What happened

NewsGuard researchers tested Sora 2 using 20 prompts based on established false narratives such as Russian disinformation and health hoaxes. 

The tool successfully generated deceptive videos for 16 out of the 20 prompts, producing hyper-realistic videos showing a toddler detained by U.S. immigration officers, a Moldovan election official destroying ballots, and fake news anchors delivering "breaking news" about non-existent corporate scandals, amongst others.

The ease of creation allows bad actors to produce high-quality propaganda and misinformation and disinformation at speed and scale and little if any cost. Furthermore, the prevalence of such videos fuels the "Liar’s Dividend," where public figures can dismiss genuine evidence of misconduct as being "AI-generated."

Why it happened

Sora 2 is a powerful generative model that will faithfully visualise whatever text prompt it is given, including well‑known false claims, unless safety systems effectively block those uses.​

NewsGuard’s prompts were explicitly built around falsehoods, yet the model still produced convincing videos for most of them, indicating that content filters and guardrails did not reliably detect and stop misinformation scenarios.​​

Researchers also noted that while OpenAI adds a floating Sora watermark and metadata, these signals can be obscured or removed using simple tools, so origin‑marking alone does not prevent misuse or guarantee transparency to viewers.

Commentators argued that limited public visibility into how safety systems are tested, tuned, and audited makes it harder to hold developers accountable for such failure modes.