Skip to content

README.md

Latest commit

 

History

History
183 lines (143 loc) · 9.7 KB

README.md

File metadata and controls

183 lines (143 loc) · 9.7 KB
Raw

label-crop-check — does the AI label survive the crop?

🤖 Made with AI — no editorial review. This text was produced by AI agents and machine-verified against the official sources. It has not undergone editorial review by a human with relevant subject-matter expertise; the exception in Article 50(4), second subparagraph of the AI Act is therefore not claimed. What was verified and what was not: Provenance.

⚠️ Not legal advice. This playbook is a technical and editorial working aid. The European Commission's guidelines (C(2026) 5054 final) are legally non-binding; only the Court of Justice of the European Union can interpret the AI Act with binding effect. There is no case law on Article 50 yet — this playbook builds a defensible position, not a safe harbour. As of: 24 August 2026.

A zero-dependency CLI (Node ≥ 18) that checks geometrically whether an AI label placed inside an image survives the common platform crops. No image processing, no uploads, no installation — the tool computes with dimensions, not with pixels.

The problem in practice

Article 50(5) AI Act requires the disclosure to be made

"in a clear and distinguishable manner at the latest at the time of the first interaction or exposure."

For images, the Code of Practice recommends the label inside the image, Sub-measure 1.2.2 lit. a:

"in an appropriate place where no intervening overlay elements exist (e.g., in the top right corner of an image or video deep fake)."

That is exactly the corner responsive crops cut away first: feed 1:1, portrait 4:5, story/reel 9:16, header 16:9, thumbnail. A label that is visible on the desktop and missing from the mobile crop is not perceivable where the person first sees the content — and perceivable is what the deployer owes (guidelines paras 141–143). Whoever merely disseminates or transmits — hosting services, online platforms, broadcasters — is not a deployer as long as they hold no authority over the use of the AI (para 16); the platform therefore does not step in here. The machine-readable provider marking under Article 50(2) does not save it: deployers "cannot rely on the machine-readable marking" (para 117).

This tool makes the pre-check before the upload reproducible: same input, same result, verifiable in the ticket, in the PR or in CI.

Usage

node check.mjs --image BxH --label X,Y,BxH [--ratios 1:1,4:5,9:16,16:9] [--focal X,Y] [--json]
node check.mjs --self-test
node check.mjs --help
Argument Meaning
--image BxH dimensions of the original image in pixels, e.g. 1600x1200
--label X,Y,BxH label box in image coordinates, origin top left
--ratios LISTE target ratios, comma-separated; default 1:1,4:5,9:16,16:9
--focal X,Y focal point; the crop is placed around it and clamped to the image bounds (default: image centre)
--json result as JSON instead of a table
--self-test run the built-in self-tests
Exit code Meaning
0 all crops checked show the label in full
1 at least one crop cuts into the label
2 input error (unusable dimensions, label outside the image, malformed ratio)

Example 1 — label top right, default crops

$ node check.mjs --image 1600x1200 --label 1080,40,160x60

label-crop-check — Bild 1600x1200 · Label 1080,40 160x60 · Zuschnitt mittig (Center-Crop)

Ratio  Zuschnitt (BxH @ X,Y)  Status         Labelfläche verloren  Hinweis
1:1    1200x1200 @ 200,0      OK             0 %
4:5    960x1200 @ 320,0       OK             0 %
9:16   675x1200 @ 462.5,0     ABGESCHNITTEN  64.1 %                Label ragt über den Rand des Zuschnitts hinaus
16:9   1600x900 @ 0,150       ABGESCHNITTEN  100 %                 Label liegt vollständig außerhalb des Zuschnitts

Ergebnis: 2 von 4 Zuschnitten schneiden das Label an. Exit 1.
Label näher zur Bildmitte setzen, einen Fokuspunkt vorgeben oder je Zuschnitt eine eigene Fassung ausspielen.

In the feed the label holds, in the story it is clipped, in the header it is gone.

