AI Chatbots Fail to Detect Sora-Generated Videos Without Watermarks
A NewsGuard test found that major AI chatbots including ChatGPT, Grok, and Gemini failed to reliably detect videos generated by OpenAI's Sora system, particularly when watermarks were absent or removed. The failure rate reached 95% for Grok and 92.5% for ChatGPT, raising concerns about the spread of convincing synthetic media that these systems cannot effectively identify.
What happened
NewsGuard tested three major chatbotsโGrok, ChatGPT, and Geminiโto detect videos generated by Sora, OpenAI's video generation system announced in 2024.
When videos lacked Sora's watermark, Grok failed to identify them as AI-generated in 95% of cases (38 of 40), ChatGPT in 92.5% of cases (37 of 40), and Gemini in 78% of cases (31 of 40).
Even with watermarks present, Grok failed in 30% of cases and ChatGPT in 7.5% of cases, while the watermark can be easily removed using freely available online tools.
Reported impact
The failure to detect AI-generated videos creates risk for misinformation spread, as Sora produces ultra-realistic video clips from text prompts that are difficult to distinguish from genuine footage.
OpenAI's own ChatGPT struggled to recognize outputs from Sora despite both systems being developed by the same company, revealing poor integration between the AI systems.
Index analysis
The inability of chatbots to detect AI-generated video represents a failure in content verification systems that are increasingly relied upon to identify synthetic media.
Governance controls
OpenAI applies watermarks to all Sora-generated content, but these can be removed easily with freely available tools, limiting their effectiveness as a detection mechanism.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.