Can You Put Jewellery on an AI Model? Disclosure Rules for Sellers

Yes. Photograph the piece, composite that photo onto an AI-generated hand or neck, check the scale against measured dimensions, then disclose it twice: a caption next to each image saying which part is AI and which is the real piece, and AI metadata kept in the file. No caption cures a picture showing the wrong stone, colour or size.

Sellers tend to mix up three jobs. Making the picture accurate is a craft job. Labelling it is a platform and law job: Google's product feeds ask for machine-readable metadata, social platforms read the same kind of signal, and advertising rules apply however the picture was made. Writing the caption a shopper actually reads is a copywriting job. Each has its own check, and a seller who does only one of them is exposed on the other two.

Follow me on Instagram@sagnikteaches

Three ways to get jewellery onto an AI model

The method decides what you must say. Take a made-up online jeweller with 40 listings who wants model images for 12 pieces: a silver ring with a 6.0 mm round stone on a 2.0 mm band, a pendant on a 45 cm chain, and drop earrings.

Connect on LinkedInSagnik Bhattacharya

Method A is the safest because the product itself is a photograph, so the risk sits in the surroundings. Method B is the easiest to start with, since you upload a photo and type a prompt, and it is also the one that can produce an extra prong or a stone cut that does not exist. Method C is fine for made-to-order work, provided the caption makes clear that the buyer is seeing a render. A tool that lets you stay in method A is worth more to a seller than one that makes prettier method B images. The true-scale tests of AI jewellery model generators compare what tools do with scale in practice.

Subscribe on YouTube@sagnikteaches

Scale accuracy: a ratio check you can do with a screenshot

Scale is the complaint that follows jewellery from photoshoots to AI images. A real ring looks the size it is because a lens saw it. An AI generator guesses proportions from its training, and it can flatter: bigger stones, slimmer bands, longer chains. You need a check that uses numbers, not opinion.

The check uses two features of the piece whose real sizes you know. Here is the worked example for the made-up ring. Its stone is 6.0 mm wide and its band is 2.0 mm wide, so the stone is three times the band. Open the AI image, zoom until the ring fills the screen, and measure both in pixels with any screenshot tool. Say the band measures 38 pixels and the stone 150. That is a ratio of about 3.9, not 3.0. Working from the band, 38 pixels stands for 2.0 mm, so the stone is drawn at roughly 7.9 mm instead of 6.0 mm: about a third too big.

A second check uses the hand. The ring's inside diameter is 17.3 mm, and a finger at the base is about that wide. Real stone to finger is therefore about 6.0 to 17.3, or 0.35. If the AI finger measures 330 pixels at the ring and the stone 150, the ratio is 0.45, and the two checks agree that the stone is about 30% too large. Fingers vary, so use this as a cross-check, not a measurement. As a rule of thumb of my own, anything more than about 10% off gets regenerated or composited again.

Other pieces need their own versions of the check.

• Necklaces. Length names vary between sellers. One reference lists a princess length as 45 to 50 cm (18 to 20 inches), but another seller's "choker" may be a different figure. State the chain in centimetres and inches in the caption. Then check that the pendant sits where that length would put it, by hanging the real chain on a person or mannequin and measuring from the base of the neck.
• Earrings. Give the drop length in millimetres and check it against the ear, since generators can lengthen drops.
• Stones. Describe size in millimetres. GIA's own page defines a metric carat as 200 milligrams and says carat measures weight, not size, so a "one carat" caption does not tell a buyer how wide the stone looks.

Put one real size photograph in every listing, such as the piece on a ruler or next to a coin, and reference it from the caption of every AI image. That single photo answers most "is it really that big?" questions before they turn into returns. For colour, which AI images also bend, the tutorial on jewellery photography with AI backgrounds that stay true to life covers the checks.

Seven visual checks before an AI image goes live

Put the real photograph and the AI image side by side at the same size and go down this list.

1. Hands and ears. Count fingers, check knuckles and nails, and look at where the ring sits relative to the knuckle.
2. Metal colour. Yellow gold, rose gold and white metal can drift towards each other in generated light.
3. Stone shape and cut. Round stays round; facets do not need to match, but a cushion cut must not become an oval.
4. Setting detail. Number of prongs, bezel edges, the shape of the gallery.
5. Engraving, hallmarks and clasps. Generators can smooth these away or invent them.
6. Reflections. A stone that shows a window that is not in the room gives the game away, and an over-bright stone suggests more sparkle than the real one has.
7. Anything extra. A second ring, a bracelet or a watch the AI added; if the picture implies a set you do not sell, remove it.

