AI transparency

Scoreflection uses artificial intelligence for three things: hearing what you play, reading a photographed score, and translating text. It uses it for nothing else. This page says which models those are, where they run, what they get wrong, and what they are never used for.

It is the information required by Art. 50 of the EU AI Act (Regulation (EU) 2024/1689). The same text is in the app, under About → Artificial intelligence. Section 9 is the exception: it is about this website rather than the app.

1. Where the models run

• Listening — on your device. When you play, a note-detection model turns the sound into notes: All models run inside the app on your phone. The audio is analysed as it arrives and never leaves the device.
• Scanning a score — on our server. A photo or PDF page you choose to scan is uploaded to the Scoreflection server and read by an optical music recognition model. It is used for that one transcription. The image is not kept afterwards, is never added to the shared library, and is never sent anywhere else.
• Translation — third-party services. Posts in the community forum can be machine-translated on request. Knowledge-base articles were machine-translated when the app was built.

None of these models learns from your playing. The only audio ever used to improve a model is a clip you deliberately send with the audio-contribution switch in Settings → Data Sharing, which is off until you turn it on.

2. What these models get wrong

They are pattern recognisers, not readers, and they are sometimes confidently wrong. Knowing where the edges are is more useful than a headline accuracy figure, so:

• A scan is a best-effort transcription, not a copy. Expect wrong, missing and occasionally invented notes. Accuracy falls on phone photos, on dense piano writing, on handwritten music, and further down a page that has a lot of music on it — the model reads a whole page at once and gets tired towards the end of a busy one. Expression marks (hairpins, dynamics, articulation) are recovered far less reliably than the notes themselves. Always compare a scan against the original before you play from it, print it or pass it on.
• Note detection while you play depends on your room. Microphone quality, background noise and an out-of-tune instrument all cost accuracy. A note shown as wrong may simply not have been heard.
• Machine translation can change meaning. Where a translated text has legal effect, the original version governs.

3. AI output is marked as AI output

Every score produced by a scan carries the disclosure inside the file itself, in the document's own metadata — not only on the screen that created it. If you export the score, open it in another notation program or send it to someone, the marking travels with it and says what produced it and that no person has checked it.

4. What Scoreflection never does

These are design decisions, and they are why most of the AI Act's harder questions do not arise here:

• No emotion recognition, and no inference of mood or state of mind from your playing, your voice or your face.
• No biometric identification and no biometric categorisation. Audio you contribute is used to improve note detection; it is never used to recognise or characterise you.
• No profiling, scoring or ranking of people. (The highscore output is a strictly point-based evaluation of in-game performance mechanics and does not involve personal data.)
• No automated decision with legal or similarly significant effect (Art. 22 GDPR). Nothing an AI model produces changes your account, your subscription or your access to anything.
• No manipulative or subliminal techniques, and nothing designed to exploit the age or inexperience of a child. Practice streaks and badges are visible, optional and can be ignored without losing access to any feature.
• No advertising, no ad SDKs and no third-party trackers, so there is no profile being built to sell.

5. Practice feedback is information, not assessment

Scoreflection marks your playing bar by bar and keeps statistics on what you practised. That is feedback for you. It is not used to grade you, to unlock or withhold content, or to report on you to anybody — and Scoreflection is not used by schools or examiners to decide admission, progression or results. If that ever changes, it would fall under Annex III of the AI Act as a high-risk use, and this page would say so before it happened.

6. You stay in charge

Nothing a model produces is applied behind your back. A scan is saved to your library only once you accept it, and you can edit or delete any transcribed score afterwards. You can ignore any feedback the listener gives you, turn the microphone off, and use the whole sheet-music library without an account.

7. Training data

The score-recognition model was built on own data, as well as on public-domain engravings and on photographs of them that we produced ourselves, together with the openly licensed research datasets its authors published. It was not trained on user content, and the shared Scoreflection library is public-domain music by design.

8. This website

Everything above is about the app. This section is about the page you are reading, because some of what is on it was made with AI as well.

• The illustrations are AI-drawn and hand-finished. The icons, the section dividers and the error pages were drawn with an AI design tool and then reworked by hand — recoloured to the app's palette, redrawn where the geometry was wrong. They are stylised artwork; they depict nothing that exists.
• The screenshots are real. Every picture of the app on this site is an unretouched capture from a device running it: cropped and resized, and nothing else. Never generated, never composited, never tidied up by a model. Where one shows a score marked up after a session, those are the marks the app made. This is a rule we hold ourselves to, not a description of the pictures that happen to be up today.
• The text is written by a person, this page included.

None of it carries a mark on the image itself, because the case Art. 50 (4) of the AI Act is aimed at does not arise here: nothing on this site is artificial content resembling a real person, place or event closely enough that you could take it for the real thing. So we say it in words instead.

9. Questions

If something here is unclear, or you think a model has done something this page does not describe, tell us — see Contact. Security problems have their own route: Product security.

10. Changes

We update this page when the product changes. The date at the top is the last change.