EU AI Act Compliance - Bulk labelling

If you have a lot of images to label, there are much faster routes than going one by one. This page walks through all of them.

Four routes at a glance

1. CSV import – bring in many labels at once from a spreadsheet.
2. Bulk marking per folder – set every media item in a media folder to one status.
3. Folder defaults – label new uploads in a folder automatically.
4. Metadata scan + review queue – review the suggestions found in the image data.

1. CSV import

With a spreadsheet (CSV file) you can label any number of media in one pass.

Step 1: Open EU AI Act Compliance → AI register.

Step 2: Click CSV import.

Step 3: Pick your prepared CSV file and start the import.

How rows are matched to images

Each row says which media item it means. You have three options and pick one per row:

• the technical media ID (exact, if you know it),
• the file name of the image,
• product number and image position (the position is the order of images on the product, starting at 1).

On top of that, every row needs the desired status (required). The tool used is optional.

The columns at a glance

The header row decides what happens, and capitalisation does not matter. For media labels the import knows these columns:

For register entries these apply:

Which of the two kinds you are uploading is detected from the header row, so there is nothing to switch. A file containing status, media_id, media_filename, product_number or image_position counts as media labels; one containing vendor, purpose, labellingStatus, responsible or just name counts as register entries.

A label file therefore looks like this:

media_filename;status;tool
hero-banner.jpg;ai_generated;Midjourney
product-01.jpg;standard_edit;
product-02.jpg;not_ai;

What the format has to provide

• A header row is required. Without one the import stops.
• The delimiter can be a semicolon or a comma – the plugin detects it from the first line, so an Excel export works either way.
• The BOM Excel writes when you save as "CSV UTF-8" is tolerated.
• Quoting follows RFC 4180, so "text; with a semicolon" works.
• Limits: 2 MB and 5,000 rows per file. Split the file if you have more.

When a row does not go through

The import always runs to the end and then shows four numbers – processed, created, updated, skipped – plus a list of the skipped rows with their line number. That number is the line in the file including the header row, so you can jump straight to it in Excel. One bad row does not stop the rest of the import.

Four messages abort the whole import instead, because the file as a whole cannot be processed: Empty CSV content., CSV has no header row., CSV exceeds the maximum of 5000 rows. and CSV exceeds the maximum size of 2097152 bytes. On top of that there is Could not auto-detect CSV type from header columns. – then the header row contains none of the columns listed above, usually because the file was saved with a delimiter other than a semicolon or comma and therefore arrives as a single column.

2. Bulk marking per media folder

This sets every media item in a given media folder to the same status in one go.

Step 1: Open EU AI Act Compliance.

Step 2: In the tools section, click Bulk marking.

Step 3: Pick the media folder.

Step 4: Pick the AI status that should apply to every media item in that folder, and optionally the tool.

Step 5: Confirm. All matching media get the status you chose.

3. Folder defaults for new uploads

With a folder default, every new upload into that folder automatically gets the status you stored – you never have to set it by hand again.

Step 1: Open EU AI Act Compliance → Folder defaults.

Step 2: Click Add default.

Step 3: Pick the media folder and the default status (optionally a tool).

Step 4: Decide whether an existing status should be overwritten or not.

Step 5: Save.

From then on, new images in that folder are labelled automatically.

4. Metadata scan and review queue

On upload the extension reads technical image metadata (C2PA/IPTC, as written by DALL·E, Adobe Firefly and others) and can suggest a status from it. How sure the suggestion is is expressed as a confidence – a value between 0 and 1, the higher the surer.

You have two ways to handle suggestions.

A) Apply automatically

In the configuration you can set suggestions to be applied automatically above a certain confidence (threshold). Anything below that is only suggested.

B) Review them in the queue

Suggestions that weren't applied automatically collect in the review queue.

Step 1: Open EU AI Act Compliance → Review queue.

Step 2: You see a list with file, suggested status, confidence and current status. Use the minimum confidence filter to decide which suggestions are shown.

Step 3: Decide per row:

• Accept – the suggested status is set.
• Reject – the row is hidden, the status stays as it is.

Step 4 (optional): With Accept all above threshold you take over every suggestion that reaches a confidence you choose.

Which route fits when?

• A few specific images: label them individually (see Labelling images & video).
• An existing collection sorted into folders: bulk marking per folder.
• A steady stream of new uploads: folder defaults.
• Data from another system: CSV import.
• Unclear cases: metadata scan + review queue.