A July 2026 wave of AI deepfake videos falsely showed Union Minister Piyush Goyal and Prime Minister Narendra Modi threatening student protesters as though those were their real words; the most viral clip took genuine PTI footage of Goyal and swapped in a fabricated audio track, and the government's PIB Fact Check unit dismissed the videos as AI-generated, with an FIR registered on 25 July

Why the clips were convincing

It is worth being honest about why these videos worked, because dismissing everyone who paused over them as gullible misses the mechanism. The Goyal clip was not a cartoonish fabrication. It was built on real footage, a genuine PTI recording of a sitting minister speaking to reporters outside Parliament, and a viewer who had never seen the original had no visual tell to catch. The setting was authentic, the face was authentic, the framing said “news.” Only the sound was false.

There is also a reason clips of officials behaving badly spread so fast. Footage that seems to catch a powerful person saying the quiet part out loud feels like a rare glimpse behind the curtain, and it flatters the viewer's sense of having seen through them. Drop that into a live protest, where feelings are already high and one side is primed to believe the worst of the other, and the clip does not read as a claim to be checked. It reads as confirmation.

None of that makes the videos real. It explains their reach. A fake that hides inside a trusted format, lands in a charged moment, and tells an audience what it half expected to hear is engineered to outrun the correction, and this one did, for a day or so.

How it was debunked

The debunk did not depend on taking anyone's word. The decisive move was recovering the original footage. Because the Goyal clip was built on a real PTI recording from 23 July, fact-checkers could line the viral version up against the source and show that the inflammatory audio simply was not in it. The visuals matched; the words did not. That is about as clean as a manipulated- media debunk gets.

Officialdom reached the same place. The government's PIB Fact Check unit dismissed the Goyal and Modi videos as AI-generated deepfakes that falsely portrayed the leaders threatening students. Goyal said on X that his remarks had been maliciously doctored using AI. An FIR was registered at Chanakyapuri police station in Delhi on 25 July, and government sources said many links to the clip had already been taken down.

On the origin of the clips, PIB pointed to what it called Pakistani propaganda accounts. That is an official attribution, and this file carries it as exactly that, a claim by the government about who was spreading the videos, not an independently confirmed forensic finding. The authenticity question and the who-made-it question are separate; the first is settled, the second is an attribution.

The clincher was the original recording. Set the viral clip beside the real PTI footage and the manufactured audio has nowhere to hide.

The anatomy of an audio swap

Strip away the drama and the construction is almost mundane. Begin with authentic footage of a public figure, the more routine the better, because a familiar setting reads as trustworthy. Keep the picture. Delete the real sound. Drop in an AI-generated audio tracksaying something incendiary, timed loosely to the speaker's mouth. Add a caption that tells viewers what they are hearing. Push it out while the topic is hot.

The scripted line was not chosen at random either. It was crafted to play on the protest movement's own name and to turn a live grievance into fury, which is why this file describes the fabricated remark rather than repeating it. The point of recounting the recipe is not the content of the fake; it is how little the recipe costs. Real footage of officials is abundant, generative audio is cheap, and the turnaround is hours.

That speed is the danger. A minister spoke to reporters on 23 July; a weaponised version of that same clip was circulating by the next day, and the correction did not arrive until the day after. The technique does not need to fool everyone forever. It needs to reach enough people during the window before the debunk lands, and an audio swap on genuine footage is built precisely to win that race.

Why fakes like this travel, even after the correction

Corrections almost never move as fast or as far as the thing they correct. The viral clip does its work in the first hours, carried on emotion and instinct; the fact-check, however airtight, arrives later and reaches a smaller, cooler crowd. By the time the recovered PTI footage is circulating, the impression, that maybe a minister really said something monstrous, has already lodged with people who will never see the retraction.

The psychology does the heavy lifting. We treat video of a recognisable person as close to proof, and we extend extra weight to anything that looks like official news footage. A fake that borrows both, a real face in a real setting, inherits credibility we grant automatically. Layer on a satisfying villain and a moment of genuine anger, and the clip outpaces sober doubt.

The goal is not to be believed forever. It is to be believed long enough, by enough people, to leave a residue the truth then has to scrub out.

There is a second-order cost, too. Every convincing fake of an official erodes the baseline assumption that real video of a real person is reliable, and it hands anyone caught on camera a ready escape: dismiss the genuine article as “probably a deepfake.” The harm is not only the specific lie. It is the slow corrosion of the idea that seeing can be believing at all.