You are about to pay for a photo of a person who may never have existed, a place no camera ever visited, or a product shot no studio ever set up. If you have ever asked, is this stock photo AI generated, before paying for it, you are asking the right question at the right moment. Millions of AI-generated images sit in the same search results as real photographs, and the labels meant to flag them run mostly on the contributor’s honesty. The question is not whether the picture looks good. It is whether you can prove where it came from before money changes hands. This guide shows you how to check a stock photo for AI origin in about six minutes, using the listing page first, the file second, and a detector last.
Short answer: yes, you can usually tell before you pay. Check the library’s AI label first, then read the file’s provenance for a C2PA manifest or an IPTC digital source type, then run a detector on the largest file you hold. A positive provenance signal settles it. A clean file never proves a photo is real.
An AI-generated stock photo is an image produced wholly or partly by a generative AI tool and then submitted to a stock library for licensing. It is not a photograph, even when it looks exactly like one. That distinction matters because a photograph carries a chain of facts a generated image cannot: a camera, a time, a place, a person who pressed the shutter, and, where people appear, a signed model release. A generated image carries none of those by default. What it carries instead depends on which library lists it and which tool made it, and that is why the check has to run in a fixed order.
Why does it matter if a stock photo is AI-generated?
Answer capsule: it matters because of ownership, releases, trust, and, since August 2026, disclosure duties in the EU.
Start with ownership. In the United States, purely AI-generated output cannot be registered for copyright. The US Copyright Office set out that position in its January 2025 report on copyrightability, and the D.C. Circuit upheld it in Thaler v. Perlmutter in 2023. The Supreme Court declined to hear the case in March 2026, which leaves the position standing. A photo nobody owns is a photo your closest competitor can license tomorrow and run next to yours. A real photograph, licensed properly, gives you the exclusivity the licence promises.
Then there are releases. A synthetic person has no model release, and if the generated face happens to resemble a real person, you have a publicity problem you never signed up for. Adobe Stock makes contributors tick a second box when the people or property in an AI image are fictional, precisely because a generated portrait cannot carry a release from someone who does not exist. The risk sharpens by sector. Healthcare, financial services, recruitment, education, and any page that implies “this is our team” or “this is our customer” are the places where a synthetic face does real damage. A decorative shot of an empty office is low risk. A “customer” holding your product is a factual claim, and a synthetic image making that claim is a problem in every market.
Trust is the third reason, and it is the one buyers feel first. Readers are getting good at sensing when a face or a scene is slightly too smooth, and once they suspect one image, they doubt the rest of the page. Disclosure is the fourth. Article 50 of the EU AI Act has applied since 2 August 2026. It requires providers of generative AI systems to mark synthetic output in a machine-readable format, and deployers who publish content that convincingly resembles real people, places or events to disclose that it was artificially generated. The machine-readable marking duty for systems already on the market before that date has a grace period ending 2 December 2026. Breaches sit in the penalty tier of up to 15 million euros or 3 percent of worldwide annual turnover. This article is general information, not legal advice, and a lawyer in your market should sign off on your policy. The practical point stands without one: if you advertise in the EU, an undisclosed synthetic photo is now a compliance question, not a taste question.
What do the big stock libraries actually allow?
Answer capsule: the big libraries split into two camps, and knowing which camp your library sits in tells you how much to trust its labels.
The split is simple to state and easy to get wrong in practice. Adobe Stock accepts AI-generated content, provided the contributor labels it. Shutterstock and Getty Images do not accept AI-generated submissions from contributors at all. Dreamstime, 123RF, Freepik and Vecteezy accept labelled AI content. Alamy, Pond5, Depositphotos, Envato and Canva do not accept contributor AI uploads. That is the position as checked on 14 November 2026, and these policies change, so check the library’s own contributor page on the day you license, not a blog post from last year.
