Deepfakes and misinformation: how to tell what is real

Learn in your browser · Chapter 01 of 06

Deepfakes and other AI-made media are not a future problem. This chapter defines the terms properly and sets the honest baseline: detection is getting harder, not easier, so a verification habit matters more than a sharp eye.

A deepfake is a video, image or audio clip that has been generated or altered by AI to make a real person appear to say or do something they never said or did. The wider category is synthetic media: any image, audio or video created by generative AI, whether or not it depicts a real person. Neither term is new any more, but the tools that make them have kept improving, and that changes what a sensible verification habit needs to look like.

People make deepfakes and synthetic media for a wide range of reasons, and only some of them are malicious. Film studios use the same underlying technology for visual effects and dubbing. Marketers use it for adverts. The concerning uses are the ones this workbook focuses on: fake endorsements used to sell a scam investment, cloned voices used to impersonate someone in an emergency, and fabricated video used to make a public figure appear to say something they never said.

A worked example (invented). A video circulates on social media showing a well-known public figure enthusiastically endorsing a new investment app, promising fast returns. The account posting it is new, the video has an unusual number of shares for its age, and the comments are full of people asking how to sign up. None of that proves the video is fake by itself, but every one of those details is a reason to pause and verify before you believe or share it, which is exactly the habit the rest of this workbook builds.

Learn in your browser · Chapter 02 of 06

Some visual tells still catch a fake today. None of them are permanent, and none of them are proof on their own. This chapter teaches you what to look for and how much weight to give it.

Visual artefacts are the small, often subtle mistakes a generative model leaves behind because it is predicting pixels statistically rather than filming a real scene. They cluster around a handful of details that are genuinely hard for current models to get consistently right.

Scroll sideways to see every column.

Learn in your browser · Chapter 03 of 06

Cloned voices and dubbed video have their own tells, separate from the visual ones. This chapter covers what to listen for, and a live technique that works well on a call.

Voice cloning copies a real person's voice closely enough that a scammer or hoaxer can make it say anything they type or script. On video, a mismatched lip sync happens when the mouth movements do not quite match the words being spoken, often because the audio and video were generated or edited separately.

• Flat or uneven emotional tone: fear, urgency or excitement that sounds slightly wrong for what is being said
• Unnatural pacing: pauses in the wrong places, or breathing sounds that are missing entirely
• A mouth shape that does not quite match certain sounds, especially ones made with closed lips like 'm' or 'b'
• Background noise that does not fit the claimed location, or that stays identical for an unusually long stretch
• A voice that shifts slightly in quality or pitch partway through a clip

Learn in your browser · Chapter 04 of 06

A picture or video clip did not appear from nowhere. Reverse image search finds its earliest known appearance, which is often the fastest way to unravel a fake.

Reverse image search takes an image, rather than a typed query, and searches for other copies of it online. It is free, takes under a minute, and often does more to unravel a fake than any visual tell in this workbook, because it can show you where an image actually first appeared, and in what context.

1. Save or screenshot the image, or copy its web address if it is already online.
2. Open Google Images, Google Lens or TinEye and upload the image or paste its address.
3. Sort the results by date where the tool allows it, and look for the earliest match.
4. Read the caption and context on that earliest match. Is it describing a completely different event, date or person?
5. Check whether the same image has been reused with a different caption elsewhere. This is one of the most common misinformation patterns: a real photo, attached to a false claim about when or where it was taken.

Learn in your browser · Chapter 05 of 06

The strongest single verification move is often the simplest: stop reading the page in front of you, and go and check what everyone else is saying.

Lateral reading is a habit professional fact-checkers use in place of staring harder at a single page or clip. Instead of scrutinising the content itself for longer, you open new tabs and check what independent, credible sources say about the same claim. It works because a convincing fake is designed to survive close reading. It is much less likely to survive being checked against several other, unrelated sources at once.

• Search the claim itself in a few words, not just the source's name, and see who else is covering it.
• Look for at least two or three outlets that appear to have done their own reporting, not just repeated each other or the original post.
• Check whether any official body, the person supposedly shown, or a relevant organisation has issued its own statement.
• Be suspicious if the only coverage you can find traces back to a single original post with no independent reporting at all.

Provenance is a technical complement to lateral reading, not a replacement for it. Content Credentials, an open standard from the Coalition for Content Provenance and Authenticity (C2PA), attach a signed, tamper-evident record to a file, stating where it came from, what was edited, and any AI generation step the tool recorded.

Scroll sideways to see every column.

Learn in your browser · Chapter 06 of 06

Everything so far comes together into one short routine. This chapter gives you the checklist in one place, then a final exercise that puts the whole thing to work.

1. Pause. Notice if you are reacting to strong emotion, urgency or a famous name before you have checked anything.
2. Look for visual tells if it is an image or video, and audio tells if it has sound, treating every one as a clue rather than a verdict.
3. On a live call, ask something unscripted that only the real person would know or could answer naturally.
4. Reverse image search it, and read the earliest match's original context and caption.
5. Check whether independent, credible sources are reporting the same claim, using lateral reading rather than reading the original post again.
6. Check for a Content Credential if the platform shows one, remembering it proves origin, not truth.
7. If you are still unsure after all of that, do not share it, and say so if you pass it on: label it unverified rather than presenting it as fact.

What you can do now. You have a working verification checklist, practice reading visual and audio tells honestly, and a habit of checking a claim against its original source and against other independent, credible reporting before you trust or share it. That habit outlasts any single detection trick, because it does not depend on generation tools staying exactly as good as they are today.

Keep this workbook. It is yours to keep, print and write in. Come back to this chapter's checklist whenever something crosses your feed that gives you pause.

https://trust-agent.ai/exam/deepfakes-and-misinformation

This is free educational content published by Mickai LTD, not professional, legal, financial or compliance advice, and Mickai accepts no liability for how it's used.

Trust Agent publishes workbooks like this one for free because Mickai believes a public that can verify what it sees, rather than simply trust or distrust it by instinct, is what keeps AI's benefits real and its harms in check.