Compliance

Verankila is the kind of app that has to earn trust structurally, not just promise it — it reads your face.

EU AI Act — emotion recognition system

Verankila voluntarily classifies itself as an emotion recognition system under Article 3(39) of the EU AI Act (Regulation (EU) 2024/1689): it infers tentative emotional states from facial biometric data by automated means. This classification triggers the transparency and prohibited-use obligations below.

Article 50 — transparency

• Active-use indicator: a persistent, visible on-screen indicator is shown whenever the camera/emotion-recognition pipeline is running.
• Tentative framing: every reading is presented as a tentative cue ("this expression reads as...", "does this fit?") — never as a fact or a clinical determination. Language implying certainty or clinical determination is avoided throughout, enforced by an automated build check on the app's own copy.
• Consent gate: a consent screen at first run informs you that Verankila is an AI emotion-recognition system before the camera is ever used.

AI-generated reference faces

Verankila's expression-practice feature shows reference faces — one way a given expression can look. Every one of them is AI-generated. They are labelled as such wherever they appear, in all seven supported languages, including the label a screen reader announces.

• No real person is depicted. Each image was generated from a written description of an expression and a demographic — no photograph was used, and no reference image. "Bears no resemblance to an identifiable real person" is one of the checks a human reviewer has to record for each individual image before it can ship.
• Nothing is generated on your device. The images were produced offline, in the build pipeline, by FLUX.1-schnell (Apache-2.0). There is no image-generating model in the app; these are fixed picture files that a person has reviewed one by one.
• Each image records how it was made. Model, seed, and the exact prompt travel with the image, so any of them can be reproduced and checked rather than taken on trust.
• They reach a public release only once reviewed. Two sign-offs are required per image — one on rendering quality and whether the face reads as unambiguously an adult, one from a clinician on whether the expression is plausible and not caricatured for the demographic shown. An automated release gate refuses to ship any image missing either sign-off, or whose bytes have changed since it was approved.

Whether Article 50(2) or 50(4) of the EU AI Act attaches here is a question for counsel and is not settled on this page: 50(4) concerns content resembling real people, places, or events, and 50(2) concerns providers of generative systems marking their output, while Verankila generates nothing at runtime and uses a third-party model offline. The disclosure above is in place either way — someone looking at a synthetic human face should be told that it is synthetic, whether or not a regulation compels it.

Article 5 — prohibited contexts

Verankila is not intended for workplace or educational deployment, which the EU AI Act prohibits for emotion-recognition systems outside medical/safety contexts. A use-context gate at onboarding is designed to refuse those contexts. Verankila is built for personal, individual use only.

GDPR data minimization

Real-time processing uses only 52 blendshape coefficients per frame — not raw camera frames or a full facial mesh — held in a short rolling window and discarded when a reading session ends. Session recording (face crop, blendshapes, readings) happens only if you explicitly opt in, and stays on your device. See Privacy for the full picture.

Scientific limitations

Verankila's classifier reaches approximately 50% accuracy on a held-out test split — well above chance for seven categories, but not clinically reliable. A reading does not and cannot indicate why you feel something; no inference from face to life circumstances is made or implied. Cross-demographic fairness (accuracy across skin tone, age, and gender presentation) has not yet been formally evaluated — this is required before any public release and is tracked as open engineering work.

What we don't do

• No cloud or server-side processing of your face
• No stock or third-party photographs of anyone — the only photographic images are the ones your own camera makes, and they stay on your device
• No clinical assessment or treatment claim
• No workplace or education deployment
• No accounts, analytics, or advertising
• No third-party data sharing