.jpg)
The Deepfakes Analysis Unit (DAU) analysed three videos, which separately feature the likeness of Prime Minister Narendra Modi, External Affairs Minister Dr. Subrahmanyam Jaishankar, and Home Minister Amit Shah. The narrative spun through these videos is that the Indian leaders are accepting defeat against Pakistan and seeking apology. After putting the videos through A.I. detection tools and getting our expert partners to weigh in, we were able to establish that the videos are deepfakes.
Each video is about one minute and 10 seconds in length. The language used in the videos is primarily Hindi with some Urdu words while the text in the body of the posts carrying the videos is Urdu. The links to the videos were escalated to the DAU by many of its fact-checking partners.
The fact-checking unit of the Press Information Bureau (PIB), which debunks misinformation related to the Indian government recently posted a fact-check, from their verified handle on X, debunking the video featuring Dr. Jaishankar’s likeness.
All the videos were posted on May 10, 2025 from different accounts; we do not know if the video uploaders were also the creators of the content. The video featuring Mr. Modi’s likeness was uploaded from an account with the display name of “Zahoor Khan” and a display picture of a man. The profile details indicate that the account belongs to a “digital creator” in Pakistan who “works” at “Facebook”.
The video featuring Mr. Shah’s likeness was posted from an account with the display name of “Awais Ahmad” and the display picture of a man. The profile details suggest that the account belongs to someone from Pakistan who works at “Student Struggle”.
The video featuring Jaishankar’s likeness was posted from an account with the display name of “Shahzad Younas Sheikh PTI” and a display picture of a man seated in a room with the Pakistani national flag and another flag in the backdrop. The profile details suggest that the account belongs to a “politician” who is an “ex spokesperson” for the government of Punjab, Pakistan.
So far, the video with Jaishankar’s likeness has received more than 31,000 views, the supposed Shah video has been viewed some 19,000 times, and the purported Modi video has about 300 views.
The caption with the video caricaturing Jaishankar roughly translates to: Jaishankar is seeking apologies and making excuses, should he be forgiven, or should we attack them until the last dhoti catches fire? The approximate translation for the caption accompanying Modi’s purported video is: The effects of Operation Bunyan Marsoos have made Modi cry out, should we accept his apology? And the text with Shah’s supposed video loosely translates to: Finally India has admitted defeat. If they had done so earlier, this situation might not have arisen. After Jaishankar and Colonel Sofia, Amit Shah has also admitted defeat. India is now apologising. Pakistan Zindabad!
Each of these videos has been packaged like a public announcement with one apparent speaker, their framing alternating between a medium shot and a medium close-up. All the purported speakers appear to be in a similar formal setting with the Indian flag in the background and a podium with at least one microphone in front, in their respective videos. Their gaze shifts in different directions as if addressing an audience.
There are multiple jump cuts in the videos, as a result, the head and body movements of the purported speakers appear to change abruptly. Their facial expressions lack natural spontaneity and seem to be at odds with the audio track in some instances. Each video has a waxy finish, commonly seen in A.I.-generated content; and lacks the naturalness typically associated with recorded footage featuring a real person. The eye movements of all the three subjects look especially unnatural.
The opening and closing of the mouth appears hastened for each subject, creating the visual effect that while their mouth is moving the rest of their face is static. The inside of their mouths, respectively, appears unnaturally dark. Their lip movements seem to align with their respective audio tracks, however, in some frames the mouth movements continue even as there is a brief pause in the audio, which creates a strange fluttering motion in their lips. Their lips also seem unnaturally bloated in many frames.
The teeth for all the three subjects look blurred and warped, they tend to disappear and re-appear within seconds. For Modi’s and Shah’s likeness the upper set of teeth is not visible at all. In Jaishankar’s case, the upper set is visible in some instances but appears to change shape. The right side of the jaw of Modi’s likeness changes shape in some frames.
Three distinct male voices are recorded over the video tracks. The voice attributed to the purported speakers bears some resemblance to their natural voice as heard in their recorded speeches and interviews. The voice with Jaishankar’s likeness does capture his mellow tone but the signature voice modulation of Modi and Shah does not come through in the videos. The diction, accent, and syntax in all the audio tracks seems odd. Overall the delivery sounds very scripted, hastened, and lacks natural pauses.
Through our assessment reports we have consistently been highlighting the patterns that we notice in videos and audios that have been manipulated or generated using A.I. We have shared below a table with the translations of the audio tracks only to inform our readers about how a certain narrative has been built using the audio. By no means do we want to give any oxygen to the bad actors behind this misleading and harmful content.
