NORDIS tested what kind of information Google’s AI overview provides for AI-generated images submitted through reverse image search. The results were concerning.
Reverse image search is one of the most important – if not the single most important – fact-checking tool. Fact-checkers, journalists, and ordinary citizens alike use it to investigate and verify images and videos circulating on social media. According to Google, people do over 100 billion visual searches a year with Google’s tools globally.
However, Google’s reverse image search now has a serious problem: according to tests conducted by NORDIS, the AI overview added to Google Search almost systematically provides incorrect information about images uploaded for reverse image search. Instead of helping with fact-checking, it obscures and complicates it.
AI overview also provides different information about the exact same image in nearly every test run, even if only a few seconds have passed between tests.
The Finnish fact-checking organisation Faktabaari, which is part of the NORDIS team, tested reverse image search and its AI overview using nine images that international fact-checkers had confirmed to be fake. In addition, one AI-generated image created by Faktabaari itself was included. The AI overview achieved a clean record in only one out of ten tests in Finnish.
NORDIS then repeated the test in Swedish. This time, Google’s AI overview performed slightly better: it provided correct information in 4 out of 10 tests. On one occasion, it suggested that the image was “likely” generated by AI, but still linked it to real news events from elsewhere.
According to a recent analysis commissioned by The New York Times, AI overviews for text-based Google searches are somewhat more reliable. More information about this analysis can be found at the end of this article.
The AI overview was almost always wrong
NORDIS submitted ten AI-generated images to Google’s reverse image search. These images had previously been checked and identified as fakes by fact-checkers belonging to the international IFCN fact-checking network. The tenth tested image was an AI-generated image from Faktabaari’s own AI-to game, in which players are asked to assess whether images are real or AI-generated.
The original tests were done using Google in the Finnish language. NORDIS then repeated the test changing Google to Swedish language.
The test results were worrying in terms of Google’s reliability and especially the usefulness of reverse image search.
1. Prime Minister of India on a coconut plantation
Google’s AI overview suggests that an AI-generated image of India’s Prime Minister
Narendra Modi on a coconut plantation is authentic. Based on several fact checks, the
image is AI-generated. (See, for example, fact checks by Newsmobile and Vishvasnews.)
2. Alleged seizure of Ukrainian cash transport vehicles in Budapest
Google’s AI overview claims that the image below shows a real situation in which Hungarian authorities stopped Ukrainian cash transport vehicles in Budapest. According to fact checks by AFP and others, the image is AI-generated.
3. Mourners in Iran after a missile strike
In the third example, an AI-generated image claimed to depict grieving people in Iran after a missile strike was submitted to reverse image search. Numerous fact checks of the image, which spread widely in early March 2026 (see, for example, Snopes), have concluded that it is AI-generated. Despite this, in both tests the AI overview claims it is a real image.
4. Alleged missile strike in Tel Aviv
The fourth example also allegedly originates from the Middle East. According to Lighthouse Journalism, an AI-generated video has been shared on social media claiming to show an Iranian missile strike on Tel Aviv, Israel. Another fact-checking service, The Quint, has also reviewed the same video and determined it to be AI-generated.
NORDIS took a screenshot from the video and submitted it to Google reverse image search. In Finnish language no AI overview was provided, so Google’s AI mode was activated for the test. Google’s AI mode claimed the image was authentic and from Iran.
In the Swedish language the AI Overview provided false information claiming that the image shows “a historical event involving a large explosion and fire” in a building.
5. Artemis spaceflight image
Google also interprets a recent AI-generated image related to the Artemis space mission as authentic,even though AFP, among others, has identified it as AI-generated. Google’s AI overview claims the image is real despite the fact that it contains Google’s own SynthID AI watermark. This suggests that Google’s AI overview is unable to verify even the company’s own watermark, which is intended to identify AI-generated images.
6. Image allegedly from Kuwait
Regarding an AI-generated image connected to Kuwait, in the Finnish language test Google’s AI mode was uncertain if the image was real or AI generated, even though several fact checks have determined the image to be AI-generated. In both tests Google’s AI overview provided no information about the image, so the test proceeded by clicking the “AI mode” button in Google Search. Finnish AI mode stated that “it is difficult to say for sure whether this image was created or modified by AI”, after which it misleadingly connected the image to “real news stories” from around the world.
Swedish AI Mode identified the image to be “likely” AI-generated but it also connected the image to other news events, like a Ukrainian drone attack on a Russian oil terminal.