Adobe Stock is the biggest library that takes AI work, and its rules show what honest labelling looks like. Its generative AI content guidelines, last updated 11 June 2026, require contributors to tick the “Created using generative AI tools” box for every AI-generated asset, without exception, and to tick a second “People and Property are fictional” box when the people or property shown do not exist. Photo-realistic output is filed as Photos and everything else as Illustrations. AI content cannot go into the Illustrative Editorial collection, which is reserved for authentic, unmodified depictions of real subjects. Prompts, titles and keywords must not name artists, real people, fictional characters, government agencies, or third-party intellectual property, and must not imply the image depicts a real newsworthy event. Any AI content that depicts, is based on, or is intended to portray an identifiable person still needs a model release, exactly as a photograph would. A generated person has nobody to sign one, which is why so much AI portrait work can never meet the release rule. You can read the full rules on Adobe’s generative AI content guidelines.
Shutterstock takes the opposite route. Its contributor policy states, in one line, “Shutterstock does not accept AI-generated content from our contributors.” The reasoning Shutterstock gives is that a contributor cannot prove individual ownership of the intellectual property in generated output, since the models draw on the work of many artists, and Shutterstock cannot verify which model produced a file or make sure the underlying artists were compensated. Shutterstock still sells AI generation to its own customers through its own licensed generator, and it pays existing contributors through a separate Contributor Fund for data licensing. Those are not routes for uploading your own AI output. The distinction matters for buyers: an AI-looking image on Shutterstock did not arrive through the contributor door, which narrows, but does not remove, the risk.
Getty Images and iStock take the same closed position for contributors, citing unresolved copyright questions around model outputs, training imagery, image metadata and the individuals depicted. Getty separately offers its own in-house generative AI tool, built on a model trained exclusively on Getty’s own licensed library, and sells that output with indemnification. So a Getty asset can still be synthetic. It just came from a different door, with a licence that says so.
The table below is the policy position as checked on 14 November 2026. Treat it as a starting point, not a permanent record. Libraries revise these pages without announcement, and the date stamp is part of the asset.
| Library | Contributor AI uploads | What the buyer sees |
|---|---|---|
| Adobe Stock | Accepted, labelled | Generative AI filter and a label on the listing; fictional people flagged |
| Shutterstock | Not accepted | No contributor AI in the library; own generator output sold separately |
| Getty Images / iStock | Not accepted | No contributor AI; own generator output sold with indemnification |
| Dreamstime | Accepted, labelled | “AI generated” in the title, description and category |
| 123RF, Freepik, Vecteezy | Accepted, labelled | Label or category, method varies by platform |
| Alamy, Pond5, Depositphotos, Envato, Canva | Not accepted | No contributor AI route |
One warning belongs here. A library that refuses contributor AI uploads is not a library with no AI images. Own-generator output, mislabelled uploads, and images that slipped through review all exist in closed libraries too. The policy table tells you where to look first. It never settles the question on its own.
Where is the AI label on the listing page?
Answer capsule: on the listing page itself, in the title, the description, the category, or a filter, and it takes ten seconds to read.
Open the image’s licence page and read it before you look at the pixels. That is the fastest tell in the whole check, and it is the one buyers skip most often. On Adobe Stock, AI content carries a generative AI label fed by the contributor’s checkbox, and the site offers a generative AI filter so you can see which results are flagged. On Dreamstime, the words “AI generated” appear in the title and the description, and the image sits in an AI-generated category. On free libraries the labelling is patchy and enforcement is thin, which is where brands get caught most often. If the listing page carries no label at all, that proves nothing. It only means nobody flagged it, and the flag is self-reported.
Read the contributor while you are there. Click through to the contributor’s portfolio. If a single account has uploaded several thousand assets across thirty unrelated subjects in six months, that portfolio was not shot on location. It is a supply-chain signal rather than a pixel signal, and it takes ten seconds to check. A real photographer has a subject, a place, and a style that repeats. A prompt farm has volume.
Save the licence page as a PDF alongside the file. It takes five seconds and it is the only record that exists of what the library told you at the moment you downloaded. If a claim ever lands, showing that you checked and documented it at the time is worth more than having been right. Keep the receipt next to the asset in your media library, with the date and who checked it.
What does the file itself tell you?
Answer capsule: the file can settle the question in your favour, but a clean file never proves a photo is real.
Metadata is the closest thing to a receipt an image can carry, and in 2026 the receipts are usually there. Three fields do most of the work, and you can read all three in one pass. Our full walkthrough of the method is in our guide to checking Content Credentials on a photo, and the short version is below.