We undertook a reverse image search using screenshots from the videos. The results that returned led us to the same videos, which had been shared across different social media platforms. We did not find any evidence online that pointed to a source video for either of the videos.
To discern the extent of A.I. manipulation in the videos under review, we put them through A.I. detection tools.
The voice tool of Hiya, a company that specialises in artificial intelligence solutions for voice safety, indicated that there is a 95 percent probability that the audio track in the video featuring Shah’s likeness was generated or modified using A.I. The tool indicated a 34 percent probability of the audio with Jaishankar’s likeness having been generated or modified using A.I., and it suggested a 15 percent probability of the use of A.I. in the audio with Modi’s likeness.
Hive AI’s deepfake video detection tool indicated that the entire video track featuring Modi’s likeness is A.I.-generated. For the videos featuring the likeness of Shah and Jaishankar, respectively, the tool highlighted several markers of manipulation throughout the video tracks.
Their audio detection tool found the entire audio track with Modi’s and Shah’s likeness to be A.I.-generated. For the audio track in the video caricaturing Jaishankar, the tool only highlighted a 30-second segment as A.I.- generated.
We also ran the audio tracks from the videos through Deepfake-O-Meter, an open platform developed by Media Forensics Lab (MDFL) at UB for detection of A.I.-generated image, video, and audio. The tool provides a selection of classifiers that can be used to analyse media files.
Of the seven audio detectors that we used to analyse all the audio tracks, six indicated that it was highly likely that the audio purported to be Shah’s voice was A.I.-generated, only four indicated the same for the voice with Jaishankar’s likeness, and only two did so for Modi’s purported audio.
AASIST (2021) and RawNet2 (2021) are designed to detect audio impersonations, voice clones, replay attacks, and other forms of audio spoofs. The Linear Frequency Cepstral Coefficient (LFCC) - Light Convolutional Neural Network (LCNN) 2021 model classifies genuine versus synthetic speech to detect audio deepfakes.
RawNet3 (2023) allows for nuanced detection of synthetic audio while RawNet2-Vocoder (2025) and RawNet2-Vocoder-V1 (2023) are useful in identifying synthesised speech. Whisper (2023) is designed to analyse synthetic human voices.
For a further analysis on the audio tracks we put them through the A.I. speech classifier of ElevenLabs, a company specialising in voice A.I. research and deployment. Their classifier returned results as “very likely”, suggesting that there is a high probability that the audio tracks with the likeness of Modi, Shah, and Jaishankar, respectively, were generated using their platform.
We reached out to ElevenLabs for a comment on the analysis. They told us that based on the analysis of technical signals, they were able to confirm that the audio track with Modi’s likeness is A.I.-generated. They added that they have taken action against the individuals who misused their tools to hold them accountable.
For the audio tracks recorded over Jaishankar’s and Shah’s likeness, however, they told us that they were unable to conclusively determine that the synthetic audios originated from ElevenLabs.
They noted that both Jaishankar and Shah are already included in their safeguard systems to help prevent unauthorised cloning or misuse of their voice, and that if their voice had been cloned or otherwise misused on the ElevenLabs platform to create the audio files in question, those safeguards would be activated.
For expert analysis, we escalated the videos to our detection partner ConTrails AI, a Bangalore-based startup with its own A.I. tools for detection of audio and video spoofs. The team ran all the videos through audio and video detection models, which indicated that the three audio tracks are A.I. manipulated. They noted that the voice of all the subjects sounds relatively monotonous, the pacing is unnatural, and each of these is most likely a voice-cloning attack.
Their model also predicted A.I.manipulation in the video track featuring Shah’s likeness, however, it did not detect any manipulation in the video tracks featuring the likeness of Modi and Jaishankar, respectively. The team at Contrails noted that while their model gave higher confidence scores to the two video tracks being real, there are visible signs of manipulation near the lip region of the subjects in their respective videos, indicating a lip-sync attack. Adding that, this an advanced lip-sync attack and that they will update their models for better detection.
In their report they noted that in the purported Shah video the lips of the subject look animated, and the portion of the cheeks surrounding the moustache has unnatural lighting at certain points.
For the video featuring Jaishankar’s likeness, the team stated that the subject’s entire mouth region seems animated, and the facial wrinkles unnaturally appear and disappear.
