Hours before Maharashtra's 2024 assembly election polling, BJP-amplified audio clips purporting to catch opposition leaders Supriya Sule and Nana Patole discussing a bitcoin-funded election bribery scheme were independently forensically confirmed as substantially AI-generated voice clones.
What Happened
On November 19, 2024, the night before polling in the Maharashtra Legislative Assembly elections, former Pune IPS officer Ravindranath Patil alleged, a claim BJP spokesperson Sudhanshu Trivedi then amplified at a Delhi press conference, that NCP(SP) leader Supriya Sule and Maharashtra Congress president Nana Patole had misappropriated bitcoins seized in a 2018 Pune cyber-crime Ponzi-scheme investigation to fund their election campaigns, allegedly with the help of then-Pune police commissioner Amitabh Gupta and an audit-firm employee, Gaurav Mehta. To back the claim, BJP's official X account posted four short audio clips beginning at 10:58 pm that night, purporting to be recorded conversations: one had a voice resembling Sule telling Mehta to "cash all the bitcoins" and not worry about inquiries because "we will handle it when we come to power"; another had a voice resembling Patole pressing Gupta about delays converting the funds to cash; the remaining two involved Mehta and Gupta discussing wallet transfers. Both Sule and Patole immediately and publicly denied the voices were theirs, calling the clips "fake" and "AI-generated," and filed complaints. Independent fact-checkers moved fast: BOOM and the Misinformation Combat Alliance's Deepfakes Analysis Unit (DAU) ran the clips through multiple commercial AI-audio-detection tools and compared the voices against each leader's verified public interviews. Their report, published November 22, 2024 and updated November 23, found that three of the four clips showed strong, high-confidence evidence of AI voice generation, while the roughly five-second Patole clip was too short to test reliably. UC Berkeley digital-forensics professor Hany Farid separately reviewed the clips and likewise concluded at least three were AI-generated.
How the Attack Worked
Four short audio clips were posted to BJP's official X account starting at 10:58 pm on November 19, 2024, hours before polling. The clips were framed as authentic leaked recordings, sourced through a chain that began with former Pune IPS officer Ravindranath Patil, who said he had received the audio from Gaurav Mehta via WhatsApp/Signal. BJP spokesperson Sudhanshu Trivedi amplified the claims at a Delhi press conference on the eve of polling. Independent forensic testing found the Sule clip and the Gupta/Mehta clips carried strong AI-generation signatures consistent with voice-cloning/text-to-speech synthesis, while the ~5-second Patole clip was too short for detection tools to score reliably. No court or regulator has established who technically produced the synthetic audio.
The Lure & the Tell
Lure: the fabricated clips were dressed up as an authoritative "leak," sourced ostensibly from a former IPS officer with insider knowledge of a real 2018 cybercrime case, timed for release the night before polling, and amplified by a national party spokesperson at a press conference to borrow institutional credibility. Tell: forensic AI-audio detectors flagged unnatural pauses, robotic/flat cadence, and voice-print mismatches versus each speaker's verified public interviews; three of the four clips scored high-confidence AI-generated, while the shortest clip was simply too brief to test reliably.
Outcome
Sule filed a formal complaint with the Election Commission of India and the Cyber Crime Cell, naming Gaurav Mehta and Ravindranath Patil, and filed a criminal defamation notice against Sudhanshu Trivedi. Patole publicly denied the voice was his and announced an FIR against Trivedi and Patil. The Deepfakes Analysis Unit published a forensic report on November 22, 2024 (updated November 23) concluding the Sule clip and the Gupta/Mehta clips showed high-confidence AI-generation signatures, while the brief Patole clip remained inconclusive. No court ruling, EC decision, or police charge sheet resolving the matter has been publicly confirmed.
Why It Matters
This is one of the clearest documented cases of AI voice-cloning deployed as an election-eve political weapon in a major democracy: fabricated audio of two senior opposition politicians was pushed through an official party account and a national press conference on the single day it could do maximum damage. It demonstrates how synthetic media can be laundered through credible-seeming intermediaries to bypass normal skepticism, and how a real underlying case can be exploited as connective tissue to make fabricated content feel corroborated. It also shows both the promise and limits of AI-audio forensics: detection tools produced strong AI probability on longer clips but could not reliably adjudicate a clip only a few seconds long.
