Google Images has no AI filter, no toggle, no setting, no operator that removes AI-generated pictures. You cannot switch them off, but you can stack workarounds: a date limit, the Photograph filter, negative keywords, trusted libraries and a one-minute spot-check. Used together, they cut most AI pictures from your results.
This guide shows you how to filter out AI images in Google search using the date limit, the Photograph filter, negative keywords, better sources, and a one-minute spot-check on any image before you use it. That matters because the mix has changed. Search for a puppy, a living room or a mountain lake and the first screen is now full of pictures that look like photographs but were never taken by a camera. Some are harmless decoration. Others end up in blog posts, presentations and product listings where a real photo was needed. The techniques below are the ones that actually move the needle, tested on live searches, including the ones that do not work as well as blog posts claim.
AI image generators became free and easy in late 2022, when ChatGPT opened to the public in November 2022 and tools such as Midjourney and DALL-E reached a mass audience. Since then, AI pictures have flooded the same stock sites, blogs and Pinterest boards that Google Images indexes, and Google ranks them like any other image. A picture does not need a camera, a location or a photographer to reach the top of the results, it needs a page that Google can crawl and a file that matches your words.
Google’s answer so far has been labelling, not filtering. Its “About this image” feature can tell you whether a picture was created or edited with AI tools, but only when the file carries C2PA provenance metadata, and most AI pictures on the open web do not. Google’s own blog on C2PA transparency says the feature shows the information “if an image contains C2PA metadata”, which is a big if. There is no equivalent of the safe-search toggle for AI content, and nothing in Google’s image search settings lets you hide generated pictures. Every workaround below works around that gap.
One more thing to accept before you start: no single trick is enough. The date limit misses AI pictures uploaded before 2022’s cut-off and kills genuinely new photos. The Photograph filter removes drawings, not AI. Negative keywords only catch pages that mention AI in their text. The method that works is the ladder, several weak filters stacked until the AI share of your results drops to something manageable.
What is the real-photo sourcing ladder?
The real-photo sourcing ladder is a sequence of five rungs, used in order, that moves you from the open web towards sources where a real photograph is the default. Each rung removes a different kind of AI picture, and each one costs you a little reach, that trade is the point.
| Rung | What it does | What it cannot do |
|---|---|---|
| 1. Date limit | before:2022 keeps only pages indexed before the AI flood | Kills genuinely new photos; misses AI pictures re-uploaded to old pages |
| 2. Type filter | Type menu (photo-only view via tbs=itp:photo) drops clipart, drawings and GIFs | AI photorealism is classified as a photograph |
| 3. Keyword surgery | Negative terms (-ai, -prompt) and positive terms (photograph, site:) reshape the query | Only catches pages that mention AI in text |
| 4. Trusted libraries | Search inside Unsplash, Pexels, Pixabay or Getty instead of the open web | Each library has its own AI policy, check the table below |
| 5. Spot-check | “About this image” plus a reverse search on the one picture you want | No metadata means no verdict |
Use the ladder as a routine, not a menu. Rungs 1โ3 take seconds and belong in the search box. Rung 4 is where you go when the open web keeps disappointing you. Rung 5 is the check you run on the single image you are about to download, publish or pay for. The rest of this guide works through each rung in turn, with the limits stated plainly.
How does the before:2022 date limit work, and where does it fail?
Google supports a before: operator that restricts results to pages indexed before a date you choose. Add before:2022 to any image search and Google drops everything it first saw from 2022 onwards, which is, roughly, everything from the AI-image era. For a query such as golden retriever puppy before:2022, the results are drawn from the pre-AI web: older blog posts, old Flickr pages, old news photos.
It works because the AI flood has a start date. ChatGPT opened to the public in November 2022, and the wave of AI pictures in search results followed it. Anything Google indexed before 2022 cannot be a Midjourney or DALL-E picture, because those tools did not exist in their public form. That makes the date limit the single strongest rung on the ladder for subjects that do not change, a breed of dog, a mountain, a historic building, a classic car.
