Deepfakes Explained: How They Are Made and How to Check a Suspicious Video, Photo or Voice Note

By Ravindra Bagale, trainer · LinkedIn · Last updated: 19 October 2026

Generative AI course · Chapter 30 · Estimated reading time: 13 minutes

1. Suppose a video says a cricketer recommends an app

Suppose Uncle Anil, our uncle in Sangli, forwards a 40-second video to the family group. A well-known cricketer smiles and says: "I invested 10,000 rupees in this app and got 30,000 in a week. Click the link and start today."

Uncle Anil asks: "He is so famous, why would he lie?"

He would not. Most likely, he never said it. Someone took a real interview, changed the lip movement and replaced the voice with a clone. The video looks right because it was built from real footage of him.

We will not argue with Uncle Anil. We will check, and show him how we checked.

2. How deepfakes are made

Face swap

1. The creator collects many photos or video frames of a person's face.
2. A model learns that face from many angles.
3. It replaces the face in another video, frame by frame.

Lip sync

1. The creator takes a real video of the person talking.
2. A model changes only the mouth area so it matches new audio.
3. The rest of the video stays real, which makes it very convincing.

Voice cloning

1. The creator collects a recording of the person's voice, often from public videos, reels or even a voice note.
2. A model learns the voice.
3. It can then say any new sentence in that voice, in seconds.

Fully generated video and images

New video and image models can create a whole scene from a text prompt. No real footage is needed.

The important idea: most of a deepfake is often real. Real background, real body, real clothes. Only the face, the mouth or the voice is changed. That is why "it looks real" proves nothing.

3. Why deepfakes fool us

• We trust faces and voices we know. Our brain checks "is this him?", and the answer is yes.
• Phone screens are small. Blur, compression and low light hide the errors.
• Messages come with emotion. Fear ("your son is arrested"), greed ("double your money") or urgency ("only today") stop careful thinking.
• Forwards feel trusted. It came from Uncle Anil, and Uncle Anil is family.
• Old tips stop working. "Deepfakes don't blink" was true for early tools. New tools fix such signs every few months.

So we do not rely on spotting a mistake. We rely on a routine.

4. The 5-step check

1. Pause. Do not forward, click or pay. A real opportunity or emergency survives 10 minutes of checking.
2. Source. Who posted it first? Is it on the person's official, verified account, or only in forwards?
3. Second channel. Check through a different route: call the person on a number you already have, open their official website, or search reliable news sites.
4. Signs. Look and listen for clues (see the table below). A clue is a warning, not proof.
5. Tools. Run the file through a checker: Gemini's AI verification, Content Credentials, reverse image search.

5. Worked example 1: check a viral photo

Suppose a photo is going around: a "flooded" famous temple, shared with the line "Happening right now!"

Step 1: reverse image search

1. Save the photo (or take a tight screenshot of only the photo).
2. Open Google Lens (in the Google app, or Google Images on a computer) and search with the image.
3. Look for the oldest copy and for news sites that use it.

Expected output (yours will be similar, not the same): either the same photo appears from an old event in another year or place (an old photo with a new story), or you find similar images marked as AI, or nothing at all. "Nothing found" does not mean real; it can also mean newly made.

Step 2: look at the details inside the file. A photo straight from a camera often carries metadata (EXIF): phone model, date, editing software. Forwarded images usually lose it. We tested this with a small Python script on a test photo and a re-saved copy:

# Read the camera details (EXIF) inside a photo file, if any are left.
import sys
from PIL import Image, ExifTags
for path in sys.argv[1:]:
    exif = Image.open(path).getexif()
    tags = {ExifTags.TAGS.get(k, k): v for k, v in exif.items()}
    keep = {k: tags[k] for k in ("Make", "Model", "DateTime", "Software") if k in tags}
    print(f"{path}: {keep if keep else 'no camera details found'}")

Real output (we ran this on 19 October 2026 on a test image with camera details, and a re-saved copy of it):

original.jpg: {'Make': 'ExamplePhone', 'Model': 'Model X1', 'DateTime': '2026:10:12 18:42:10'}
forwarded.jpg: no camera details found

Re-saving removed every detail. So missing metadata proves nothing: most forwards look like forwarded.jpg. Metadata is useful only when you have the original file, for example from the person who took it.