Record what you find in a row per image. This one is for the ring (illustrative):

A chat assistant that accepts images can speed up the comparison, though it can miss things. Try a prompt like this and treat the answer as a list of things to look at, not a verdict.

I am attaching two images. Image 1 is a real photo of our ring. Image 2 is an AI-generated image of the same
ring on a hand. List every visible difference in: number and shape of prongs, stone shape and facet pattern,
band width and profile, engraving or marks, metal colour, and apparent size relative to the finger.
Be specific. Write "cannot tell" when unsure. Do not say the images match unless you checked each item.

A realistic answer (illustrative) says image 1 shows four claw prongs and image 2 six, the band in image 2 looks rounded and thicker, the metal is warmer, the stone looks larger against the finger, and it cannot tell whether the engraving survived. That is four useful catches and one honest "cannot tell". What the assistant cannot do is know which differences matter to a buyer. That judgement stays with you, and the zoomed-in comparison stays yours too.

Platform rules: what each surface asks of you

Rules differ by platform and change often, so this section records what the platforms' own pages said when I read them in October 2026, with the things I could not confirm marked as such.

Two practical points follow. First, the file label and the caption label are different things. The metadata tells machines. The caption tells shoppers. You need both, and the metadata is the one that can vanish when an image is resized or compressed by the wrong tool. The IPTC vocabulary that Google refers to has separate values for an image "created using generative AI", one "edited using generative AI" (inpainting or outpainting, such as a background painted around a real photo) and a "composite including generative AI elements", which describes a method A image. Google's page lists the created and composite values among those to preserve. Check which value your tool writes, then read the metadata of the exported file before upload. The tutorial on Merchant Center requirements for AI-generated product photos goes into the feed side.

Second, a labelling tool is not the same as a labelling policy. If a free generator adds its own visible mark to downloads, that mark is the vendor's, not your disclosure, and a visible stamp can conflict with a feed's overlay rules. Decide your own wording and put it where shoppers read.

Advertising rules: the picture is the claim

An advertising rulebook cares what a picture makes a shopper believe, not which tool made it. One advertising regulator's own June 2026 guidance on AI says its code is media-neutral, so the rules apply regardless of how the content was created, and that the advertiser remains responsible for any output used in an ad. In a 2026 ruling the same regulator upheld a complaint about a toy advert whose footage appeared to show real or AI-generated puppies, because the advertiser had no evidence the toy looked and behaved as shown.

Translate that to jewellery. The things a shopper relies on in a picture are size, colour, finish, stone and setting. If an AI image makes a stone look a third bigger, the advertiser needs to be able to explain why, and the explanation "the generator drew it" will not help. Typical ways AI model images mislead a jewellery buyer:

• A stone drawn larger or cleaner than the real one.
• Gold plate or vermeil shown with a colour depth that only solid gold has.
• A piece shown on a model alongside other jewellery that is not included.
• An AI person presented as a customer ("worn by [first name], a happy customer") or paired with a review that implies a real wearer.
• A made-to-order piece shown as if in stock.

The EU AI Act adds one more thing to think about if you sell to customers there. Article 50 requires anyone publishing a deep fake to disclose that the content was artificially generated. The Act defines a deep fake as AI-generated image, audio or video that resembles existing persons, objects or places and would falsely appear authentic. Whether a catalogue photo of a real ring on a person who does not exist counts is an untested question, and I am not giving legal advice on it. The cautious reading is cheap to follow: label the image.

A label does not turn an inaccurate picture into an accurate one, and the toy ruling above turned on whether the advertiser could substantiate what its pictures showed. That is the reason the scale check comes first. For the wider legal picture on AI images in marketing, the tutorial on whether you can legally use AI-generated images in marketing is the broader read.

Disclosure wording to copy, by surface

Good wording says three things: the image is AI-generated, what in it is real, and where to see the real product. Vague phrases such as "enhanced with AI" fail because they do not say which part. Here is wording for each place, all illustrative.

Placement matters as much as phrasing. Put the label next to the image, not in a footer. Label each AI image, not only the first one, since each is separately seen. Use the same sentence structure everywhere so that shoppers learn it. For a site-wide statement about AI use, the tutorial on writing an AI disclosure statement for your website gives a longer template, and the one on labelling AI-generated images and video on social media handles the platform switches.

A weak caption and a better one show the difference. Weak: "Visual for illustration only." It says nothing about AI or about what is real. Better: "AI-generated model image. The ring shown is our real product; the actual stone is 6 mm wide." The second tells the shopper what they are looking at and how big the stone really is.