But the limits are real, and blogs that present before:2022 as a cure-all are overselling it:
- The date is the page’s date, not the photo’s date. Google dates a result by when it indexed the page, not when the shutter fired. An AI picture uploaded to a page Google first crawled in 2019 sails straight through.
- AI existed before 2022. GAN-generated faces and early text-to-image research predate the public tools. The flood started in late 2022; it did not start from zero.
- It kills new photos. Searching for a 2025 car model, a new phone or this year’s festival with
before:2022returns nothing useful at all. - It narrows the pool hard. For niche subjects you may get a handful of old, low-resolution results instead of a page of AI pictures. That is the trade, and sometimes it is worth it.
Practical rule: use before:2022 for timeless subjects, and use a nearer cut-off, before:2024, for example, when you need recent photos but still want to cut out the worst of the flood. You can combine it with the other rungs: golden retriever puppy photograph -ai -prompt before:2022 is a legitimate query, and it is the shape of search this guide recommends.
What does the Type โ Photograph filter actually exclude?
Open Google Images, run a search and click Tools under the search bar, and you will see a Type menu. In the desktop interface we tested on 12 November 2026, that menu offered only “Any type”, “Clip art”, “Line drawing” and “GIF”, there was no “Photograph” entry to click at all. The photo-only view still exists, but you reach it through the address bar: add tbs=itp:photo to the search URL and Google keeps only photographs. It is the closest thing Google Images has to a “real photos only” switch, and it is not one.
What it actually excludes is style, not origin. Google’s classifier looks at what a picture looks like, not how it was made. An AI picture that mimics a photograph, a golden retriever puppy on grass, a living room at dusk, a portrait with soft window light, is classified as a photograph and stays in the results. The filter is honest about drawings and useless about photorealistic AI, which is exactly the kind of AI picture that causes the problem.
That does not make it worthless. For many searches, a large share of the AI results are illustrations, digital art and “concept” images rather than photorealism, and the photo-only view clears those at a stroke. It is also the fastest rung on the ladder once you know the parameter, and it stacks with the date limit and negative keywords. Use it every time you want a photograph rather than an image in general, just do not mistake it for an AI filter, because it is not one.
Two related filters are worth knowing. Size โ Large quietly helps: AI pictures are often served at odd or small sizes, while real photographs from cameras tend to be large files. And Usage rights โ Creative Commons licences narrows results to pages that declare a licence, which skews towards real stock libraries and away from AI-image farms, though it is a licensing filter, not a truth filter, and AI pictures can carry Creative Commons licences too.
How to filter out AI images in Google search with keywords
Search engines match words, so the words around your subject change what comes back. Two kinds of surgery help: negative keywords that push AI pages out, and positive words that pull real photography in.
Negative keywords: subtract the AI vocabulary
Put a minus sign directly before a word and Google excludes pages containing it. A practical starter set for image searches is:
-ai -generated -midjourney -dall-e -prompt -"ai generated" -"ai art" -diffusion
A query such as landscape painting -ai -generated -midjourney -dall-e drops the pages that advertise themselves as AI art. The catch is that the minus sign works on page text, not on pixels. It removes pages that mention AI, including useful articles about AI, and it does nothing to AI pictures sitting on pages that never mention AI at all. An AI puppy photo on a blog post titled “cute puppies” has no AI vocabulary to subtract, so it stays. Treat negative keywords as a nudge, not a wall.
Be careful with two of them. -digital art and -concept art also remove genuine human-made digital art, which may be exactly what you wanted. And -ai on its own can remove pages that use “ai” as part of another word or abbreviation. Add terms one at a time and watch what disappears before you build a long blacklist.
Positive words: pull real photography towards you
Adding photograph or photography to a query biases results towards pages that describe themselves as photographs. Adding a source helps more: site:unsplash.com, site:flickr.com or site:gettyimages.com restricts results to one library, and adding a photographer’s name or a photography competition (“bird photography contest”, “National Geographic”) points Google at pages that curate human-made work. These are the oldest tricks in this guide, and they still work, because they change where Google looks, not just what it matches.