Step 3: Content Credentials. Upload the image at verify.contentauthenticity.org. If the file carries Content Credentials (an industry standard called C2PA), you see who made it, which apps edited it and whether AI was used. Many files have none, especially forwards.

6. Worked example 2: check a video with Gemini

Google's Gemini app can check whether an image, video or audio file was created or edited by Google's AI, using Google's invisible SynthID watermark, and it can read Content Credentials.

Steps:

1. Sign in at gemini.google.com (or the Gemini app).
2. Add the file: one image, video or audio file at a time, up to 100 MB. Videos must be under 90 seconds.
3. Ask: "Was this video created or edited by Google AI?" or "Is this AI-generated?"

Expected output (yours will be similar, not the same): one of three kinds of answer:

• SynthID detected: all or part of the file was made or edited with Google AI. For video, Gemini says which parts.
• Not detected: it was not made or edited by Google AI, but it could still be made by another company's AI.
• Not clear: for example, the content is too simple, or the edit too small.

Google lists a daily limit of about 10 checks each for images, videos and audio.

What this means for the cricketer video: "not detected" does not make it real. A scammer may have used a tool that adds no watermark. That is why steps 1 to 3 of the routine (pause, source, second channel) come before the tools. For the cricketer, step 3 settles it quickly: his verified accounts and reliable news sites say nothing about this app.

7. Worked example 3: a voice note from "your brother"

Suppose a WhatsApp voice note arrives from an unknown number: it sounds exactly like your brother Sagar. "I'm in trouble, my phone broke, please send 15,000 to this UPI ID, don't tell Mum."

The check, step by step:

1. Pause. Do not pay. Notice the pattern: unknown number, money, secrecy, urgency.
2. Second channel. Call Sagar on his old, saved number. If he does not pick up, call someone who is with him, or ask a question only he would know (or your family code word, if you have one).
3. Tools, optionally. You can upload the audio to Gemini and ask if it was made by Google AI. A "not detected" answer still does not clear it.

Expected output (yours will be similar, not the same): Sagar answers his saved number, confused, and says he is fine. The voice note was a clone, probably made from his public reels.

Then: do not reply to the number, report it (see section 9), and warn the family group with a simple message: "A fake voice note in Sagar's voice is asking for money. It is not him. Always call his saved number."

8. What the tools can and cannot prove

No single tool can prove a file is real. Tools can only add evidence. The strongest proof is still the second channel: the real person, the official account, a reliable news report.

9. Where to report, in short

• Lost money, or shared bank details? Call 1930 or report at cybercrime.gov.in at once, and call your bank. Our guide How to Report Cyber or UPI Fraud in India and the cyber crime reporting checklist walk you through it.
• A fraud call, SMS or WhatsApp, but no money lost? Report it on Sanchar Saathi's Chakshu page (sancharsaathi.gov.in) under suspected fraud communication.
• A fake video on a platform? Use the platform's report button (impersonation or fake content).

For the full list of steps, habits and recovery-scam warnings, see the deepfake and voice clone scams guide.

10. Beginner mistakes and quick fixes

11. Try at home

Recap

Run one Gemini check and one reverse image search today, so you know how they work before you need them. In a real emergency, the first step is always the same: pause, and call a number you already trust.

Previous chapter: Is ChatGPT Safe? Privacy Settings and What Never to Paste into AI

Next chapter: AI Detectors and Humanize AI Tools: What Works and What Is Honest

Ravindra Bagale, trainer: linkedin.com/in/ravindra-bagale. The people, messages and videos on this page are made up for learning; the cricketer is not a real person. Gemini's verification steps and limits were checked on Gemini Apps Help, and reporting channels on cybercrime.gov.in and Sanchar Saathi, on 19 October 2026 and can change. The Python output was produced on 19 October 2026 with test images. Tool answers change, so your output will be similar, not the same.