How do you filter AI art out of Google search results?
Google has no button that hides AI images — as of 2026-07 the switch does not exist. What works: add -ai -aiart -"ai generated" to your query, restrict to trusted sites with site: or Google's Usage Rights filter, and check individual images with 'About this image.' Partial fixes only; none of them are a real filter.
Why — the first-principles explanation
The reason there's no toggle is not laziness — it's that the toggle would require Google to know which images are AI, and mostly it can't. Google's actual mechanism for this is C2PA Content Credentials: a cryptographically signed record of how an image was made, attached to the file. Google joined the C2PA steering committee on September 17, 2024 and wired it into 'About this image' across Google Images, Lens, and Circle to Search.
But Content Credentials only appear if the tool that made the image attached them and the manifest is signed by a certificate on the C2PA Trust List and nobody stripped the metadata along the way. That last condition kills it in practice. Screenshot an image, re-upload it to a social platform, or run it through most editors, and the credentials vanish — while the picture looks identical. The signal is opt-in, fragile, and trivially removable, which means a filter built on it would silently miss most of what you're trying to exclude. Google itself says there's no silver bullet.
Google's other tool, SynthID, is a watermark baked into the pixels rather than the metadata, and it survives cropping, filters, and compression. But it only marks content from Google's own AI products. It says nothing about images from Midjourney, Stable Diffusion, or anyone else — and about 80% of AI images historically came from Stable Diffusion, which anyone can run offline with no watermark at all.
So the honest strategy inverts the problem. Stop trying to exclude AI and start restricting to sources that predate it or verify their images. Filtering by date is the bluntest reliable tool available: before 2022, text-to-image barely existed at consumer scale. Site restriction to museums, archives, and stock libraries with human-only policies does more than any keyword trick. Negative keywords help at the margins, because AI content is often labeled AI by the people posting it — but that only catches the honest ones.
An example that makes it click
Imagine asking a librarian to pull every book that wasn't ghostwritten. She can only check the ones where the ghostwriter signed a confession in the front cover. Honest publishers include the note. Everyone else doesn't — and photocopying any book erases the note anyway.
That's Content Credentials. It works perfectly on the images that were going to tell you anyway, and not at all on the ones you're worried about. Which is why the practical move isn't asking for "no ghostwriters" — it's saying "only show me books from before 2022, or from these four publishers I trust." You lose some good stuff. You also stop getting fooled.
How to do it
- Add negative keywords to your query: -ai -aiart -"ai generated" -midjourney -"stable diffusion". This catches self-labeled AI content and nothing else, but it's free and instant.
- Restrict by date. In Google Images, use Tools → Time → Custom range and set the end date to 2021. Consumer text-to-image barely existed before 2022, so this is the bluntest reliable filter you have.
- Restrict by site with the site: operator — site:metmuseum.org, site:loc.gov, site:si.edu — or search museum, archive, and library collections directly instead of the open web.
- Use Tools → Usage Rights → Creative Commons licenses, which skews results toward Wikimedia, Flickr, and institutional collections that have human-provenance norms.
- Check any specific image with 'About this image': click the three dots on the image in Google Images, or use Lens or Circle to Search. It shows where the image first appeared, where else it's been seen, and any C2PA Content Credentials.
- Cross-check with a reverse image search. An image with no history before 2023 and no original publisher is a strong hint, though not proof.
- For research or licensing work, use stock libraries and archives that publish an explicit human-only or AI-labeled policy rather than filtering the open web at all.
Key facts
- Google Search offers no setting to exclude AI-generated images from results as of 2026-07; the available tools are per-image inspection and query restriction.
- Google joined the C2PA as a steering committee member on September 17, 2024 and integrated Content Credentials (C2PA version 2.1) into 'About this image' across Google Images, Lens, and Circle to Search.
- C2PA labels appear only when the manifest is signed by a certificate on the C2PA Trust List — and the metadata is removed by screenshots, many re-uploads, and most editors.
- 'About this image' launched in the U.S. in English in May 2023 and shows when an image was first indexed, where it may have first appeared, and where else it has been seen online.
- SynthID watermarking covers only Google's own AI products, while an estimated 80% of AI images through 2023 came from open-source Stable Diffusion, which can be run offline with no watermark.
- Google states there is 'no silver bullet solution for all content online' for provenance.
▶ The 60-second explainer (script)
How do you filter AI art out of Google? Straight answer: you can't. As of July twenty twenty-six there is no button, no setting, no toggle. And the reason is more interesting than the disappointment. To build that filter, Google would need to know which images are AI — and mostly it can't. Its actual tool is C2PA Content Credentials: a signed record of how an image was made, baked into the file. Google joined the C2PA steering committee in September twenty twenty-four and put it in About This Image. But credentials only show up if the tool attached them, the certificate is on the trust list, and nobody stripped the metadata. Take a screenshot, re-upload it anywhere, and the credentials evaporate while the picture looks identical. Google's other tool, SynthID, is a watermark in the actual pixels — that one survives cropping and compression. But it only marks Google's own AI. Meanwhile, roughly eighty percent of AI images historically came from Stable Diffusion, which anyone runs offline with no watermark whatsoever. So flip the problem. Stop excluding AI, start restricting to sources that predate it. In Google Images, Tools, Time, custom range, end it at twenty twenty-one — the bluntest reliable filter that exists. Use site colon for museums and archives. Add minus AI, minus Midjourney, which catches the honest people who labeled it. And check any single image with About This Image. Partial fixes. But they're what's real.
What authoritative sources say
People also ask
Does the udm=14 trick filter out AI images?
No. That parameter forces Google's classic 'Web' results and hides AI Overviews — the AI-written summary box. It does nothing about AI-generated images in Google Images.
Will Google ever add a real filter?
It would require reliable detection, which doesn't exist for unlabeled images. Google is expanding C2PA support, but any filter built on opt-in metadata misses everything that strips it.
Do browser extensions that block AI images work?
They use the same weak signals — filenames, keywords, known AI domains — plus their own guesses. They catch some obvious cases and miss the rest, with false positives on real art.
Why does the date filter work better than keywords?
Because it doesn't depend on anyone labeling anything. Consumer text-to-image barely existed before 2022, so an end date of 2021 excludes it by physics rather than by honesty.
Is Pinterest or DeviantArt filtering better than Google?
Some platforms require creators to tag AI content and let you hide it, which works on compliant users. Enforcement against people who simply don't tag remains weak everywhere.