For the video featuring Modi’s likeness, the team noted that the subject’s lips look animated and his facial hair moves unnaturally as he speaks.
We also reached out to our partners at RIT’s DeFake Project to get expert analysis on the video featuring Shah’s likeness. Saniat Sohrawardi and Kelly Wu from the project pointed out several visual inconsistencies. They observed that for every scene transition or cut there is a lip-sync error at the beginning and end of the scene where the mouth does not start following the audio at the correct moment. They added that this is fairly common even for the best lip-sync methods.
Ms. Wu and Mr. Sohrawardi corroborated our observations by highlighting that the visuals exhibit a general smoothness, which is consistent with that seen in generated images and generated videos. They pointed to a specific time code toward the end of the video where the lighting and reflection of the subject’s hands on the table or podium in front seems implausible. They mentioned that the noticeable presence of other visual artefacts in that moment could be due to compression in the video.
Wu and Sohrawardi observed that the starting position for each scene in the video track is the same. According to them, the people who generated the video may have used the same image with cropped variations to generate several image-video scenes and then animated them using lip-sync. The team could not confirm whether a single tool was used or two separate tools were manually used to generate the content; they noted that a thorough look at the trends and available tools was needed to investigate that.
To highlight what was rather well done in the synthetic video the team noted that the facial movements within the scenes are quite good and there are less obvious giveaways; accounting for the smoothened consistency of the video as a possible filter on top of the video. They added that the consistency of the glasses on Shah’s likeness was impressive.
The team concluded that the video is a deepfake as it imitated the likeness of Amit Shah and used it to generate media to showcase actions he has never performed himself.
To get another expert to weigh in on the video featuring the likeness of Jaishankar, we escalated it to the Global Deepfake Detection System (GODDS), a detection system set up by Northwestern University’s Security & AI Lab (NSAIL).
The video was analysed by two human analysts and run through 22 deepfake detection algorithms. Of those, 12 predictive models gave a higher probability of the video being fake and the remaining 10 indicated a lower probability of the video being fake.
In their report, the team pointed to several signs of A.I. manipulation in the video. They noted that the video features multiple cuts with the subject in different angles, which could indicate that the media was edited. They highlighted a specific instance in the video where after a cut the mouth movements do not match the subject’s speech.
They also pointed to a five-second segment where the subject’s forehead and hairline move in an exaggerated manner while speaking, appearing unnatural. Additionally, they noted that at certain points in the video, the bottom of the subject’s tie is a different shade and image quality than the top.
Furthermore, they highlighted that the bottom of his glasses also seem to blur into his face at a specific moment in the video. They also pointed out that his mouth moves unnaturally, appearing artificial or robotic, and occasionally it blurs.
The GODDS team concluded that the video is likely generated with artificial intelligence.
To get another expert to weigh in on the video featuring the likeness of Modi, we reached out to our partner Validia, a San-Francisco based deepfake cybersecurity service. They used their proprietary software to check the authenticity of the video and audio being associated with Modi.
The team at Validia noted that the video is a deepfake and that it is of a much higher quality than those previously analysed by their team. However, they identified multiple visual indicators that helped them in their analysis, including non-natural lip movements and an animated nature of the video.
They suggested that this deepfake was not created using traditional lip-sync tools such as Sync Labs, and rather used a platform like RunwayML to create an animated effect of the prime minister.
To analyse the audio they compared the audio track from the video with a sample of Modi’s real voice and arrived at a similarity score of 0.88. A similarity score of 1 would mean perfect similarity and anything between 0.8-0.99 would be a likely deepfake voice. A further analysis of the heat-maps that they used for the comparison of the two voices pointed to numerous parts that trailed away from Modi’s normal voice.
Based on our findings and analyses from experts, we can conclude that the three videos we assessed through this report are deepfakes of Modi, Shah, and Jaishankar, respectively.
(Written by Debraj Sarkar and Rahul Adhikari, edited by Pamposh Raina.)
Kindly Note: The manipulated audio/video files that we receive on our tipline are not embedded in our assessment reports because we do not intend to contribute to their virality.
You can read below the fact-checks related to this piece published by our partners:
AI Videos Of PM Modi, Amit Shah & Jaishankar Apologising To Pakistan Viral
AI Clips of PM Modi, Amit Shah and Jaishankar Apologising for Op Sindoor Viral
Fact Check: पाकिस्तान से फैलाए गए पीएम मोदी, अमित शाह और जयशंकर के डीपफेक वीडियो
.png)
