Academy›Special courses: eight courses for day-to-day selling›AI images and video: product accuracy, labelling and rights
Special courses: eight courses for day-to-day sellingAI in everyday selling

AI images and video: product accuracy, labelling and rights

Course lesson 2/3 ⏱ ~15 Min. 🏅 up to 215 points By Enes Kurt Updated November 2026 Check your listing free →
What you'll take away

An image model builds a kitchen scene in a minute that used to take a photo studio a full day. Lifestyle images and use-case scenes now cost almost nothing. But the same model can quietly turn your grinder into a different grinder, invent a person who needs a label, or put someone else's logo in the background. This lesson shows where AI helps your image stack, where the limits are and which labelling applies since August 2026. The basics for the main image and gallery are in the lessons on the main image and the image stack, the legal overview in lesson 15. This one is about working with AI.

1

Main image and gallery: how far AI may go

Amazon judges images by the result: they must show the product the customer receives. For the main image Amazon requires the actual product being sold on a pure white background, without graphics, illustrations or placeholders. Product safety law adds to that: the EU General Product Safety Regulation requires online offers to include information that identifies the product, explicitly including a picture of it (Article 19). An image showing a different product than the one delivered is therefore more than a marketing problem.

ImageWhat AI may doWhere the line is
Main imageCut out, remove dust and scratches from the sample, correct light and white balance, set the background to pure whiteNever regenerate the product: shape, color, surface and contents stay as photographed
Use case and lifestyleGenerate room, light, season and propsProduct unchanged and at the right scale, no function it doesn't have
InfographicLayout ideas, backgrounds, iconsEvery number and every line of text comes from the fact sheet, not the model
Size comparisonHand, countertop, kitchen settingScale measured, dimensions as text from the fact sheet
VideoScenes, camera moves, transitionsProduct unchanged in every frame, no faked functions

The rule of thumb from the main image lesson applies to the whole stack: AI as a tool yes, AI as an inventor no. What's new is how much a tool can do now, and therefore how easily it becomes an inventor by accident.

2

Product accuracy: the real product is the fixed point

Image models redraw the product on every run, even when you give them a photo as reference. Details shift that you don't see at first glance: proportions, the shape of a button, the color of an edge, the lettering on the housing. So work in this order:

  1. Real reference photos: the production unit, not the pre-production sample, from at least four angles, in even light, with a tape measure in the frame for scale.
  2. Insert instead of generate: where your tool allows it, place the cut-out real photo into the generated scene instead of letting the model redraw the product. The AI then builds only the setting, shadows and reflections.
  3. Compare like an inspector: new image and reference photo side by side, at full resolution, feature by feature.
  4. Document the approval: one row in the image register (see the expert chapter).
Product accuracy per image
  • Shape and proportions match the reference photo.
  • Color and surface are right (brushed stainless steel stays brushed, doesn't turn chrome).
  • Controls identical in number, position and shape.
  • Lettering and logo legible and correct, no invented characters.
  • No more parts in the image than in the box.
  • Scale against hand, pan or countertop calculated.
  • No faked function: no light, display or cable that doesn't exist.
Example

Test run with the AURELO set: from one reference photo, an image model generates ten kitchen scenes in which it redraws the product itself. Four fail the comparison: twice an LED light ring at the burr that the set doesn't have, once a black instead of silver button, once three grinders on the counter instead of two. A fifth scene looks right until you do the math: the grinder is 19.5 cm tall, the pan next to it is 28 cm across, but in the image the grinder looks almost as tall as the pan is wide, so roughly 26 cm. The buyer expects a third more grinder than they get. Every “smaller than expected” return costs about €4.65 to €9.15 on the €24.99 set with a €4.50 purchase cost (calculation from the bullet lesson). Five of ten scenes remain. With the original photo inserted instead of a redrawn product, far fewer scenes fail in practice.

3

Scenes and lifestyle: what AI is really good for

AI plays to its strength in the setting: scenes that show what the product is for, in variants a photo shoot could never deliver economically.

