India is pouring effort into AI deepfake detection ahead of future elections, but detection alone is structurally too slow: the most damaging fakes are released 24–72 hours before voting, when fact-checkers cannot keep up — which is why provenance and pre-registered identity matter more than any detector.
How good is deepfake detection technology now?
Better than it was, but still reactive. Modern detectors combine computer vision, audio forensics and signal analysis, and the strongest are multimodal — checking video, image and audio together, as India-focused surveys by firms like Paladin Tech describe. Yet detection is a probability game played after content already exists, and adversarial generators are explicitly trained to defeat the classifiers built to catch them.
Why is the timing of a deepfake the real problem?
Because the attack is designed to exploit the clock. The pattern is consistent across geographies: a fabricated clip of a candidate lands 24 to 72 hours before polling, spreads through WhatsApp forwards, and does its damage before any verification can catch up. India saw a surge of WhatsApp deepfakes during the 2024 general election, and a viral forward reaches millions long before a fact-check reaches thousands. Detection that arrives after the vote is a post-mortem, not a defence. The broader scale of the threat is captured in our India deepfake statistics.
What works better than chasing fakes after they spread?
Establishing what is real, in advance. If a public figure's authentic likeness, voice and official channels are registered and verifiable ahead of time, the question flips from "is this fake?" — slow and uncertain — to "is this the genuine, consented source?" — fast and provable. Provenance and content credentials let a platform or voter check authenticity at the point of sharing rather than days later. This proactive posture is the logic behind how the Zimorta model works: register identity first, then detect misuse against that baseline.
Does the law help close the timing gap?
Partly, and it is tightening. India's IT Rules amendments in February 2026 introduced labelling duties and faster takedown timelines for synthetically generated content, and personality-rights injunctions increasingly bind platforms directly. But a takedown order, like a fact-check, is inherently after the fact — useful for accountability, weak as prevention during the critical 48-hour window. Our guide to personality rights in India covers how these legal tools are being stitched together.
What should candidates, creators and platforms do now?
Move from detection to a layered defence built before the campaign starts. That means pre-registering authentic likeness and voice, publishing verifiable official channels, watermarking genuine content, and running continuous monitoring so a fake is flagged in hours rather than discovered after polling. Detection remains a necessary layer — but it is the last one, not the first. For talent whose identity carries commercial as well as political value, quantifying what that identity is worth is part of taking it seriously; our likeness rate calculator helps put a number on it.
The bottom line
Better detectors are welcome, but India cannot detect its way out of election deepfakes when the whole attack is engineered to beat the clock. The durable answer is to establish authenticity in advance — provenance, registration and monitoring — so the truth is already verifiable before the fake ever drops.
