By Isis Blachez, Sofia Rubinson, and Ines Chomnalez | Published on May 8, 2026

AI detection tools have been known to occasionally misidentify AI-generated images as real, but a new NewsGuard audit of leading AI detection tools found what may be a more concerning weakness: Three of the five leading tools we tested often get fooled by real images. According to NewsGuard’s findings, the tools collectively declared authentic images to be AI-generated 13.33 percent of the time, and one tool got it wrong 40 percent of the time. 

This vulnerability has potentially significant real-world consequences, empowering bad actors to dispute reality by citing a detection tool to argue that a given image is fake. It also unfairly maligns those posting the real images.

In late April and early May 2026, NewsGuard tested 15 authentic photos related to the U.S.-Iran war published by credible outlets, running them through five leading AI detection models — Hive, AI or Not, ZeroGPT, Sightengine, and ScamAI. All five tools market themselves as being able to distinguish AI-generated images from authentic ones. 

Of the five models, ScamAI — which proclaims its “Industry-leading accuracy” on its website — misidentified the most images, declaring that six out of 15 (40 percent) of the authentic images were AI-generated. ZeroGPT classified three of the 15 real images (20 percent) as fake, and AI or Not got fooled one time (6.67 percent). Hive and Sightengine correctly identified all 15 authentic images as real.

ScamAI co-founder Dennis Ng, in a May 6, 2026, video interview with NewsGuard, acknowledged that false positives can happen, saying one reason detection tools can be fooled is because they generally require high resolution and high pixel levels to detect a real image. “We can certainly adjust and fine-tune our models according to specific use cases to help reduce these occurrences,” Dennis said. (ScamAI’s website states, “Accuracy varies by media type and attack technique,” adding that its model generally achieves “95.3% detection accuracy.”)

ZeroGPT CEO Rawad Baroud told NewsGuard in a May 6, 2026, email that image-processing techniques such as resizing and compression can result in real images being incorrectly flagged as AI. He added that images “that contain visual characteristics similar to synthetic images,” including unusual lighting, high contrast, and blur, can result in inaccurate results. “News images from conflict zones can often contain several of these qualities,” he said. 

While ZeroGPT does not disclose the accuracy percentage of its image detector, the company states on its website that the tool should be used “to make informed decisions about the image’s authenticity.”

AI or Not CEO Anatoly Kvitnitsky told NewsGuard in a May 7, 2026, email that “in the case of a false positive, low image quality can sometimes affect the response.” On the company’s website, AI or Not states that it has a “98.9% AI Detection Accuracy” based on “AI or Not’s evaluation of a recent public academic data set containing real and AI-generated images.”

IRAN WAR IMAGES TRIP UP TOOLS

The 2026 Iran war has resulted in a flood of suspect visuals on social media, NewsGuard found, from fabricated scenes of missile strikes and urban destruction, to staged casualties, all designed to manipulate public perception. 

Bad actors have also weaponized AI detection tools to cast doubt on authentic visuals and to discredit legitimate reporting. One high profile example of this phenomenon was the authentic “proof-of-life” video of Israeli Prime Minister Benjamin Netanyahu at a cafe, which he posted online to counter false claims that he had been killed or maimed in an Iranian missile strike.

In response, users uploaded the clip alongside a Hive response stating that the video was “likely to be AI-Generated” with 96.9 percent probability. Anti-Israel and pro-Iran social media users claimed that the Hive result proved that Netanyahu was actually killed. In fact, comparisons of the cafe video to other images and videos taken at the same locale confirmed the video’s authenticity. It appears that a filter or a light edit was made to the video, slightly blurring its background, apparently prompting Hive to classify it as AI-generated. Hive did not respond to two emails from NewsGuard seeking comment on the matter.

UNDETECTED MANIPULATION

NewsGuard also stress-tested the models to determine whether they identified images that had been slightly altered by AI in merely cosmetic ways as AI-generated or real, as well as how they performed with images altered with AI to fundamentally change their meaning. In the first category, ScamAI was the tool most likely to identify lightly edited images as AI. In the second category, Sightengine was most likely to classify the significantly altered images as real. 

Asked about this, ScamAI’s Ng told NewsGuard: “If our AI model detects there’s some type of [AI] filter involved on that real image, we will still classify it as an AI-manipulated image.” Minor enhancements, just like more substantial AI edits, “will leave the trace of AI editing,” he explained.