  • Make use cases visible: barbecue evening, small kitchen, gift on the dining table. The use cases from your bullets get a picture, and copy and gallery tell the same story.
  • Seasons and markets: Christmas decor in November, a patio in May, a kitchen that fits the target market (sockets, countertops, food).
  • Consistent look: define light, color palette, angles and materials once and use the same settings for every scene. Otherwise the gallery looks like it came from five different shops.
  • Props with care: anything in the frame can be mistaken for contents. Never show a second device that looks like an accessory.
Tip

If an image only needs to carry a mood, generate the scene slightly stylized rather than deceptively real, for example with a soft background or graphic elements. A recognizably designed scene doesn't appear to anyone as a real photograph and needs no label under the EU AI Act. A deceptively real photo of a situation that never happened may well need one.

4

People in AI images

Hands seasoning a dish, guests at a barbecue: people bring use-case images to life, and they are the area with the most rules.

  • Amazon's labelling: since July 2026 Amazon requires, in all its stores worldwide, the keyword contains-synthetic-performer in the dc:subject metadata field for images, videos and A+ content showing photorealistic, fully AI-generated people. Many image editors call that field “Keywords”. Amazon uses it to show shoppers a disclosure. Not covered are real people, even if edited with AI, recognizably artificial characters and images without people.
  • EU AI Act: a photorealistic AI person who appears to be a real customer is usually a deepfake under Article 50. Amazon has not said whether its shopper disclosure fulfils that duty. According to the Commission's FAQ, a machine-readable marker alone is not enough, so add a visible notice as well (next chapter).
  • Real models: if you alter real photos of people with AI, such as clothing or setting, you need consent that covers such edits. Check the model release before placing existing photos into new scenes.
  • Resemblance to real people: a generated face that resembles a well-known person touches their personality rights and is a textbook deepfake at the same time. Never name real people in prompts, and discard results with a noticeable resemblance.
  • Hands: too many fingers or impossible grips cost more trust than any mediocre photo. Check hands at full resolution.
In plain terms

An AI person in a product image is like an actor in a TV ad: perfectly fine, as long as nobody believes they are watching a real customer filmed by chance. The label is the small note in the credits that makes this clear. Without it, the staging becomes deception.

5

Labelling under the EU AI Act

Since 2 August 2026 the transparency obligations of Article 50 of the EU AI Act apply. The Digital Omnibus (Regulation (EU) 2026/1744, in force since 27 July 2026) did not postpone anything for deployers. What counts for you as a seller is Article 50(4): anyone who uses AI to generate or manipulate image, audio or video content that qualifies as a deepfake must disclose it.

Under Article 3(60), a deepfake is content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful. Objects are named explicitly, so a product image can qualify even without people. This is how to place the notice correctly:

  • Recognizable at first exposure: in the image or right next to it, not in the terms, legal notice or a footnote. In a listing that means in practice: a small, legible notice in the image itself, such as “AI-generated scene”. Text is prohibited on the main image, one more reason to use only real photos there.
  • Perceivable without tools: a tool's watermark or metadata do not fulfil your duty.
  • Help from the Commission: the voluntary code of practice on labelling (June 2026) offers a common EU icon for such notices, and guidelines on Article 50 followed in July 2026.
  • Existing content: according to the Commission, content created before 2 August 2026 has no retroactive labelling duty. Labelling it anyway is still recommended.
  • Fines: up to €15 million or 3% of worldwide annual turnover, for small and medium-sized companies whichever is lower.
Image in the AURELO setAmazon keywordNotice under Art. 50
Photorealistic AI cook seasoning with the grinderyesyes, visible in the image
Real product photo, cut out and cleaned of dust with AInono, the image shows a real photograph
Real grinder in a photorealistic AI kitchen, no peoplenocase by case: rather no if the scene clearly sets a mood; yes if it fakes a real situation (customer photo, test, before and after)
Recognizably drawn illustration of the burrnono
Important

A label does not cure deception. If an AI image shows the grinder in a color that doesn't exist, it stays misleading, with or without a notice. And the middle case in the table is legally young: Commission guidelines are not law, and courts have hardly ruled on it. If you're unsure, label it, or ask a lawyer specializing in IT or competition law when in doubt.

6

AI video: same rules, more sources of error

Video sells, and AI makes it cheap. Since November 2025, Amazon Ads in Germany offers a video generator for Sponsored Brands video campaigns that creates several short clips from one product image, and AI image features for ad creatives have been there longer. Product videos on the detail page follow the same principles as images, except you are not checking one picture but hundreds.

