Stock agencies require generative AI images to be labeled in two places: a flag in the submission form — Adobe's "Created using generative AI tools" checkbox, Freepik's AI button, Vecteezy's AI-generated dropdown — and, on Dreamstime and 123RF, the words "AI generated" inside the description, keywords and category. Shutterstock and Getty Images do not accept contributor-submitted AI-generated content at all.
That is the compressed version, and it is enough to keep most submissions out of trouble. The longer answer matters if you publish AI-assisted work at volume, because the rules are not consistent between agencies, the penalties for getting them wrong range from a single rejection to account termination, and the file-level metadata standard that underpins all of it is something most contributors have never looked at.
Two layers of labeling: the form and the file
Almost every disagreement about AI metadata comes from conflating two separate things. The first layer is the agency's own submission form — a checkbox, a category, a required phrase in the description. This is what moderators check, and it lives in the agency's database, not in your file. The second layer is the metadata embedded in the image itself: XMP fields written by the generator, and increasingly a signed provenance manifest. Agencies are starting to read this layer automatically, but the two layers are set separately and a tool that handles one does not necessarily handle the other.
Practical consequence: if you use a third-party upload tool, check whether it can actually set the agency's AI flag, or whether it only writes IPTC into the file and leaves the checkbox for you. This is the single most common way an otherwise compliant batch gets rejected.
What each agency requires
Adobe Stock — accepts generative AI images, vectors and video. You must select the "Created using generative AI tools" checkbox in the Contributor Portal for every AI-generated asset, without exception. Photo-realistic output is submitted as Photos; everything else as Illustrations. Generative AI content cannot go into the Illustrative Editorial collection. If the image shows a person or property that is entirely fictional, tick "People and Property are fictional"; if it depicts or is based on an identifiable real person, you need a signed model release.
Adobe also draws a useful line on partial AI use. You must label the asset when a generative tool in Photoshop or Illustrator changed, augmented or added a primary subject — generating a new person, animal or object, significantly altering a human subject, or recoloring the main subject. You are not obliged to label for extending a background, removing a distracting object or copyrighted design, or recoloring the background. That distinction is worth internalising, because it is the one most contributors get wrong in both directions.
Freepik — the AI label is mandatory for any resource fully or partially developed with artificial intelligence. You can set it with the AI button in the resource editing panel, which generates the keyword automatically, or add the _ai_generated tag via CSV for bulk submissions. Freepik also rejects assets that closely resemble your own previously submitted work in composition, framing or main elements, which matters a lot for generative batches where prompt variations produce near-identical output.
Vecteezy — tick the "AI-generated" checkbox in the Add Data section, then use the dropdown to name the program you generated the content in. Vecteezy asks for more than disclosure: AI content has to be perfectly rendered and technically flawless, with no distortion, pixelation or warping, and it has to be unique.
Dreamstime — accepts AI-generated images under defined conditions. The description must clearly state that the image was generated using AI, and one of the selected categories must be "Illustrations and Clip Art/AI-generated". You must hold all rights to the output. Generated faces should only be uploaded if providing a model release is possible; without one, Dreamstime asks you not to upload them. Dreamstime has said its framework may change as AI regulation develops, so this is one to re-check periodically.
123RF — requires "AI Generated" in both the description and the keywords, and the "AI Generated Images" category only. This is a case where the disclosure is genuinely part of your metadata rather than a form checkbox, so it has to survive whatever pipeline you use to write titles and keywords.
Shutterstock — does not allow AI-generated content to be submitted by contributors for licensing. Shutterstock's stated reasoning is consistent and worth understanding rather than working around: it requires contributors to be able to prove IP ownership of submitted content, and since generative models draw on the IP of many artists, ownership cannot be assigned to an individual. Shutterstock operates its own AI image generator and a Contributor Fund separately from contributor submissions — those are not a route for uploading your own AI output.
Getty Images and iStock — prohibit AI-generated content, citing unresolved copyright questions around model outputs, training imagery, image metadata and the individuals depicted. The ban is broad: Getty's position is that contributors must own the copyright for everything in a file, including AI-generated artwork appearing as a prop in the background. 3D renders and conventional digital editing are unaffected. Getty has said it relies on user reporting and on filters developed with the Coalition for Content Provenance and Authenticity, which brings us to the file layer.
The field underneath it all: IPTC DigitalSourceType
IPTC — the body that maintains the photo metadata standard the whole stock industry uses — recommends that software creating images with trained AI algorithms writes a Digital Source Type value of trainedAlgorithmicMedia into the XMP packet. The property is Iptc4xmpExt:DigitalSourceType and the full controlled-vocabulary URI is http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia. C2PA picked up the same vocabulary, using digitalSourceType inside its Actions and in its IPTC photo metadata assertion, so a C2PA manifest and an XMP field say the same thing in two formats.