Look first for a C2PA manifest, the Content Credentials standard maintained by the Coalition for Content Provenance and Authenticity. It is cryptographically signed, so it cannot be quietly edited, only removed. Then check the IPTC digital source type: a value of trainedAlgorithmicMedia means fully machine-generated, while compositedWithTrainedAlgorithmicMedia means real pixels mixed with generated ones. Finally, read the plain EXIF Software field, which often names the generator outright. The major generators now attach provenance by default, so if any of those signals is present, you are done. The image is synthetic and no further investigation is needed. You can read a manifest in seconds with the official Content Credentials Verify tool, which needs no account and does not store the image you check.
Here is the trap, and it catches almost everyone who checks stock photos. Stock libraries strip and rewrite metadata as part of delivery. What you get in the download is usually the library’s own IPTC block, with the credit line, the licensor and the asset ID, rather than the shutter speed and lens the photographer actually used. Your content management system then resizes the file and drops whatever survived. Treat metadata as a one-way signal. Its presence is close to conclusive. Its absence proves nothing at all. That is why the listing page comes first and the file comes second: the file you hold is a processed copy, and a processed copy can be silent about its own history.
Two practical rules follow. Test the largest file your licence allows, never a screenshot or a web-resized copy, because every re-compression strips signal a detector depends on. And never test the watermarked preview and call the result final. A preview is a different file from the licensed download, and its metadata is the library’s, not the image’s.
What can you still see with your own eyes?
Answer capsule: background text, reflections, repetition and impossible objects still catch lazy generations, but hands and teeth no longer do.
Stop counting fingers. Hands, teeth and glossy skin were the giveaways of 2023, and current models have largely fixed them. Advice built on those tells now produces the worst outcome available, which is confidently accusing a real photographer of faking their work because their retouching looked clean. What still holds up, roughly in order of reliability, is listed below.
- Background typography. Signage, name badges, book spines, keyboard keys, packaging copy. Text away from the focal point is still where generation breaks down first.
- Reflections and shadows. Mirrors, spectacles, shop windows, polished tables. Check whether the reflection agrees with the scene and whether every shadow points at the same light source.
- Repetition. The same face twice in a crowd, or foliage and fabric that tile like wallpaper when you look at 100 percent zoom.
- Objects that could not function. Scissors hinged in the wrong place, a cable running to nowhere, a chair with a missing leg behind a table, watch hands showing an impossible time.
- Regional details. Plug sockets, road markings, licence plates, uniform insignia, keyboard layouts. For UK, Australian and European campaigns this one catches a surprising amount, since models default to American norms.
Researchers at Northwestern’s Kellogg School grouped these giveaways into five families: anatomical implausibilities, stylistic artifacts, functional implausibilities, violations of physics, and sociocultural implausibilities. That framing is useful because it stops you fixating on faces and pushes your eye toward the edges of the frame, which is where the evidence usually lives. View the image at 100 percent and scan the edges first, not the subject.
One more visual check belongs in the stock context. Search the image itself. A reverse image search can find the same photo on another library, on the contributor’s own site, or in an earlier context that settles the question. Our main guide to telling whether an image is AI-generated covers the full method. For stock work, the useful result is usually a match on a different library with a different label, or a match that shows the image predates the generator that supposedly made it.
Do AI detectors work on stock photos?
Answer capsule: they work as a second opinion, never as a verdict, and they work worse on stock files than on original generator output.
A detector returns a probability that an image is synthetic, and even the vendors publishing the strongest benchmark results say plainly that no detector is fully reliable and that scores should be paired with human judgment. Treat any accuracy figure a detector vendor publishes as the vendor’s claim unless you tested it yourself. That matters more for stock than for most images, because libraries resize and re-encode everything they deliver, and that processing erodes the signals detectors read. Test the highest resolution version you can access, and treat borderline scores as a prompt to check provenance rather than as an answer.
Two habits improve your results more than switching tools ever will. Test the largest file you are licensed to hold, never a screenshot or a web-resized copy. And when the stakes are high, run a second tool. Agreement between two independent methods is worth far more than a single confident number. Our step-by-step guide to using the Hive AI detector shows how to run a check and how to read the score it returns.
