Are Stock Photos AI Generated? How to Check a Photo Before Your Website Uses It
Some stock photos are AI generated even when the library's filter says otherwise, and the reliable answer sits in the file itself. How I check a photo before it goes on a website, and what each label in the file means.
Mike Lopez
Designs and builds websites and B2B software
On this page
Some stock photos are AI generated, and I have found them behind a stock library’s own no-AI filter, so I think the only reliable way to know is to read the photo’s file before it goes on your website. In September two of the photos which I had shortlisted for covers on this Journal, both found with that filter switched on and both listed as photographs, turned out to be AI generated when I downloaded the files, and one file also named an AI upscaler as the tool which had made it.
Both looked like ordinary photographs in the preview, a hallway of pillars with light falling between them and three white cylinders, and that is really the point of this post, because the answer sits in a few lines of text which travel inside the file while the picture itself looks like any other photo.
Why a photo’s source matters on a website
A visitor reads a face on your website as a real person, so on a clinic’s website a face which never existed is a promise that nobody can keep, right beside the booking path which I wrote about for dental clinic websites. I use photographs for people on the sites which I design, and every face on the clinic website in my Softglow case study is a photograph which I checked before it went into the design, which I would ask of any stock photo that shows a person.
The label which the file carries
The photo industry’s metadata standard, IPTC, has a field called Digital Source Type, and an image made by a generative model is meant to carry the value trainedAlgorithmicMedia, which the IPTC vocabulary defines as media created by an AI model trained on captured content. A photograph from a camera carries digitalCapture instead, and there are values in between for a real photo which an algorithm has cleaned up or a generative model has extended.
| Digital Source Type in the file | What it means | Would I use it for a face? |
|---|---|---|
digitalCapture |
Captured from real life with a digital camera | Yes |
algorithmicallyEnhanced |
A real photo, sharpened or cleaned by an algorithm with its content unchanged | Yes |
compositeWithTrainedAlgorithmicMedia |
A real photo extended or retouched by a generative model | Only after a close look |
trainedAlgorithmicMedia |
Made by a generative AI model | No |
| No value at all | The file says nothing either way | Check the camera details next |
Why a library’s filter can miss them
I think a stock library’s AI filter mostly relies on the label which the contributor chose when they uploaded the image, while the Digital Source Type can be written into the file by the tool which made it, so the file can be the more honest of the two. I suppose a contributor who has upscaled or finished an AI image in other software may well think of the result as their own photo, which is how a listing and a file come to disagree, and it’s really the reason that I read the file even when the listing looks certain.
The two covers which I found that way never went on this site, and both are now on a list of rejected images which I check before choosing any cover, so the same picture cannot come back through a different search.
How I check a photo
The check takes a couple of minutes for each photo, and I run it on every image which shows a person, a place or a product before anyone sees it in a design.
Download the full file
The preview images on stock sites are mostly resized copies with their metadata stripped, so the check only works on the file which you would actually use, and I download each shortlisted photo before I show it to anyone.
Read the source type
A free tool such as ExifTool lists every field in the file, and the line to look for is Digital Source Type. I also read the Creator Tool line, since an image generator or an AI upscaler named there tells the same story when the source type is missing.
Look for the camera
A real photograph mostly carries the make and model of the camera which took it, and often the editing software, such as Lightroom or Capture One. On a Mac those details show in the TIFF and Exif tabs of Preview’s inspector, and on Windows they are listed as camera maker and camera model on the Details tab of the file’s properties, and I think a camera and a lens in the file are the best sign that a person stood somewhere and took the picture.
Check when it was published
When a file has no camera details and no source type, I look at when the photo was first published, and I treat a photo which was online before 2022 as the safer choice, because I think AI images which pass for photographs only became common in stock libraries after that.
Where to start
I would download the three photos which matter most on your website, which are mostly the home page hero and the faces on your about or team page, and read their files with ExifTool before anything else. If a face shows trainedAlgorithmicMedia, I would replace it with a photograph whose file names its camera, and a redesign is a good moment to check every other photo too, which is why it sits in my website redesign checklist.
Frequently asked questions
Are stock photos AI generated?
Some are, since stock libraries now sell AI images beside photographs, and I have found two which came through a library’s no-AI filter. The file’s own Digital Source Type and its camera details are the reliable answer.
How can I tell if a stock photo is AI generated?
Download the full file and read its metadata with a tool such as ExifTool. The value trainedAlgorithmicMedia in the Digital Source Type field means an AI model made it, and a camera make and model mean a camera did.
Can I use AI images on my website?
You can, and for abstract shapes or textures I think it matters less, but I would keep faces, places and products as photographs, because a visitor reads those as real. The covers on this Journal are monochrome shapes, and I still check each file, so the whole set stays photographic.
Do preview images keep their metadata?
Mostly no, because previews are resized copies, which is why the check needs the downloaded file. A preview with no metadata tells you nothing either way.