The limit, again, is honesty about what words can do. AI pictures can carry the word “photograph” in their captions, file names and alt text. A prompt-generated image uploaded with the caption “a photograph of a golden retriever puppy” matches your positive keywords perfectly. Words alone cannot prove a camera was involved, which is why the ladder ends with a spot-check, not a search term.
Where should you search instead of Google Images?
When the open web keeps serving AI pictures, the strongest move is to leave it. Stock and reference libraries each publish an AI policy, and the policies differ sharply, which is useful, because you can choose a library whose rules match what you need. The table below is built from the libraries’ own published pages, checked in November 2026. Policies change, so treat the links as the source of truth and re-check before you rely on one for paid work.
| Library | AI images in the library? | What the policy actually says |
|---|---|---|
| Unsplash | No | Its Help Centre says all content on Unsplash is created by artists, and its Unsplash+ submissions page says it cannot accept content created or edited by AI generative models. |
| Pexels | Banned, with a caveat | Section 7B of its terms bars contributors from uploading content made with generative AI, but the same terms say Pexels cannot guarantee that all content is human-produced. |
| Pixabay | Yes, labelled | Its AI guidelines accept generated work that meets its standards, as long as contributors apply the AI label before submitting. Pixabay also offers a filter to exclude AI-tagged images from search results. |
| Getty Images | No | Its AI-free imagery policy states plainly that it does not accept AI-generated visuals in its Creative or Editorial libraries, and that detection, human review and contributor requirements enforce that. |
| Adobe Stock | Yes, labelled | It accepts generative AI submissions, but every one must be labelled, “Generative AI” in the title and as a keyword, plus the generative-AI checkbox in the Contributor Portal. |
Read the table as a ladder of trust. Unsplash and Getty are the strictest: if you need a real photo and you need to be able to say so, start there. Pexels bans AI uploads but admits it cannot guarantee every file, so it is a strong second choice rather than a guarantee. Pixabay and Adobe Stock both allow AI content but require it to be labelled, which makes them usable, use Pixabay’s own AI filter to hide labelled AI images, and on Adobe Stock look for the “Generative AI” label in the title or keywords before you download.
For subjects no stock library covers, a specific person, a specific place, a specific event, use the site: operator to search inside trusted publishers instead: news agencies, museums, archives and official tourism sites. A search such as site:loc.gov golden retriever or site:nhm.ac.uk butterfly looks inside the Library of Congress or the Natural History Museum, where every picture has a catalogue record behind it. It is slower than a general search, and that is the point: the slowness is the provenance.
How do you spot-check one image before you use it?
Search tricks narrow the field. The spot-check decides the single image. It takes about a minute and it has three steps: check what Google knows about the picture, check where else it appears, and check whether it carries any provenance at all.
Step 1: Open “About this image”. In Google Images, click the three dots on a result and choose “About this image”. Google’s own support page says the panel shows when Google may have first seen a similar version of the image, other pages that use similar versions, and, in the “How the image was made or edited” section, details associated with the image itself, including whether it was created or edited with AI tools. That last part only appears when the file carries C2PA metadata. If the panel says the image was created with AI tools, you have your answer. If it says nothing about how the image was made, you have no answer yet, only an absence of evidence, which is not evidence of a real photo.
Step 2: Reverse-search it. Run the image through a reverse search and look at the earliest date and the earliest source. A photograph that first appeared on a photographer’s portfolio in 2019 is a photograph. An image whose earliest appearance is a 2025 AI-gallery page, a prompt-sharing site or a page full of similar “perfect” pictures is not. Our guide to Google reverse image search walks through the steps, and the free reverse image search tool on our homepage runs the same picture through several engines at once, so you can compare what Lens, TinEye, Bing and Yandex each find. If you want to understand the provenance metadata itself, our Content Credentials check guide explains what a valid credential proves and what it does not.
