex_life", "sexual_orientation",
];

type Feature = {
  id: string; infers: string; contexts: string[];
  categories?: string[]; notice_id?: string; ancillary_to?: string;
  medical_or_safety?: boolean;
};

const { features } = parse(readFileSync("ai-features.yaml", "utf8")) as { features: Feature[] };
const errors: string[] = [];

for (const f of features) {
  if (f.infers === "emotion") {
    const bad = f.contexts.filter(c => PROHIBITED_EMOTION_CONTEXTS.includes(c));
    if (bad.length && !f.medical_or_safety)
      errors.push(`${f.id}: emotion inference in ${bad.join(",")} (Art. 5(1)(f))`);
  }
  if (f.infers === "biometric_category") {
    const bad = (f.categories ?? []).filter(c => PROHIBITED_INFERENCES.includes(c));
    if (bad.length) errors.push(`${f.id}: infers ${bad.join(",")} (Art. 5(1)(g))`);
  }
  const needsNotice =
    f.infers === "emotion" || (f.infers === "biometric_category" && !f.ancillary_to);
  if (needsNotice && !f.notice_id)
    errors.push(`${f.id}: no exposure notice configured (Art. 50(3))`);
}

if (errors.length) { console.error(errors.join("\n")); process.exit(1); }
console.log(`AI Act 50(3) check passed for ${features.length} features`);
Enter fullscreen mode Exit fullscreen mode

medical_or_safety: true should need a written justification in code review. It's an exception, not a toggle.

The notice: before the sensor starts, not after

"At the latest at the time of the first interaction or exposure" means the notice has to come before or alongside the first frame or audio chunk the model sees. It can't be in a privacy policy link the person finds later. In a browser, that means gating getUserMedia:

function AffectNotice({ onAccept }: { onAccept: () => void }) {
  return (
    <div role="dialog" aria-labelledby="affect-title" aria-describedby="affect-body">
      <h2 id="affect-title">This session uses emotion recognition</h2>
      <p id="affect-body">
        An AI system analyses your facial expressions to estimate emotions
        (for example, frustration or satisfaction) during this session.
        Results are used to {/* purpose */}improve support quality.
        <a href="/privacy#emotion-recognition">How your data is processed</a>
      </p>
      <button onClick={onAccept}>Continue</button>
    </div>
  );
}

async function startCamera(noticeShown: boolean) {
  if (!noticeShown) throw new Error("Art. 50(3) notice must be shown before capture");
  return navigator.mediaDevices.getUserMedia({ video: true });
}
Enter fullscreen mode Exit fullscreen mode

Practical notes:

  • Name the system type in plain words: "emotion recognition" or "estimates your age range", not "enhanced AI insights". The duty is to inform people "of the operation of the system."
  • Accessibility is part of the duty (Art. 50(5)). Use a real dialog with ARIA labelling, keyboard focus, readable contrast, and an audio version for voice-only channels. For phone lines, play a short spoken notice before analysis begins.
  • Physical spaces need physical notices. For in-store cameras doing demographic categorisation, put signage at the entrance, before people enter the field of view.
  • An Article 50(3) notice is not GDPR consent. The Article 50(3) text points separately to GDPR for the processing itself, so you still need a lawful basis and, for special-category data, an Article 9 condition. Treat these as two jobs.

Log that the notice happened

If a regulator or customer asks, you want proof that every analysed session was preceded by a notice. Keep the log separate from the biometric data itself, and don't store embeddings in it:

type ExposureEvent = {
  ts: string;               // ISO 8601
  session_id: string;       // pseudonymous
  feature_id: "call-sentiment";
  notice_id: "notice-voice-affect-2026-08";
  channel: "web" | "phone" | "in_store";
  notice_delivered_before_capture: boolean;
};

function logExposure(e: ExposureEvent) {
  if (!e.notice_delivered_before_capture) alertOnCall(e);  // treat as an incident
  appendJsonl("/var/log/ai-act/exposure.jsonl", e);
}
Enter fullscreen mode Exit fullscreen mode

Version your notice text (notice_id) so you can show exactly what someone saw on a given date.

If you're the vendor, not the deployer

Article 50(3) is formally on your customer, but you can make compliance the default:

  1. Ship the notice component and spoken-notice audio with your SDK.
  2. Refuse to start inference unless the client sends a notice_id and a timestamp.
  3. Expose an intended_context parameter, and reject workplace and education for emotion endpoints unless the contract covers a medical or safety use.
  4. Document in your docs whether your feature counts as "ancillary" under recital 16, and why.

Quick checklist

  • [ ] Each biometric feature is classified: emotion / category / expression-only / none
  • [ ] No emotion inference in workplace or education contexts (Art. 5(1)(f))
  • [ ] No sensitive-attribute categorisation (Art. 5(1)(g))
  • [ ] Notice shown at or before first exposure, accessible (Art. 50(3), 50(5))
  • [ ] Separate GDPR lawful basis documented
  • [ ] Exposure log with versioned notice IDs
  • [ ] Annex III high-risk readiness on the roadmap for 2 December 2027

This article is general technical information, not legal advice. Check your specific situation with qualified counsel.

If you want a quick read on which Article 50 duties apply to your product, chatbot disclosure, content marking, deepfake labels or the 50(3) notices covered here, RuleReddy has a free Article 50 check at rulereddy.com.

Sources

Top comments (0)