Sightengine founder David Lissmyr told NewsGuard that its base model “is made to flag fully AI-generated images or heavily edited images” and that the company has more advanced models that can flag lighter edits. Sightengine markets itself on its website as having the “highest accuracy for AI-generated media detection,” citing a study by the University of Rochester and the University of Kansas.

As mentioned above, Hive and Sightengine were the only two tools that did not misidentify authentic images. At the same time, the two tools performed the worst when detecting images that had been significantly manipulated, declaring them authentic. Sightengine correctly identified just five of 15 heavily altered images, a 33 percent detection rate. Hive found that nine of those heavily altered images were fake, a 73.33 percent detection rate. ScamAI correctly identified 12 of them (80 percent), ZeroGPT was correct on 14 (93.33 percent), and AI or Not scored a perfect 100 percent.  

Hive did not respond to two emailed requests seeking comment. On Hive’s website, the company claims that “a 2024 independent research study found that our AI-generated detection model outperforms competing models as well as human expert analysis.” 

When it came to assessing slightly touched-up images, ScamAI labeled 93 percent of these images as AI-generated, followed by AI or Not at 87 percent and ZeroGPT at 80 percent. For Hive and Sightengine, the light enhancement was less likely to trick the tools, which flagged only 27 percent as AI-generated.

The variation in results — ranging from 27 percent to 93 percent for the same set of images — suggests the tools have different thresholds for what counts as AI manipulation. When detection tools label images with light touch-up edits as fully AI-generated, they risk giving consumers a false basis to dismiss authentic visual evidence.

MULTIPLE TOOLS, INCONSISTENT RESULTS

The five tools rarely agreed with each other — in 35 of the 45 images tested, at least one tool reached a different verdict than the rest. So a user turning to more than one tool for clarity would likely be left unsure about what to believe.

For example, NewsGuard tested a photograph of an Iranian missile that was manipulated with AI to inscribe the message “No Kings.” Three of the five tools classified it as AI-generated, while two labeled it as real. (NewsGuard found that pro-Iran accounts garnered millions of views on a different image of an Iranian missile manipulated with AI to contain the phrase “No Kings,” a reference to anti-Trump protests, as previously reported.)

The variability among these tools’ classifications of the same visuals points to a deeper issue: There is no industry consensus on what counts as AI-generated content, and the detection tools themselves do not state what level of manipulation will lead to an AI-generated classification.  Faced with a lack of clarity on how AI detection tools work, users can easily misinterpret results. 

NewsGuard’s findings suggest that human verification could be an important element in any effort to classify images as authentic or AI-generated. Indicators such as visual anomalies, contextual inconsistencies, and an image’s provenance — which can elude the tools — can be just as important as results from AI detection tools when assessing authenticity. 

Indeed, ZeroGPT’s Baroud said that outputs from the company’s model should be only one part of a larger verification process. “We encourage journalists, researchers, and fact-checkers to combine AI-detection results with source verification, reverse image search, metadata analysis, publication history, and contextual reporting before reaching a conclusion,” Baroud said.

METHODOLOGY

NewsGuard selected 15 images relating to the Iran war from credible sources, including news agencies Reuters and The Associated Press, daily newspapers The New York Times and The Guardian, as well as Google Earth satellite imagery. 

Each image was edited in two ways: 

• Lightly edited: NewsGuard submitted the image to an AI tool with the prompt “Enhance the lighting in this image and blur out unnecessary elements in the background.” 
• Significantly edited: NewsGuard submitted the image to an AI tool with a prompt asking it to edit the image in a way that would change the meaning of its content, inspired by false claims that circulated online during the Iran war. Such prompts included, “Edit this image to show a missile hitting the ship and it beginning to sink, realistically;” “Add smoke rising as if the nuclear facility was demolished;” “Change the flags so it looks like this image was taken in Israel;” and “Edit the writing and flag on this plane to make it look like it is operated by Iran.” 

Out of the 15 images, five were edited using OpenAI’s ChatGPT, five with Google’s Gemini, and five with xAI’s Grok. Images edited with Gemini were cropped to remove the watermark. 

NewsGuard submitted the 45 images — 15 real, 15 lightly edited, and 15 significantly edited — to the AI detection tools Hive, AI or Not, ZeroGPT, Sightengine, and ScamAI, using the free or least expensive model for each tool. 

An image that was found by a model to have a 50 percent or higher likelihood to be AI-generated was classified as an AI-generated result. Images that were found to have a 49 percent or lower likelihood of being AI-generated were classified as real.

Edited by Dina Contini and Eric Effron