Several major generators write this automatically — OpenAI sets DigitalSourceType to trainedAlgorithmicMedia on its output, for instance. The practical implication for contributors is uncomfortable but simple: the file may already declare how it was made, whatever you type in the form. Editing or export steps can strip XMP, which is not the same as clearing your obligation to disclose. If you are tempted to strip the field to get past a ban, understand that you are choosing to misrepresent a submission to an agency that requires proof of IP ownership, and that agencies are actively building detection around exactly this.
The reverse case is more common and easier to fix: your generator wrote the field, your editor stripped it, and now nothing in the file indicates AI use even though you have honestly ticked the agency's checkbox. Writing DigitalSourceType back in during your metadata pass costs nothing and makes the file self-describing for every downstream system.
Prompt, title and keyword restrictions
Disclosure is only half the compliance problem. Adobe explicitly prohibits certain content in generative AI prompts, titles and keywords: names of artists, real people or fictional characters; references to creative works still in copyright; names of government agencies; references to third-party intellectual property; and descriptions implying the content depicts an actual newsworthy event. Adobe states that violations may result in content removal or account termination — a materially heavier penalty than a rejection.
This is the rule that automated keywording most often breaks. A vision model looking at a generated portrait in a particular painterly style will happily suggest an artist's name, because that is a genuinely accurate description of the visual reference. Accurate and permitted are different tests. If you generate keywords automatically, you need a banned-terms layer that runs after generation, not a hope that the model behaves. More on that in our guide to titles and descriptions that comply with stock moderation policies.
Tools that handle AI labeling across a batch
Agency-native portals (Adobe Contributor Portal, Freepik editing panel, Vecteezy Add Data) — Best for: anyone submitting to one agency, and the only place some flags can be set at all. Pricing: free, part of contributing. Limitation: the label applies to that agency only, and you repeat the entire process — and the judgement calls — for every additional platform.
Xpiks — Free and open-source desktop metadata editor covering photos, vectors and video, with a paid Pro tier. Best for: contributors who want direct, inspectable control over IPTC and XMP fields, including writing DigitalSourceType by hand. Limitation: it is a metadata editor, not a compliance engine — it will not know that 123RF wants the phrase in the description and Adobe wants a checkbox, and per-file attention does not scale.
StockSubmitter — Long-established desktop application built around submitting to many agencies from one interface, including agency-specific form fields. Best for: contributors whose metadata is already written and correct, and whose real problem is distribution breadth. Limitation: authoring the metadata, and deciding what counts as AI-assisted, remains your job; the tool is Windows-centric.
Wirestock — Aggregator that takes your files, generates basic metadata and distributes on your behalf, charging a commission on sales instead of a subscription. Best for: contributors who want zero upfront cost and are content to hand off distribution decisions. Limitation: the commission is permanent and applies to every sale, and you have less direct control over how each agency's AI fields get populated.
PhotoTag.ai — Web-based AI keywording with solid visual recognition and credit-based pricing. Best for: fast keyword and title generation for stills without installing anything. Limitation: it describes what it sees; it does not know your file was generated rather than photographed, and delivery plus agency flags remain separate steps.
Rastock AI — Generates metadata, embeds it as IPTC, maps fields per agency and delivers over FTP to 25+ agencies in one pass, with banned and mandatory keyword controls so a required disclosure phrase can be enforced across a whole batch and prohibited terms filtered out after generation. Best for: contributors publishing AI-assisted work to several agencies at volume. Limitation: it cannot decide for you whether a generated face is "based on" a real person, and it cannot tick a checkbox that exists only inside an agency's web interface — those remain human decisions on the agencies that require them. For a single agency and a handful of files a month, this is more infrastructure than the problem needs.
When to use which
Agency portal vs any third-party tool. Use the portal alone if you publish to one agency, or if your AI work is a small share of an otherwise photographic portfolio. The moment you are submitting the same generative batch to three or more agencies with three different disclosure conventions, hand-repeating the process is where mistakes enter.
Xpiks vs an automated pipeline. Choose Xpiks when you want to see and edit the raw XMP, when your volume is low enough for per-file attention, or when you are correcting a specific problem rather than running a workflow. Choose an automated pipeline when the same disclosure rules have to be applied identically to hundreds of files, because consistency — not creativity — is what compliance rewards.
StockSubmitter vs Rastock AI. Both solve multi-agency delivery. StockSubmitter is the better fit if your metadata is already written and vetted and you simply need it distributed widely. Rastock AI fits better if writing the metadata is the bottleneck and you want per-agency field mapping and keyword rules applied in the same pass as delivery.
