Sora 2 makes convincing fake crime footage (bsky.app)

A journalist demonstrated that OpenAI’s Sora (Sora 2) can generate highly convincing fake body‑camera video in under a minute: using an interactive web interface (a JavaScript‑driven demo rather than a simple HTML form), he prompted the model to produce “body cam footage of cops arresting a dark skinned man in a department store,” and Sora returned a realistic clip quickly. The result highlights the model’s ability to synthesize complex multimodal scenes — plausible camera perspective, photorealistic people and environment — from a single natural‑language prompt. That capability matters because it lowers the technical and time barriers for producing realistic deepfake evidence, amplifying risks to individual reputations, legal processes, and public trust in police footage. Technically, the incident underscores rapid generation speed, interactive web deployment, and powerful scene/rendering priors in modern video models, while revealing gaps in content safety and provenance controls. For the AI/ML community this is a clear call to accelerate robust watermarking, metadata provenance, detection tools, stricter prompt/content filtering, and policy guardrails — especially for racially sensitive or law‑enforcement contexts where fabricated visual evidence could cause outsized harm.