Amazon’s New Rule for AI-Generated Person Is Here: This Time It’s Not a Suggestion, It’s a Fine

If you’ve recently seen posts in seller groups like “AI images got taken down” or “$5,000 fine,” no need to second-guess it. It’s real.

With New York State’s Synthetic Performers Disclosure Act (SB 8420-A) now officially in effect, Amazon has updated its backend rules as well: if your main product image, gallery images, or A+ page contains realistic AI-generated people, you must add a disclosure. If you fail to do it properly, the lighter consequence is a policy violation notice or listing removal. The heavier consequence is a fine of up to $5,000 per violation. If you run a multi-SKU catalog model, that risk can multiply fast.

For cross-border sellers who have long relied on AI model photos and AI lifestyle images to cut shooting costs, this rule probably feels abrupt. But once you break it down, it’s not as scary as the rumors make it sound. What matters is getting three things clear: whether your image actually needs labeling, where the label should go, and how to update your existing image library with the least effort.

1. Don’t panic, first figure out which images actually need disclosure

When this rule first came out, the most common misunderstanding was: “Any image made with AI has to be labeled.” That is not accurate. What actually requires disclosure is a realistic person generated by AI appearing in the image, including:

• AI-generated models in main images or secondary images, even if only part of the body is shown, such as a hand wearing a ring, an arm displaying skincare products, or a shoulder showing how clothing fits;
• AI-generated real-person model images on A+ pages or brand storefronts;
• AI digital humans, AI voiceover presenters, or AI-generated “customer showcase” people appearing in product videos.

On the other hand, the following cases do not require labeling:

• The image was shot with a real person, even if AI was later used for retouching or skin smoothing;
• The character is a fictional figure from film, TV, or games;
• There is no person in the image at all, or the person is clearly in a cartoon or illustrated style and is obviously not meant to simulate a real human.

The standard is not “whether the face is shown.” It is “whether AI was used to generate body parts that look like a real person.” Many sellers got this wrong at first and assumed they were safe as long as the face was hidden. In practice, close-ups of hands and arms are often the easiest to catch in spot checks.

2. How exactly should the label be added, and will it affect the image itself?

The second common mistake is thinking that “labeling” means adding a watermark or visible text onto the image. That is also what many sellers disliked at first. After spending time making polished visuals, nobody wants a huge “AI-generated” stamp pasted across the image.

That’s not how it works. The disclosure Amazon requires is written into the image file’s metadata. Technically, it goes in the dc:subject field in XMP, using the keyword contains-synthetic-performer. This is a hidden tag embedded in the file information. The image itself does not change at all, and customers will not see any visible text or watermark. Once Amazon’s system detects the tag, it will automatically show customers a notice that the image contains an AI-generated person, and your compliance obligation is considered fulfilled.

Also, if you upload assets through A+ Content Manager and check the “AI-generated content declaration” option during upload, you do not need to add a separate metadata tag. You only need to choose one of those two methods. But if the image is used directly as a main image, gallery image, or uploaded through other channels, then you do need to add the tag properly inside the file.

3. The real headache is your existing image library

New images are easy enough to handle. You can just ask your designer to add the tag during export. The real pain point is existing listings that have already been live for months or even years. This hits especially hard if one listing has more than ten images, or one account manages hundreds of SKUs. Checking images one by one and editing metadata manually becomes a huge workload, and it is very easy to miss some.

At that point, sellers usually see two options in front of them:

The first is manual processing image by image: find each image, open it in professional metadata editing software, add the keyword one at a time, and then verify one at a time that the tag was successfully written. Once the image count grows, just tracking which ones were already fixed and which ones still need work can become overwhelming.

The second is uploading images to a third-party online tool for batch processing: this saves the effort of editing metadata by hand, but many sellers hesitate for a reason. If images have to be uploaded to someone else’s server, could product selection images or exclusive creative assets be retained? Could account-related information be collected along the way? For unreleased new product images that need to stay confidential, that concern is valid.

If you’re stuck between those two choices, there is actually a third option that is easier to manage: process everything on your own computer without sending files through any server. We ran into the same problems when cleaning up the existing images in our own account, so we ended up making a free tool and sharing it with other sellers dealing with the same update: AI-Generated Person Tagging Tool.

It’s simple to use, just four steps: import product images that contain AI-generated people, with support for dragging in an entire folder at once → check the images you want to process in the list → click “Validate and Tag,” and the tool will write the disclosure tag locally on your device and automatically verify the result → after processing, click “Export Tagged Images” to download them as a package, ready to use for listing updates.

A few details sellers usually care about:

• No image upload: the entire process runs locally in your browser. Not a single byte of your images leaves your computer, and there is no need to authorize or connect an Amazon account;
• The image stays exactly the same: the tag is written only into the hidden metadata area. There is no recompression and no watermark. The exported image is pixel-for-pixel identical to the original;
• Already tagged images are skipped automatically: if an image already contains the disclosure tag, the tool will mark it as “Already Tagged” after validation, without reprocessing it or causing errors;
• Exceptions are flagged separately: if the tool encounters an unsupported format or an image that fails write verification, it places that file into an “Exceptions” category for easier review, without interrupting the processing of other images. Supported formats are also being expanded continuously.

Final note

This new rule is really one small example of a broader shift toward AI compliance in cross-border ecommerce. From image disclosures to title length limits, Amazon is turning “AI use must be transparent” into more specific and enforceable rules. Instead of waiting for a spot check and then scrambling, it makes more sense to review your existing assets now and turn image tagging into one standard step in your product launch workflow.

Rules will keep changing, but one approach stays practical: get the things you can confirm today done properly first.