  • Product in every frame: video models deform products between frames: a button moves, the housing briefly widens, lettering melts. Watch the video at half speed and pause at every cut.
  • Physics: pepper drifting upward, salt falling out of a closed burr: anything physically wrong looks dubious and can fake a function.
  • No faked operation: a video in which the grinder starts with a tap on the lid when it actually has a push button describes a different product.
  • Sound and voice: a synthetic voice-over that resembles a real, well-known voice falls under the deepfake rules. Music from a tool only with a clear usage license. Amazon's keyword applies to AI people in video just as in images.
  • Usage rights: read the terms of every tool. Some allow outputs only in specific channels.
Most common mistake

Checking only the thumbnail. With images the check is done in a minute, and with videos many people rely on the first impression. Yet the errors arise precisely in the motion: the grinder looks right in the first and last frame and has a second button at second seven. A video gets checked scene by scene, not by gut feeling.

7

Copyright and trademark law

  • Purely AI-generated images are usually not protected. German copyright requires a personal intellectual creation by a human (Section 2(2) UrhG). The Munich local court ruled on 13 February 2026 (142 C 9786/25) that AI-generated logos were not protected: even detailed prompts and choosing among several suggestions were not enough. For you this means: if a competitor copies your fully generated scene, you often have no copyright claim.
  • Your real photos are your strongest right. In Germany, photos are protected as “Lichtbilder” even without artistic merit (Section 72 UrhG). A scene with your real product photo inserted therefore contains a protected part. One more reason to work with your own photos instead of redrawn products.
  • No third-party images as reference. Never upload competitors' photos, unlicensed stock images or search engine results as a reference to generate something similar. Whether the result infringes someone else's work is a dispute you don't want.
  • Third-party brands in the frame. Image models like to put well-known logos on kitchen appliances, bottles or clothing in the background. A recognizable third-party logo in your ad image can infringe trademark rights and is a typical trigger for complaints (brand protection). Check backgrounds deliberately and never name brands or “in the style of” instructions in prompts.
  • The tool's terms. Who owns the outputs and whether you may use them commercially is governed by the provider's terms of use. Keep the version that applied when you generated the images.

This lesson is not legal advice. For a concrete case, a lawyer specializing in copyright and media law is the safe call when in doubt.

8

Expert insight: metadata and the image register

Besides pixels, every image carries metadata, and that is becoming the layer of proof. Three standards matter for you:

  • IPTC Digital Source Type: describes how an image was made. The key values: digitalCapture (camera photo), trainedAlgorithmicMedia (fully created with generative AI), compositeWithTrainedAlgorithmicMedia (edited with generative AI, for example by extending or replacing parts of the image), compositeSynthetic (composite with at least one AI element), algorithmicallyEnhanced (algorithmically improved, content unchanged) and digitalCreation (made by a person without generative tools).
  • C2PA (Content Credentials): an open standard for a cryptographically signed provenance record inside the file: which tool, which edits. Many image models write it automatically.
  • dc:subject: the keyword field where Amazon requires contains-synthetic-performer.

The trap: metadata rarely survives the workflow. Many “export for web” functions strip it, some editors overwrite it, and a C2PA signature becomes invalid as soon as the file is changed with a tool that doesn't support C2PA. So check the final export file, not the working file, with a metadata tool. Amazon does not document whether metadata survives in the file it serves after upload, so keep your originals. The machine-readable marking under Article 50(2), by the way, is a duty of the image model providers (for systems on the market before 2 August 2026 with a transition period until 2 December 2026), and it does not replace your visible notice.

Example

The AURELO set's image register is a spreadsheet with one row per file: file name with version, IPTC value, tool and date, reference photo, reviewer, Amazon keyword yes/no, visible notice yes/no, live from/to. For seven images and one video that's eight rows. Image 1 (main image, cut out): digitalCapture. Images 2 and 3 (real photo in an AI kitchen): compositeWithTrainedAlgorithmicMedia. Images 4 and 5 (infographics): digitalCreation. Image 6 (AI cook): trainedAlgorithmicMedia, keyword and visible notice. Image 7 (size comparison with a real hand): digitalCapture. Setting it up takes about ten minutes, after that one row per new file. If Amazon asks or a cease-and-desist letter arrives, you have the answer in a minute.

When you replace an image, upload it under a new file name and note in the register when which image was live. If questions come up, what counts is what was visible on a specific day.

Continue reading for free

The first lessons of every track are open to everyone. From here on you just need a free account: no subscription, no costs.