Step 3: Read the picture, not just the metadata. Before you download, look for the tells that survive even when metadata is stripped: hands with the wrong number of fingers, text that dissolves into nonsense lettering, jewellery that melts into skin, backgrounds that are too clean and too perfect, and lighting that flatters everything equally. Newer generators have fixed the worst of these, so a clean picture proves nothing, but a picture with garbled text or impossible anatomy is settled in seconds. Our guide to telling if an image is AI-generated covers the visual signs in detail, and our ChatGPT image check shows what happens when you test a real file against OpenAI’s own verification page.
The test: four searches, one puppy, zero obvious AI pictures
We ran this ladder live on 12 November 2026 with the query “golden retriever puppy”, counting obviously AI-generated images, impossible anatomy, a too-perfect plastic look, garbled text or surreal lighting, in the first 20 results of each search. The result was not what the AI-panic headlines suggest. The plain search returned 0 obviously AI pictures out of 20: the grid was all plausibly real photos, with only a couple of very clean studio shots that had no clear AI tells. Adding before:2022 also returned 0 out of 20, but the results changed completely, towards older breeder and blog pages such as the AKC and Golden Meadows, with fewer polished Instagram and Pinterest shots. The photo-only view (tbs=itp:photo) also returned 0 out of 20 and changed the order only slightly, because almost all of the top results were photographs already. Adding “photograph -ai -prompt” also returned 0 out of 20, but shifted the results towards professional pet photographers such as Penny Whistle and BARKography.
Two lessons follow. First, for a popular, well-photographed subject, the first screen of Google Images is already real photographs, the AI pictures hide further down, where an Etsy listing just outside the top 20 was titled “Golden Retriever Puppy Photo, High Resolution Ultra Realistic (Digital Download)”. Second, the workarounds did not visibly remove AI pictures from this query; they changed which real photos surfaced, older pages with the date limit, professional photographers with the keyword surgery. That is still useful, but it is a different promise from “filter out the AI”, and this guide does not pretend otherwise.
The “About this image” check told the same story. On the first result, a puppy chewing a toy on the Ollie blog, the preview said only “No image metadata available”, and the full panel said “A version of this image is at least 1 year old”, listed four pages using it, and had no “How the image was made or edited” section and nothing about AI anywhere. The panel confirms what the picture is not; it cannot confirm what it is. No metadata means no verdict, which is why the spot-check ends with your own eyes, not with Google’s panel.
The routine is deliberately boring. No single step is decisive, and that is the honest position: a real photo can have no metadata, and an AI picture can sit on a page that looks trustworthy. What the routine does is make the AI pictures pay a cost. Each step removes a slice of them, and the ones that survive all three steps are the ones worth using.
What about browser extensions and blocklists?
Extensions are the last rung, and they work differently from everything above: instead of changing your search, they change your results page. The best-known approach pairs the ad-blocker uBlock Origin with a community blocklist of AI-image sites, a regularly updated list of domains that exist mainly to host AI-generated pictures. Once installed, most AI results from those domains disappear from Google image searches. A similar extension, uBlacklist, hides results from domains you choose. And the Chrome Web Store lists Google Search AI Image Hider, which hides or highlights images from websites dedicated to hosting AI-generated content, with three modes, hidden, highlighted with a yellow border, or off, and a disclaimer in its own listing that it is not 100% accurate.
Extensions have the same structural limit as negative keywords: they can only hide what they recognise. A blocklist knows the AI-image farms; it does not know the ordinary blog, the Pinterest board or the stock site that quietly started mixing AI pictures into its library. They can also produce false positives, hiding real photographs from a domain that happens to be listed. Install them from the official store, read the reviews and the permissions they ask for, and treat them as a supplement to the ladder, never as a replacement for the spot-check on the one image you actually use.
If you take one habit from this guide, make it the ladder order: date limit, Photograph filter, keyword surgery, trusted library, spot-check. No rung is a filter on its own, and Google still offers no official AI filter, but stacked together, in that order, they are the closest thing to one that exists today.