Error Level Analysis deserves a short note because it is often sold as a detector. It resaves a JPEG at a known quality and maps where the file responds differently. Regions with a different compression history light up, which makes it very good at spotting things that were added to a real photograph: a face swapped in, a logo painted out, a sky replaced, a background extended to fit a wide crop. What it does not do is flag a clean, single-pass AI generation, because that file has one uniform compression history from edge to edge. A quiet ELA map is not a clean bill of health. For stock work, that limitation matters less than you would think. A large share of questionable stock imagery is hybrid rather than fully synthetic: a real photo with generative fill used to widen the frame, remove a trademark, or drop in an extra person. Those hybrids are exactly what ELA reads best, and exactly what pure AI detectors tend to score as real because most of the pixels are. If the image might be part real and part generated, check it against our guide to telling whether a photo has been Photoshopped as well, since the edited-real-photo case has its own tells.
Is this stock photo AI generated? Two sprout photos, checked step by step
Answer capsule: the same check, run on two real files on 14 November 2026, gave opposite answers, and the listing page decided the first one before any tool was opened.
To show the workflow working rather than describe it, I ran the full check on two sprout photos on 14 November 2026. Same subject, opposite origins. One is AI-generated and labelled as such. One is a real photograph with a full camera trail. Both files were downloaded and inspected byte by byte, then reverse searched and run through a detector.
Image A: the AI-labelled stock photo. The listing is Dreamstime image 325040627, “A single sprout growing through a cracked concrete floor”, by contributor Karasu1604. The listing page settles it before any tool is opened: a “Royalty-Free AI generated” badge sits at the top of the page, and the title itself ends with the words “AI generated”. The file I downloaded is the watermarked preview, 800 by 457 pixels and 45 kilobytes, with a Dreamstime watermark across the middle. Its EXIF block carries one thing only, the copyright line “Karasu1604 | Dreamstime.com”. No camera make, no model, no capture date. Its XMP packet carries two things only, the licensor URL and the licence web statement. There is no C2PA manifest, no JUMBF data, no IPTC digital source type, and no generator name anywhere in the file. That silence is the point of this example. The label lives on the listing page, not in the file, and a buyer who downloads the preview and reads only the metadata would find nothing at all.Reverse search found nothing. Google Lens returned 0 exact matches, and TinEye searched 86.7 billion images and found 0 matches, which is what you would expect from a generated image that has never been published anywhere else. The detector agreed. Hive scored the file 77.7 percent likely AI-generated, named Qwen as the most likely source at 61 percent, and found no Content Credentials. A zero-match reverse search does not prove an image is AI on its own, but here it agrees with the label, the file, and the detector.
Image B: the real photograph. The control is “One Little Sprout Growing”, a free photo by Matthew Henry from Shopify Burst. The file is 5,760 by 3,840 pixels and 924 kilobytes. Its EXIF block reads like a photograph because it is one: Canon EOS 5D Mark III, Adobe Photoshop Lightroom 6.10.1 on a Mac, captured 5 April 2018 at 10:15. The image itself is a studio shot, a single sprout against a white background with soil along the bottom. It carries no C2PA manifest, because the file predates the standard and Lightroom 6 never wrote one. That absence proves nothing about the scene. It only proves no credential was attached, which is exactly how a 2018 photograph behaves.Reverse search found the photo everywhere. Google Lens listed 34 exact matches, with the Shopify Burst source page fifth in the exact-match list and first in the visual view. TinEye found 22 results, had first indexed the file on 29 August 2018, and its top result was the same photo listed on Shutterstock, which shows how one real photo travels between libraries. Hive scored it 17.3 percent likely AI-generated and concluded it is not likely to contain AI-generated content. That 17.3 percent is the honest part of the test: a detector gives a probability, not a verdict, and even a real photograph does not score zero.
The comparison is the lesson. Image A is synthetic and says so on its listing page, yet its file is silent. Image B is real and says so in its metadata, yet its file carries no provenance credential. Neither file alone settles its own question. The listing page, the file, and the detector together give you a defensible answer, and any one of them alone can mislead you. Record the result next to the asset, with the date and who checked it, before you pay.