7. Manipulated video from Erbil, Iraq
A video circulating from Erbil, Iraq, shown below, has been manipulated using AI, as confirmed by, among others, the Norwegian fact-checkers at Faktisk.
Google’s AI overview did not provide any information about the screenshot taken from the video, so the test proceeded to Google’s AI mode. The Finnish AI mode recognized that the image was AI-generated but still provided incorrect information, claiming it was from Venezuela, even though the scene is actually from Iraq. The Swedish AI mode also stated that the image was “likely” AI-generated. However, it mainly provided links to social media posts as sources, even though verified fact-checks of the image are available. Google does though link to fact-checks in the right column.
8. Jeffrey Epstein related image
In January 2026, the U.S. Department of Justice released more than three million documents related to sex offender Jeffrey Epstein. Since then, these materials have been searched and examined to identify which people in positions of power had links to Epstein. At the same time, the internet has been flooded with both genuine and manipulated photographs, emails, and other evidence showing well-known individuals posing with Epstein or with young girls. An AI-generated image of U.S. President Donald Trump, fact-checked by Snopes, has circulated showing two children kneeling in front of a young Trump. Such an image does not exist in the Department of Justice materials and is AI-generated, yet Google’s AI overview still claims in Finnish that the image was taken by Jeffrey Epstein.
In Swedish the AI overview identified the image of being AI generated.
9. Firefighting image checked by TjekDet
The Danish fact-checking organisation TjekDet recently checked an AI-generated image showing firefighters extinguishing a blaze. Google’s AI mode also correctly identified this image as AI-generated in both tests. This is the only image in the test for which Google’s AI overview or AI mode provided entirely correct information both in Finnish and Swedish.
10. Faktabaari’s AI-image
In Faktabaari’s AI-to game, players must determine whether an image is real or AI-generated. For the game, Faktabaari has created the image below of a castle surrounded by floodwater using AI. Although the castle depicted is not a copy of any existing building, Google’s AI overview claims it shows the “Leawood pumping station in England,” which the AI-generated structure only very distantly resembles. Google’s AI overview claims the image was taken in England, despite the fact that the image metadata displays the exact prompt used to generate the AI image. This reveals that the AI overview is unable to check image metadata.
NORDIS and Faktabaari also tested all the images in Google’s so-called AI mode, which is also available when conducting a reverse image search with Google Lens. All it requires is ignoring Google’s AI overview when it appears and clicking on AI mode instead.
The result is not flawless, but the test indicates that AI Mode is better than AI Overview at detecting AI-generated images, especially when asked directly whether the image was created with AI or not.
The AI overview should not be trusted
AI overviews have emerged as a new concern in the spread of misinformation.
Google’s AI overviews deliver tens of millions of incorrect answers every hour worldwide. This emerged in a recent analysis conducted by the AI company Oumi and commissioned by The New York Times. In its analysis, Oumi examined AI overviews attached to Google’s text-based searches.
According to the analysis, approximately nine out of ten of Google’s AI overviews are accurate. However, because Google processes more than five trillion searches a year, the absolute number of incorrect answers is staggering. Moreover, the analysis found that in more than half of the cases where the AI overview’s information was accurate, the information presented in the overview could not be directly found in the sources linked within it. The analysis was further complicated by the fact that the AI overview could provide completely different answers in two separate tests conducted only seconds apart.
The New York Times article did not address problems related to Google’s reverse image search and the associated AI overviews. Based on Faktabaari’s and NORDIS’ tests, however, AI overviews are particularly unreliable when it comes to image identification.
Google acknowledges the criticism
Faktabaari, on behalf of both the outlet itself and NORDIS, contacted Google for comment on the findings in this article.
Google acknowledged the criticism and said it plans to integrate SynthID – its detection tool — into Google Search.
“Lens is built to connect people with from the web to help them understand an image. In limited cases when issues arise – like if our features misinterpret web content or if there’s not a lot of available context about an image on the web – we use those examples to improve our systems, and we take action as appropriate under our policies.
Lens (the feature that powers reverse image search) uses a visual match retrieval to find visually similar images on the open web. The system works by matching pixels from your uploaded image to pixels in images that already exist online.
We reviewed the examples you shared with us and took action under our policies where appropriate.
AI Overviews are dynamic, and they can improve over time as more information or context is published on a topic. Our explainer has more details about how AI Overviews work. We announced SynthID verification on the Gemini App, and our plan to also expand these capabilities including on Search,” Joakim Larsson, communications specialist at Google, writes in a written reply.