Wirestock vs a subscription tool. This is a pricing model decision. A revenue share costs nothing until you earn and is lower risk on a small or uncertain portfolio; a flat subscription is fixed regardless of income and cheaper once volume is real. Model twelve and twenty-four months against your actual expected revenue before assuming the commission option is the cheaper one — our pricing page lists flat tiers if you want a number to compare against.
PhotoTag.ai vs Rastock AI. Both generate metadata with AI. The difference is scope rather than quality: PhotoTag.ai focuses on generation and leaves compliance flags, embedding and upload to you, while Rastock AI covers generation, IPTC embedding, per-agency export and FTP delivery together. If you already have a delivery process you trust, a generation-only tool is the smaller and simpler purchase.
A pre-upload checklist for generative batches
Before a batch of AI-assisted work leaves your machine, five checks catch most problems. First, decide per file whether generative tools touched a primary subject — that is the test Adobe uses, and it is a reasonable default elsewhere. Second, confirm the file still carries DigitalSourceType if it was generated, and write it back if your editor stripped it. Third, run your titles and keywords against a banned-terms list covering artist names, real people, fictional characters, brands and government agencies.
Fourth, resolve releases before submission rather than after rejection: if a generated person is based on a real individual you need a release, and if they are fictional you need to say so explicitly on the agencies that ask. Our guide to categorization and model release attachment covers how to attach these in bulk. Fifth, route the batch only to agencies that accept AI content — sending generative work to Getty or Shutterstock is not a borderline case, and repeated policy violations carry account-level risk rather than per-file risk.
The honest summary
There is no tool that makes generative AI submissions compliant on your behalf, because the hardest parts of these rules are judgement calls — was a primary subject changed, is this face based on someone real, does this style reference cross into an artist's name. What tooling genuinely solves is consistency: applying the same disclosure phrase, the same category, the same banned-terms filter and the same embedded fields across every file in a batch, to every agency, without drift. That is a smaller claim than "automated compliance," and it is the one worth paying for.
If you are building this into a routine, start with the fundamentals of writing metadata for stock photos, then layer the AI-specific requirements on top. The disclosure rules change; the underlying discipline does not.
Related reading
Getty Images and iStock Contributor Guide 2026: Rates, AI Rules and Metadata
Shutterstock Categories and Metadata: How to Map Files Correctly
Adobe Stock Rejection Reasons (And How to Fix Every One)
Frequently asked questions
Do I need to label an image as AI if I only used generative fill on the background?
Under Adobe's guidance, no. Labeling is required when a generative tool changed, augmented or added a primary subject — a new person, animal or object, significant changes to a human subject, or recoloring the main subject. Extending a background, removing a distracting object or copyrighted design, and recoloring the background do not trigger the requirement. Other agencies may apply stricter tests, so check before assuming.
Which stock agencies accept AI-generated images?
Adobe Stock, Freepik, Vecteezy, Dreamstime and 123RF accept AI-generated content with mandatory disclosure, each using a different mechanism — a checkbox, a category, or a required phrase in the description and keywords. Shutterstock and Getty Images, including iStock, do not accept contributor-submitted AI-generated content, citing inability to verify IP ownership of model outputs.
What is IPTC DigitalSourceType and do I need to set it?
DigitalSourceType is an IPTC photo metadata property recording how an asset was created. For AI output, IPTC recommends the value trainedAlgorithmicMedia written into the XMP packet as Iptc4xmpExt:DigitalSourceType. Some generators set it automatically. No agency currently accepts it as a substitute for their own submission flag, but keeping it intact makes your file self-describing and costs nothing.
Can I name the AI model, like Midjourney or Firefly, in my keywords?
Vecteezy asks you to name the generating program in a dedicated dropdown, so there it is required. Elsewhere, avoid putting tool names in keywords unless the agency asks — keywords should describe the image for buyers, not the process. Adobe separately prohibits artist names, real people, fictional characters and third-party IP references in prompts, titles and keywords.
Do AI-generated people need a model release?
It depends on origin. Adobe requires a signed model release for any generative content that depicts, is based on, or is intended to portray an identifiable real person — including cases where you used a real person's photo as a prompt. If the person is entirely fictional, no release is needed but you must tick the fictional people and property checkbox. Dreamstime asks contributors not to upload generated faces at all unless a release is possible.
Will stripping AI metadata get my images accepted on agencies that ban them?
Removing XMP fields does not change your disclosure obligation, and agencies that ban AI content are building detection around provenance signals — Getty has said it works with the C2PA on filtering. Submitting AI work to an agency that prohibits it is a policy violation regardless of what the file says, and enforcement operates at account level, not per file.