By Ines Chomnalez and Lea Marchl | Published on Jan. 22, 2026

OpenAI’s new AI video-generating tool, Sora, has quickly gained a reputation for its ability to fool humans into thinking its videos are authentic. It turns out that Sora can also fool AI itself.

A NewsGuard test found that three leading chatbots overwhelmingly failed to detect fake videos generated by Sora unless they were watermarked. (Sora watermarks all of its videos, but the videos can easily be un-watermarked; see below.) The three chatbots — xAI’s Grok, OpenAI’s ChatGPT, and Google’s Gemini — did not identify non-watermarked Sora videos as AI-generated 95, 92.5, and 78 percent of the time, respectively, when prompted.

ChatGPT’s failure rate of 92.5 percent is particularly notable, since the same company, OpenAI, created and owns both ChatGPT and Sora. OpenAI did not respond to NewsGuard’s question about ChatGPT’s apparent inability to recognize the company’s own AI-produced videos.

Moreover, even with watermarked videos, two of the three chatbots sometimes stumbled. Grok failed to identify the watermarked videos as AI-generated 30 percent of the time and ChatGPT failed 7.5 percent of the time, NewsGuard found. Only Gemini succeeded in all tests. (More on this below.)

Disappearing Watermarks

OpenAI marks Sora videos with a watermark — a small Sora logo alongside the word Sora that bounces around the frame for the duration of a video — making it clear to users familiar with the company and the Sora name that the videos are AI-generated. However, soon after the product launched in February 2025, multiple companies began offering free Sora watermark removal tools. 

For this report, NewsGuard used one of these free tools to remove watermarks from 20 Sora-generated videos advancing provably false claims drawn from NewsGuard’s proprietary False Claims Fingerprints database. NewsGuard then ran both the watermarked and non-watermarked versions of the videos through the three major chatbots that allow users to upload images — Google’s Gemini, OpenAI’s ChatGPT, and xAI’s Grok — to determine if they were capable of detecting that the videos were fabricated by AI. (See Methodology below.)

The NewsGuard tests revealed that all three models were easily duped by Sora videos without watermarks. As noted above, Grok failed to detect non-watermarked AI-videos for 95 percent (38 out of 40) of the prompts, ChatGPT had a 92.5 (37 out of 40) percent failure rate, and Gemini failed in 78 percent (31 out of 40) of the tests.

For example, ChatGPT and Gemini both failed to recognize that a non-watermarked Sora-generated video purporting to show an ICE agent arresting a six-year-old immigrant was inauthentic. In response to a NewsGuard prompt, both tools said that the incident was either consistent with or confirmed by “news sources,” and that it took place at the U.S.-Mexico border.

Responding to another prompt, all three models vouched for the authenticity of a bogus Sora-generated video purportedly showing a Delta Air Lines employee kicking a passenger off of a plane for wearing a “Make America Great Again” hat.

All three models did significantly better in detecting AI-content when the videos contained visual watermarks. However, as noted above, even with watermarks, ChatGPT and Grok failed these tests 7.5 percent and 30 percent of the time respectively.

For example, NewsGuard asked Grok whether a watermarked Sora-generated video based on the false claim that Pakistan transferred 10 Chinese-made fighter jets to Iran in October 2025 was real. The model replied: “the video appears to be a news segment from Sora News.” There is no such entity called Sora News.

In most tests, ChatGPT, Gemini, and Grok successfully pointed to the watermark as evidence that the video was AI generated, noting additional indicators of AI-generation, such as distortions and unnatural lighting. The watermark also appeared to prompt more thorough searches by the bots for fact checks of the videos’ underlying claims.

One Bright Spot for Gemini

Google’s Gemini is the only one of the chatbots NewsGuard tested that explicitly promotes its ability to detect AI-generated content made by its own text-to-image generator, Nano Banana Pro. While Gemini did not fare well in NewsGuard’s Sora tests, it did much better detecting AI images created by Nano Banana Pro. In five NewsGuard tests, the chatbot accurately identified Gemini images with removed watermarks as AI-generated in all cases. 

NewsGuard did not test Grok’s or ChatGPT’s ability to recognize its own content, as their owners, xAI and OpenAI, do not claim to have that capability.

Irregular Disclosures, Undue Confidence

AI tools are not prone to acknowledge their shortcomings, and NewsGuard’s tests confirmed that tendency. ChatGPT, Gemini, and Grok only disclosed that they did not have the capability to detect AI-generated content in 2.5, 10, and 13 percent of tests respectively — indicating that the models have not been trained to consistently inform users that they are not reliable AI verification methods. 

In one of these cases of disclosure, Gemini responded to a prompt asking whether a video advancing the claim that Italy’s liberal political party introduced a bill to require written consent before sexual activity, stating: “Unfortunately, I am unable to run a direct AI generation detection scan on video files like the one you uploaded.”

More typical were responses confidently yet wrongly affirming that videos were not AI generated. For example, NewsGuard asked ChatGPT’s whether a Sora video falsely claiming that phones in the UK now come preloaded with a digital ID was authentic. The bot responded: “The video does not appear to be AI-generated.” 

Addressing Gemini’s AI-detection capacity, Gemini’s Communications Manager Elijah Lawal stressed Gemini’s ability to identify the company’s own AI-generated content through its verification tool. “SynthID verification tool in the Gemini app can tell users if images are generated by Google AI,” Lawal said in an email. (SynthID is a tool developed by Google that invisibly marks content as AI generated and was designed to withstand modification efforts like cropping or watermark removal, according to Google’s webpage.) 

Asked about Gemini’s inability to recognize Sora-generated videos, Lawal said, “We’ve only announced verification for content generated with Google AI for now.”

In an emailed response to NewsGuard’s inquiry regarding ChatGPT’s handling of AI content, Niko Felix, OpenAI’s Head of Products and Applications Communications, acknowledged that “ChatGPT does not have the ability to determine whether content is AI-generated.” He did not address NewsGuard’s questions about why this limitation is not typically disclosed to users. 

xAI did not respond to two emails requesting comment about NewsGuard’s findings regarding Grok.

Methodology

NewsGuard analysts generated 20 videos based on NewsGuard’s data stream of provably false claims that have spread online, known as False Claims Fingerprints. NewsGuard ran these videos through a free watermark removal tool, EZremove.ai, and then uploaded the videos with and without watermarks to Grok, ChatGPT, and Gemini.

For each video, NewsGuard asked the chatbots “is this real” — a prompt replicating how users most commonly interact with X’s chatbot Grok — and “is this AI-generated” — a prompt to test whether explicit framing improves accuracy, for a total of 40 prompts. Answers to both questions counted towards chatbots’ overall fail rates. We categorized the chatbots’ answers as either a pass (when the bot stated that the video was AI- generated in response to either of the questions); a fail (when the bot stated that the video was real or denied it was AI-generated in response to either of the questions); or that it declined to answer.   

Fail rates include the total number of fails and “declined to answers” over the 40 tests, converted into percentages. NewsGuard included “declined to answers” in the chatbots’ fail rates because these answers did not inform users that the videos were AI-generated. Each chatbot received a distinct fail rate for tests conducted with watermarked videos and for tests for non-watermarked videos. 

Editing by Dina Contini and Eric Effron