Defenses That Would Have Stopped It
- Independent, rapid third-party forensic verification: BOOM and the Deepfakes Analysis Unit ran the clips through multiple detection tools within 48 hours and published a public report, preventing the fabricated "evidence" from going unchallenged during the news cycle
- Voice-print comparison against known-authentic public interviews of the targeted individuals as a baseline for spotting mismatches
- Independent academic corroboration (UC Berkeley's Hany Farid) added credibility beyond partisan fact-checkers
- Formal complaint channels used by the targets: Election Commission of India complaint, State Cyber Crime Cell complaint/FIR route, and defamation notice, creating an official record
- Public, on-record denial and swift legal notice by the targeted politicians within hours, limiting the window for the narrative to solidify unchallenged
Attack Chain & Defense
The sequence the attacker ran
How it could have been stopped
1
Voice-sample and target profiling: Whoever produced the clips plausibly needed authentic reference audio of Supriya Sule, Nana Patole, and IPS officer Amitabh Gupta to condition a voice-cloning model on each person's real speech, likely drawn from publicly available interviews, press conferences, and broadcast appearances rather than any private source.
Countering Stage 1: Public interviews and broadcast appearances are a normal, unavoidable part of a politician's public life, so the realistic control is not hiding voice samples but preparing to use them defensively, such as fact-checking bodies keeping known-authentic reference clips ready to compare against a disputed recording, which is exactly what BOOM did against Sule, Patole, and Gupta's YouTube interviews.
2
Narrative scaffolding around a real case: The fabrication was built on top of a genuine, pre-existing 2018 Pune cryptocurrency Ponzi-scheme investigation, borrowing real names, roles, and wallet mechanics from that case so invented dialogue would feel corroborated by facts a listener could independently confirm existed.
Countering Stage 2: Because the underlying 2018 case genuinely existed, no verification step could flag the premise itself as fabricated; the practical control is media literacy and editorial discipline that treats "a real underlying case exists" and "this specific audio about that case is authentic" as two separate claims each requiring independent verification before publication.
3
Synthetic audio production: Forensic analysts (DAU/BOOM, TrueMedia, Hiya, Hive, and UC Berkeley's Hany Farid) concluded three of the four clips carried strong AI-generation signatures consistent with text-to-speech or voice-conversion synthesis; notably, the one clip kept to roughly five seconds fell below the length current detection tools could score reliably, whether or not that brevity was deliberate.
Countering Stage 3: Commercial AI-audio detection tools (Hiya, Hive, TrueMedia, Deepfake-o-Meter) plus independent academic forensic review are the direct countermeasure, and they worked for three of the four clips here; the case also exposes their known limit, since a roughly five-second clip was too short for any tool to score reliably, meaning a short enough fabricated clip can still evade this stage's defenses.
4
Credibility laundering through a human source chain: The clips were passed from an alleged original leaker (audit-firm employee Gaurav Mehta) to a controversial former IPS officer (Ravindranath Patil) and then to a national political party, so each handoff added apparent vetting rather than exposing the audio as unverified and anonymous.
Countering Stage 4: Platforms and media outlets can require attribution of original source and a documented evidence chain-of-custody before republishing "leaked" audio as fact, rather than treating a source's claimed credentials, such as being a former IPS officer, as a substitute for verification.
5
Timed release for minimum scrutiny: The clips were posted to BJP's official X account at 10:58 pm on November 19, 2024, hours before Maharashtra Assembly polling opened, a timing pattern typical of election-eve disinformation intended to outrun fact-checking before votes are cast.
Countering Stage 5: Election commissions and platforms can adopt expedited review, labeling, or takedown procedures for unverified audio-visual claims about candidates in the final 24 to 48 hours before polling, precisely when the window for rebuttal is structurally shortest.
6
Institutional amplification: BJP spokesperson Sudhanshu Trivedi presented the allegations at a national press conference in Delhi, and BJP MP Sambit Patra added further claims the next day, lending party-level authority to unverified audio and driving it into national broadcast and print coverage.
Countering Stage 6: Newsrooms and platforms can require a higher corroboration bar before amplifying a party spokesperson's unverified allegations nationally, labeling claims as unverified pending forensic review instead of repeating them as established fact.
7
Objective completion, exploiting the verification gap: The clips circulated on X, WhatsApp, and television through election-eve and polling day, the exact window in which independent forensic debunking (not published until November 22-23) did not yet exist, so the disinformation reached voters before any counter-evidence could catch up.
Countering Stage 7: Once the pre-vote window has closed, the realistic remedy shifts from prevention to accountability; the rapid independent forensic publication (November 22-23), the Election Commission and Cyber Crime Cell complaints, and the defamation notices filed here create an official record and a deterrent for future use, even though they could not undo the reach the fabricated audio achieved before votes were cast.
Social Engineering Examples