Here is the whole gallery for the made-up pendant listing, so you can see where each line lands. Image 1 is the real front photo on a plain background, with no caption beyond the product name. Image 2 is the real pendant on a ruler, captioned "Actual size: 18 mm drop on a 45 cm (18 in) chain." Image 3 is the AI-generated neck with the real pendant composited on, captioned "Image 3: AI-generated model, real pendant. Chain length 45 cm; see image 2 for actual pendant size." Image 4 is the packaging, a real photo. The description's last line repeats the rule in one sentence: "Images 1, 2 and 4 are photographs of the actual item." A shopper who reads only the gallery, or only the description, gets the same message.

A 30-minute routine per piece, and what a mistake costs

Once the routine is set, one piece takes about 30 minutes (my estimate). The steps, in order:

1. Photograph the piece on a plain background, plus one size shot on a ruler.
2. Generate the AI hand, neck or ear first, with the ring finger bare and the light coming from the direction you used for the real photo. Then composite the real photo onto it (method A), using an image editor with generative fill or a chat assistant's image editing. Scale the piece with the ratio check, and add a soft contact shadow so it does not float.
3. Run the ratio check on the ring and the seven visual checks. Regenerate or recomposite on any fail.
4. Read the exported file's metadata and confirm the AI label survived; add it if not.
5. Write the caption from the wording table and the alt text.
6. Add a row to your image log: file name, method, tool and date, what is real, checks passed, caption used.

Be aware that a generator inside an editor may use a partner model, and a vendor's promise about commercial safety may cover only its own models. Check which model did the work before you rely on that promise.

The image log is worth the five minutes because it is the evidence you would need if a buyer or a platform queries a picture. A filled-in row might read: "ring-07-model-3.jpg; method A; editor with generative fill, 12 Oct; ring pixels from ring-07-front.jpg; ratio check 3.1 against 3.0 target; metadata confirmed; caption template A."

Here is a quick sum on why the checks pay. Assume a shop with 200 orders a month and an average return that costs $14 in postage and handling. If misleading scale causes 3% of orders to come back as "smaller than it looked", that is 6 returns, about $84 a month, plus the refunded sale and the buyer's goodwill. A ruler photo takes five minutes per piece. Repeated returns on pieces that have AI images are the signal to go back to real photographs for those listings, and the image log tells you which pieces those are.

Questions jewellery sellers ask about AI model images

Do I have to label an AI model image?

Several places ask for it. Google's product feeds want AI metadata in the image file, social platforms read similar signals, and advertising rules apply however a picture was made. Even where no rule names your exact case, a plain caption costs nothing and protects you if a buyer feels misled.

Is it fine if the ring in the image is my real photo?

It is safer than a redrawn ring, because the product pixels are real, but the hand or neck is still AI-generated. Say so, for example 'real ring, AI-generated hand', and check that scale, colour and finish still match the physical piece.

Can I use an AI model as my main listing image?

Keep the first image a real photograph of the actual piece, with measurements nearby, and use AI model images as extras that show how it might look worn. Some marketplaces set rules for main images, so read the category's image requirements before you upload.

Which size details should the caption give?

Give stone width and band width in millimetres, ring inside diameter, chain length in centimetres and inches, and earring drop, in a dimensions panel next to the images. Carat measures weight, not size, so carat alone does not tell a buyer how big a stone will look.

What if a customer says the piece looked bigger in the picture?

Honour the return your policy promises, then compare the image with the real piece. If the AI image overstated the size, replace it and note the change in your image log. Repeated complaints on pieces with AI images are your cue to go back to real photographs.

Further reads

• How to Check AI-Generated Images for Errors Before Posting — A general error check for AI images before you post.
• AI Tools for a Jewellery Business, Compared by the Job They Do — AI tools for a jewellery business, compared by the job they do.
• AI Model Shots for Boutiques: Try-On Images Without a Photoshoot — AI model shots for boutiques without a photoshoot.
• AI for Jewellery Design: Concept Sketches to Client-Approved Renders — A brief-to-render routine for bespoke jewellers: ten brief fields, three rounds of variants, a buildability check and a concept sheet the client approves.
• How to Make AI Ads for a Local Shop: One Ad Set From Start to Finish — One image ad and one short video ad for a local shop, built with AI from a real photo, checked against platform rules and tested on $84.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Google Merchant Center image requirements; Meta's help page and newsroom update on AI info labels for ads; the IPTC Digital Source Type vocabulary; EU AI Act Article 50; an advertising regulator's June 2026 guidance on AI and one of its 2026 rulings on a toy advert; GIA's carat weight page; Wikipedia's ring size and necklace pages. All read in October 2026. Marketplace policies change often, so read the current page before you upload.