Telling Real from Fake in a Synthetic World

For most of history, a photograph or a recording was strong evidence that something happened. AI has quietly ended that assumption. Systems can now generate faces, voices, and scenes convincing enough to fool careful people. This is not only a problem of being tricked by fakes; it is also that real evidence can now be waved away as "probably fake."

The goal of this lesson is a middle path: neither believing everything nor doubting everything, but building habits that let you judge what is in front of you.

Synthetic media and deepfakes

Synthetic media is any image, audio, or video generated or heavily altered by AI. Deepfakes are a subset that convincingly depict real people saying or doing things they never did. The technology has improved fast and is widely available.

Why your eyes are not enough

Early fakes had tell-tale glitches — odd hands, warped backgrounds, unnatural blinking. Detectors that rely on such artefacts date quickly as the technology improves. Spotting fakes by eye is a losing game over time.

Provenance beats forensics

The durable strategy is to ask where did this come from? rather than does this look fake? Provenance — the traceable origin and chain of a piece of media — is more reliable than inspecting pixels. Who published it first? Do trusted outlets carry it? Does it match other records of the event?

A shocking video of a public figure circulates on social media. Instead of studying the footage, you check provenance: no reputable news organisation has it, the account that posted it is days old, and no other angle of the "event" exists. Low provenance, high caution — you don't share it.

Doubt can go too far. If you dismiss a genuine, well-sourced recording as "probably AI" simply because fakes exist, you have handed wrongdoers what researchers call the liar's dividend — the ability to escape accountability by calling real evidence fake. Verification cuts both ways.

Researchers studying AI-generated faces have reported that people struggle to tell them from photographs of real people. In studies published around 2022 (for example, work by Nightingale and Farid), participants distinguished AI-synthesized faces from real ones at close to chance levels, and in some cases rated the synthetic faces as more trustworthy-looking than real ones. The finding is a warning: "it looks like a real person" is no longer evidence that it is one. Judgment has to move from the image itself to its source and corroboration.

• "I can spot a fake if I look closely." Artefact-spotting fails as models improve.
• "A detection app will tell me for sure." Detectors are useful but imperfect and can be evaded.
• "If it might be fake, assume it is." That fuels the liar's dividend and dismisses real evidence.
• "Only video can be faked." Text, voice, and images are all cheaply synthesised.

Take one striking image or clip from your feed today. Before reacting, find its earliest source and check whether any trusted outlet corroborates it. Decide: share, hold, or discard — and why.

Think Like a Maester: When you cannot trust the picture, trust the trail that leads to it.

1. What is a deepfake?
2. Why is spotting fakes "by eye" an unreliable long-term strategy?
3. What does provenance mean, and why does it help?
4. What is the liar's dividend?
5. Describe one habit for checking a suspicious image.

AI can now fabricate convincing images, audio, and video, so seeing is no longer believing — and real evidence can be dismissed as fake. Rather than hunting for pixel-level glitches, verify provenance: trace origin, check trusted corroboration, and weigh context. The aim is calibrated doubt that resists both gullibility and reflexive cynicism.

• Spotting Misinformation and Disinformation
• AI as a Thinking Partner, Not a Replacement