Example 2 — the same image area with a focal point set

$ node check.mjs --image 1600x1200 --label 1150,40,120x60 --ratios 1:1,4:5,9:16 --focal 1310,600

label-crop-check — Bild 1600x1200 · Label 1150,40 120x60 · Zuschnitt um Fokuspunkt 1310,600

Ratio  Zuschnitt (BxH @ X,Y)  Status  Labelfläche verloren  Hinweis
1:1    1200x1200 @ 400,0      OK      0 %
4:5    960x1200 @ 640,0       OK      0 %
9:16   675x1200 @ 925,0       OK      0 %

Ergebnis: Das Label überlebt alle 3 geprüften Zuschnitte. Exit 0.
Geometrische Vorprüfung — die Sichtprüfung am realen Post ersetzt sie nicht.

The focal point is clamped at the image edge (9:16 starts at x = 925, not at 972.5) — that matches the behaviour of crops with a focal or anchor point set.

For scripts and CI, --json returns the same result in machine-readable form (results[].survives, results[].label_area_lost_pct, all_survive, exit_code).

Model and assumptions — honestly

The tool is a pre-check of pure geometry. It knows neither the image nor the platform.

  • Center-crop assumption. For every target ratio the largest possible section is formed and placed centrally — with --focal around the focal point, clamped to the image bounds. Real platforms crop differently: they scale, set anchor points of their own, detect subjects automatically, overlay UI elements on the image and change their formats. An OK here is a reasoned expectation, not a commitment from the platform.
  • "Survives" means: fully inside the crop. The label box has to lie completely within the section; touching the edge counts as visible. A partially visible label counts as cut off — stating what percentage of the label area is lost.
  • No image processing. The tool does not read any image file. Dimensions and label box come from the layout or export step; state them wrongly and the result is wrong.
  • No statement about legibility. Contrast, font size, scaling down to thumbnail dimensions, being covered by play buttons or subtitles — the tool checks none of that, only position and area.
  • The ratio list is a starting value as of 22 August 2026, not a maintained platform catalogue. Your own target formats belong in the call via --ratios, because what is decisive are the channels actually published to.
  • The tool does not replace a visual check. It makes the pre-check reproducible and documentable — the check on the real post remains the last step.

What the tool does not answer

  • Whether a label is required at all. That is settled by the decision tree and the case catalog. For images, audio and video there is no editorial exception — there the deepfake test alone counts; the art exception merely relaxes the form of the disclosure.
  • What the label has to look like and be called. Wording, EU icons and placement per modality are in chapter 04, section 4.1.
  • Whether a label is enough. It is no free pass: misleading practices under the UWG (German Act against Unfair Competition), copyright and personality rights remain untouched by it (chapter 06).

⚠️ If a grey zone remains after the run (for instance: label right at the edge, crop of the target platform unclear) — record the classification with a short justification (see gate/).

Self-test

$ node check.mjs --self-test

label-crop-check — Selbsttest
  ok     1. Zentriertes Label überlebt alle vier Standard-Zuschnitte
  ok     2. Label oben rechts fällt bei 9:16 weg, 1:1 und 4:5 halten es
  ok     3. Fokuspunkt rettet dasselbe Ecken-Label im 9:16-Zuschnitt
  ok     4. Fokuspunkt am Bildrand wird geklemmt — Zuschnitt bleibt im Bild
  ok     5. Randberührung zählt als sichtbar (Label füllt den Zuschnitt exakt)
  ok     6. Degenerat: Label größer als der Zuschnitt — definierter Befund statt Absturz
  ok     7. Degenerat: Label ragt aus dem Bild — sauberer Fehler mit Exit 2
  ok     8. Degenerat: Labelbox ohne Fläche — sauberer Fehler mit Exit 2
  ok     9. Degenerat: Fokuspunkt außerhalb des Bildes — sauberer Fehler mit Exit 2
  ok    10. Unbrauchbare Eingaben (Ratio 1:0, Formatfehler, Unbekanntes) — jeweils Exit 2

10 von 10 Fällen bestanden.

Exit 0 when all cases pass — which lets the tool itself be safeguarded in CI.

Sources