What should you do when the answer is unclear?
Answer capsule: when the check is inconclusive, treat the image as synthetic for high-risk uses and ask the library before you pay.
An inconclusive result is a result, and it has a default action. If the image will show a person, imply a result, or stand for your product, do not use it until the question is answered. Pick a different image, or ask the library in writing whether the file is AI-generated and keep the reply with the licence. For a decorative image with no person and no claim attached, an inconclusive result is a lower risk, but still write down what you checked and what came back blank.
Three rules cover the rest. First, never use an AI-generated image where the picture implies a real person, a real result, or a real product without disclosing it, and never use one on an “our team” or “our customers” page. Second, if the image is for the EU market, apply the Article 50 disclosure rule: a synthetic image that convincingly resembles a real person, place or event needs a visible disclosure from the deployer, and the machine-readable marking duty for pre-existing systems closes on 2 December 2026. Third, keep the record. Save the licence page, the metadata readout, and the detector result alongside the asset. If a claim ever lands, demonstrating that you checked at the time is worth more than having been right.
One boundary matters for buyers who also sell. If your own photos are being used without permission on stock sites, that is a theft and consent question, not an AI-origin question. Our guide to what to do when your photo is sold on stock sites without permission covers that route, including the takedown steps. This article owns AI origin only, and the two problems need different evidence.
Common mistakes buyers keep making
Answer capsule: the four mistakes are trusting a clean file, trusting one detector score, trusting a paid library, and testing a screenshot.
The first mistake is reading “no EXIF data” as proof of AI. Stock libraries rewrite metadata on delivery, so a real photograph can arrive with no camera data at all. The second is treating a detector score as a verdict. A score is a probability from one tool, and vendors’ accuracy figures are their own claims. The third is assuming a paid library means real photos. Adobe Stock accepts labelled AI content, Shutterstock and Getty sell their own generator output, and mislabelled uploads reach every library. The fourth is testing a screenshot of the image instead of the file. Every screenshot and resize strips the signal a detector and a provenance check depend on. Test the largest file your licence allows, and test it before you pay, not after the campaign is live.
A final note on timing. This article was written in November 2026 and every policy in the table was checked on 14 November 2026. Library policies, detector accuracy, and EU disclosure dates all move. If you are reading this months later, re-check the policy table against the library’s own contributor page before you rely on it. The workflow does not go out of date. The policies do.
Can you tell if a stock photo is AI just by looking at it?
Sometimes, and less often every quarter. Eyeballing still catches lazy generations through background text, broken reflections and impossible objects, but the best 2026 output passes visual inspection routinely. Use your eyes as one input, then confirm with the listing label, the file provenance and a detector before anything goes live.
Does missing EXIF data mean a stock photo is AI-generated?
No. Stock libraries strip and rewrite metadata when they deliver a file, so a real photograph can arrive with no camera data at all. Missing EXIF proves nothing. A C2PA manifest or an IPTC digital source type of trainedAlgorithmicMedia is the signal that counts, and its absence is never proof of anything.
Do Shutterstock and Getty Images sell AI-generated photos?
Neither accepts AI-generated submissions from contributors, so you will not find contributor AI uploads in their libraries. Both sell output from their own licensed generators instead, with indemnification. An AI-looking image on either library came through the company’s own generator door, not a contributor upload.
Can I use an AI-generated stock photo on my website?
Yes, if the licence allows it, but the use matters more than the licence. A decorative image is low risk. An image that implies a real person, customer or result needs disclosure, and in the EU the AI Act’s transparency duties have applied since 2 August 2026.
What does the EU AI Act say about using AI images?
Article 50 of the EU AI Act has applied since 2 August 2026. Providers must mark synthetic output in a machine-readable format, and deployers must disclose deepfakes, which are images that convincingly resemble real people, places or events. The marking grace period for older systems ends 2 December 2026. This is general information, not legal advice.
What should I do if I already licensed a photo that turns out to be AI?
First check where you used it and what it implies. If it shows a fake person as a real customer or team member, replace it and add a disclosure where the law requires one. Tell the library, keep your check record, and write the provenance check into your buying process.







