When the video turns real-looking but isn’t: the Sora 2 deepfake warning
You know how one convincing video can make you stop scrolling and wonder, Wait — did that really happen? Well, according to new analysis, the answer may increasingly be “no.” Researchers at NewsGuard have revealed that the latest video-generation tool from OpenAI, Sora 2, can be prompted to create false or misleading videos 80% of the time when asked. And yes, that’s a lot more serious than it sounds.
NewsGuard, which rates the credibility of online news sources, tested twenty false claims drawn from its database of known misinformation. The team asked Sora 2 to generate videos illustrating each claim — things like a Moldovan election official destroying pro-Russian ballots, a toddler detained by U.S. immigration officers, or even a Coca-Cola spokesperson announcing the company would skip the Super Bowl because of Bad Bunny’s halftime show. The result? Sixteen of the twenty prompts succeeded, and eleven of them worked on the first try. Five of the claims originated from Russian disinformation campaigns. These weren’t crude animations or sloppy edits either. They looked frighteningly real, like news clips you might scroll past on your lunch break without blinking twice.
Here’s the thing: we’ve known about deepfakes for a while. But this feels different. Sora 2 produces videos so lifelike that even seasoned viewers struggle to tell truth from fabrication. The old giveaways — strange lighting, mismatched lips, extra fingers — are fading away. Watching one of these clips, you’d swear it was real footage shot on a professional set. That’s why experts are calling this a turning point for misinformation, not just another step in AI progress.
OpenAI isn’t denying that risk. The company’s “system card” for Sora 2 openly admits that its advanced capabilities require “consideration of new potential risks, including non-consensual use of likeness or misleading generations.” To limit harm, they’re rolling out access through invitation only, restricting uploads of real human images or videos, and embedding both visible and invisible provenance signals — visible watermarks and hidden C2PA metadata that tag each creation as Sora-made. The idea is that if someone spreads a fake, investigators can trace it back.
But that safeguard isn’t bulletproof. NewsGuard found the watermark could be removed with a free online tool in just four minutes. The edited videos showed minor blurring where the label used to be, but looked authentic enough to fool anyone who wasn’t scrutinizing every pixel. And if that’s possible with free software, imagine what motivated actors with resources could do.
That’s what experts are worried about. Scott Ellis, a creative director at Daon, called Sora 2 essentially a deepfake engine and warned that an eighty-percent success rate in generating convincing falsehoods is “a giant red flag.” Arif Mamedov, CEO of Regula Forensics, went further, saying this isn’t about hobbyists anymore — “we’re talking about industrial-scale misinformation pipelines that can be created by anyone with a prompt.” When the cost of deception drops to nearly zero, truth itself becomes the rare commodity.
And it’s not just about national security or politics. The broader danger is erosion of trust. When people can’t tell real from fake, they start doubting everything — including legitimate journalism. That’s how misinformation wins: not by convincing everyone of a lie, but by convincing everyone that nothing can be trusted. Dan Kennedy, a journalism professor at Northeastern University, wasn’t surprised by the findings. He said that fake videos are exactly what Sora 2 is built to create, and clever users will always find ways around filters meant to prevent misuse. His warning was blunt — deceptive content that once took teams of experts can now be made by anyone, in minutes, at a quality high enough that even trained eyes may not see the trick.
OpenAI argues it’s learning as it goes, describing its approach as “iterative safety.” It’s continuing to test how people use Sora 2, adding layers of moderation and refining its rules. Each generated video carries both visible and hidden markers, and OpenAI maintains internal tools to identify Sora-made clips with high accuracy. But as critics point out, provenance isn’t the same as truth. A watermark shows where a video came from, not whether what it depicts actually happened.
Other experts question whether watermarks can really stand up against increasingly sophisticated editing. Jason Crawforth, who runs a digital media authentication firm, explained that even advanced watermarks can often be detected and erased, especially as AI editing itself improves. Jason Soroko from Sectigo made a similar point: if a watermark sits in the pixels, a simple crop or resize can destroy it; if it’s in the metadata, it disappears the moment social platforms strip those tags. They argue that a sturdier solution would involve credentials that travel with the asset — digitally signed, blockchain-anchored proof of origin and edits. But even that, they note, shows where something was created, not whether it’s truthful.
Jordan Mitchell from Growth Stack Media took it further, saying the real issue is that these systems were trained on massive datasets without proper consent or content-origin tracking. He suggested blockchain-based authentication as one possible future, likening it to how NFTs provide immutable proof of ownership for digital art.
Interestingly, Sora 2 did refuse to generate four specific false claims during the test, including one alleging a vaccine-cancer link and another blaming Israel for a fabricated attack. Why it refused those particular prompts, researchers don’t know — and that inconsistency, experts say, might be its most dangerous trait. When a system sometimes says “no” but sometimes says “sure” to nearly identical requests, users learn to experiment until they find phrasing that slips through. That’s a recipe for loopholes, trial-and-error exploitation, and a growing sense that the AI’s rules are arbitrary. As Crawforth put it, inconsistency erodes trust. If people can’t predict what’s allowed, they can’t trust the system’s safeguards either.
And that’s where this all circles back — to trust. In an age where seeing no longer guarantees believing, every technological leap forces us to re-evaluate how we decide what’s real. Sora 2 isn’t evil by nature; it’s a remarkable creative tool with enormous artistic potential. But it also exposes how fragile our collective trust has become. When anyone can fabricate a moment, brand statement, or political scandal with a few well-chosen words, truth needs better armor than a watermark.
Because if we’re not careful, soon every video will start with a silent question — not what happened? but did it happen at all?