Deepfakes of Prime Minister Narendra Modi, Minister of External Affairs S. Jaishankar, and Home Minister Amit Shah apologising and admitting defeat to Pakistan are viral amidst the military conflict between India and Pakistan.
BOOM found overwhelming evidence of artificial intelligence-aided manipulation in the videos, the same further confirmed through various deepfake detection tools.
The videos have gone viral following reports of retaliatory strikes from Pakistan for Operation Sindoor, and sustained skirmishes at the border between India and Pakistan. Soon after, both the countries came to a ceasefire agreement from 5 p.m. on May 10, with allegations of violations by Pakistan, later denied by the country.
Currently, both India and Pakistan have independently declared a victory over each other.
The videos show digital impersonations of Modi, Jaishankar and Shah admitting defeat and apologising to Pakistan for conducting Operation Sindoor and is viral with a claim in Urdu that roughly translates to, "The effects of Operation Binyan Marsoos, Modi's screams have erupted. Should we accept Modi's apology!!!!!"
Click here, here and here to view archives of the above post.
A quick search with the Urdu caption suggested that the deepfake videos were being shared by many other users as real statements by the leaders.
All Three Videos Are AI-Generated
BOOM found that all three videos are deepfake, with no verified news reports of any Indian leader including Prime Minister Narendra Modi having issued an apology or admitting defeat to Pakistan.
We then analysed the three videos and found overwhelming indications of AI-generated artifacts in the vides. The voices were sharper than the usual voices of these leaders, and the diction, and speaking styles were highly unusual and mechanical. We also noticed unnatural expressive movements, and badly synchronised lip movement with the voice.
Using this, we ran several segments from all three videos through Hive Moderator's AI Video detector, which provided very high possibility of the videos being AI-generated.
We also extracted the audio from all three videos, and ran them through various voice clone and AI audio detection tools on Deepfake-o-meter, an open source detection tool by University at Buffalo's Media Forensics Lab. Most of the AI audio detection models provided significant likelihood of the audio files being AI-generated.