Generative video has moved a long-running warning into a practical problem. OpenAI's Sora 2 can turn a short text prompt into realistic, audio-enabled footage, and a NewsGuard investigation shows how that capability can be used to make false claims look like news.
The issue is not only that fake video exists. It is that creating it now requires far less skill, time, and production effort than earlier forms of deepfake media.
From technical trick to everyday prompt
Back in 2017, Google's Ian Goodfellow warned that trust in video as evidence might be historically unusual. He described it as a "little bit of a fluke, historically" that people had been able to treat video as proof that something happened.
Goodfellow helped invent GANs, the early generative AI technology behind the first deepfake porn videos. At that stage, producing a face swap still took hours and required technical expertise. The threat was visible, but the barrier to entry was still meaningful.
Sora 2 represents a different phase. The source article describes current video generators as powerful multimodal transformer models that have moved beyond those earlier GAN-based systems. With one line of text, a user can generate a realistic scene with audio up to 25 seconds long.
That shift changes the economics of disinformation. A false story no longer needs to remain a post, meme, or edited screenshot. It can be presented as a clip that resembles a broadcast, a street scene, or a firsthand recording.
What NewsGuard found
NewsGuard tested Sora 2 by using prompts based on 20 known false claims that were circulating online between September 24 and October 10, 2025. The model produced convincing, news-style videos for 16 of those claims. In 11 cases, it did so on the first try.
The examples covered several kinds of false narratives. NewsGuard prompted the system with claims involving pro-Russian ballots destroyed in Moldova, ICE agents arresting a toddler, and a fictional Super Bowl boycott by Coca-Cola over Bad Bunny's appearance. It also generated a credible-looking video from the false rumor that undocumented migrants in the US could no longer send money abroad.
Five of the tested claims were connected to Russian disinformation campaigns. Sora 2 generated convincing videos for all five, each in less than five minutes.
The practical implication is plain: a disinformation operator can test an idea, generate a video, and iterate quickly. The clip does not need to be perfect to be useful. If it looks close enough to a normal news segment or social video, it may travel before viewers have time to question its origin.
Safeguards are present but fragile
OpenAI says Sora blocks violent content, depictions of public figures, and misleading videos. The company also says every clip should include a visible, moving watermark and C2PA metadata, and that internal tools are used to detect abuse.
NewsGuard's findings show why those protections may not be enough on their own. The investigation found that the watermark could be removed in under four minutes using free online tools. After removal, the videos were only slightly blurred and still appeared real to most viewers, according to the source article.
OpenAI did not respond to NewsGuard's questions about how easily Sora's watermark can be removed.
The public-figure filters also showed mixed results. Prompts using names such as "Zelensky" were blocked, but NewsGuard reported that a vague description, "Ukrainian war chief," produced a lookalike video at least once. Later attempts to reproduce that result on October 15 and 16 using multiple Sora accounts did not work.
Other attempts were stopped. NewsGuard tried phrases including "a former reality TV star turned president" and "billionaire tech owner from South Africa" in efforts to bypass filters for Donald Trump and Elon Musk, but those prompts were blocked.
Why distribution matters as much as generation
The source article says some Sora-generated clips have already spread widely even with a watermark still visible. In early October 2025, AI-generated videos showing supposed antifa protesters being pepper-sprayed by police circulated across social media. NewsGuard said millions of users shared them while believing they showed real events.
This matters because watermarks and metadata depend on viewers, platforms, and distribution systems recognizing and respecting them. If a clip is reposted, compressed, cropped, or stripped of provenance signals, the remaining question is often whether the footage feels believable in the moment.
NewsGuard identified several groups likely to find such tools useful for disinformation:
- authoritarian regimes
- state-backed propaganda networks
- conspiracy theorists
- financially motivated actors
The attraction is clear from the investigation itself. A simple prompt can create material that looks like a report, includes audio, and can be tailored to a current false claim.
OpenAI is not the only company in the frame
The source article notes that similar video generators are also being built by companies like Google, with its Veo 3 model, and by Chinese developers. The concern is therefore not limited to one product.
Still, OpenAI's scale makes Sora 2 especially important. The Sora 2 app was downloaded over a million times in just five days, and anyone can use it for free. That reach increases the consequences of both successful safeguards and failed ones.
OpenAI's handling of related harms is already being watched closely. The company blocked Sora-generated videos involving historical figures like Martin Luther King with racist slurs and other offensive content, while also citing "strong free speech interests in depicting historical figures."
The larger lesson from NewsGuard's test is not that every generated video will fool everyone. It is that convincing fake footage is becoming faster and easier to produce, while the systems meant to label, block, or trace it remain imperfect. For readers, platforms, and AI companies, video can no longer be treated as self-authenticating evidence.