Still ahead in this lesson:
  • 2Product accuracy: the real product is the fixed point
  • 3Scenes and lifestyle: what AI is really good for
  • 4People in AI images
  • 5Labelling under the EU AI Act
  • 6AI video: same rules, more sources of error
  • 7Copyright and trademark law
  • 8Expert insight: metadata and the image register
  • ✓Quiz: 6 questions with instant feedback
✎ Practice

Your turn: 2 exercises

Solve them to unlock the quiz. Up to 30 points each when you get it right first time without a hint.

1
Sort it

Which labelling does the image need?

You are going through the new gallery and video for the AURELO spice grinder set, as of after 2 August 2026. Match each file to the labelling it needs: Amazon's keyword contains-synthetic-performer, the visible notice under Article 50 of the EU AI Act, or neither.

Photorealistic AI cook seasoning with the grinder
Real product photo, cut out and cleaned of dust with AI
AI scene without people, styled like a real customer photo of the grinder
Recognizably drawn illustration of the burr
AI video with a photorealistic, fully generated barbecue guest
Photorealistic before and after image without people, fully AI-generated
Infographic with dimensions from the fact sheet, designed on a computer
Recognizably artificial cartoon character holding the grinder
2
Put in order

Put the workflow for an AI scene in order

You want to show the AURELO set in an AI kitchen scene without the product changing. Put the steps in the right order.

  1. Document the approval as a row in the image register
  2. Upload the image under a new file name and record in the register from when it is live
  3. Photograph the production unit from at least four angles with a tape measure in the frame
  4. Cut out the real product photo
  5. Place the cut-out photo into the AI kitchen scene, the AI builds only the setting and shadows
  6. Compare the new image with the reference photo at full resolution, feature by feature

Check yourself

6 quick questions, one at a time, with instant feedback. With a free account your progress is saved.

What applies to your main image when you work with AI?
For the main image Amazon requires the actual product being sold, without graphics or placeholders. AI may cut out and clean up, the product itself stays as photographed.
In an AI scene, the 19.5 cm grinder looks about 26 cm tall next to the pan. What is the main risk?
A wrong scale creates a wrong expectation. That costs returns with the reason “smaller than expected” and poor reviews. A label changes nothing about it.
For which images does Amazon require the keyword contains-synthetic-performer?
The rule applies only to photorealistic, fully AI-generated people in images, videos and A+. Real people, even AI-edited, and images without people are excluded.
Since when does the deepfake notice duty from Article 50 apply, and what about older images?
Article 50 applies since 2 August 2026. The transition period until 2 December 2026 only concerns providers' machine-readable marking. Content created before has no retroactive duty.
What does the IPTC value trainedAlgorithmicMedia in an image's metadata mean?
trainedAlgorithmicMedia stands for fully AI-generated content. Camera photos carry digitalCapture, AI-edited photos compositeWithTrainedAlgorithmicMedia.
A competitor copies your fully AI-generated kitchen scene. Where do you stand?
Purely AI-generated images are usually not protected for lack of human creation (Munich local court, 13 Feb 2026). Real photos, by contrast, enjoy Lichtbild protection, a reason to work with them.

Frequently asked

Do I have to label AI images on Amazon?

Not every one. Amazon requires the keyword contains-synthetic-performer only for photorealistic, fully AI-generated people. Under the EU AI Act you need a visible notice since 2 August 2026 if an image counts as a deepfake, meaning it shows something existing deceptively real. Cut-out real product photos and recognizably designed graphics need none.

How can I tell whether an image contains AI metadata?

With a metadata tool or a Content Credentials (C2PA) checking service. Look for the IPTC field Digital Source Type, such as trainedAlgorithmicMedia, and for a signed provenance record. If both are missing, that proves nothing: many exports strip metadata.

Can I show AI-generated people in product images?

Yes, with two duties: Amazon's keyword contains-synthetic-performer in the dc:subject field and, if the person could appear real, a visible notice under Article 50 of the EU AI Act. Off limits are faces resembling real people and scenes that attribute abilities to the product it doesn't have.

← Previous lessonAI for listing copy: facts first, then text Next lesson →AI for analysis and customer service: reviews, returns, reports
Enes Kurt
Founder of Listimo

Everything in this academy comes from more than ten years of hands-on work with Amazon listings, the same playbook behind Listimo: the tool that finds what holds an Amazon listing back and fixes it.