AI Media LiteracyHow to Tell If an Image or Video Is AI, and What 1.4 Million Views Taught Us About Teaching It

The short answer

You can usually tell an AI image or video from a real one by checking things the model never had to get right: physics, light and continuity. A real camera records one world, lit by real light, obeying gravity. An AI model paints pictures that look like that world, one guess at a time, and the guesses do not always agree with each other.

The fastest checks:

1. Shadows. In a real photo, every shadow points back to the same light. In AI images they often disagree.
2. Reflections. Mirrors, windows and even the two eyes of one person should reflect the same room.
3. Background blur. A real lens softens the background gradually with distance. AI blur often drops off like a cliff around the subject.
4. Fine detail. Zoom into fabric, hands and hair. AI detail tends to melt into smooth paint.
5. Motion. In AI video, hair ignores the wind, wheels spin at the wrong speed and nobody breathes.
6. Continuity. Pan the camera away and back. In AI video, the room is often not the same room.

No single tell is proof. Use several together, and treat a clean result as "no evidence found," not "verified real."

Why a production and research house teaches people to spot AI

We make original series, films and shorts, building and using AI to amplify what our artists can do, and we research how far it can expand what stories can be. That puts us in an unusual position: we spend every day looking closely at exactly where generated images break, because that is where our artists have to step in.

Through October 2026 we turned that working knowledge into a series of short films, each one built around a single question about AI images and video, answered with the evidence on screen. On Instagram the campaign passed 1.4 million views in under 20 days. One film, AI leaves no fingerprint, reached 963,000 views on its own. The second strongest, Your face has a pulse, reached 110,000.

We think media literacy is part of the job for anyone making work with these tools. If audiences can't tell what's real, they stop trusting everything, including the real thing. This post is the long version of the series, written so you can use it, share it, and check our sources.

How to tell if an image is AI: six tells

1. It has no camera fingerprint

Every camera sensor is slightly imperfect, and those imperfections stamp a faint, unique noise pattern on every photo it takes. Forensic researchers call it PRNU (photo-response non-uniformity). In a 2006 study, researchers matched about 2,700 photos to the exact camera that took them using this pattern alone, even when two cameras were the same model.

An AI image was never taken by a sensor, so it carries no sensor fingerprint at all.

How to use it: this is a lab test, not a glance test. Forensic tools can check for it. But it is the clearest illustration of the core idea: a real photo is evidence of a moment; an AI image is a picture of what such a moment would probably look like.

This was the campaign's best-performing film by far. The comments asked whether an AI could learn to fake a fingerprint, and we answered in a follow-up: the pattern is specific to one physical sensor, so faking it convincingly means knowing that one sensor's noise, which the model never had.

2. The shadows don't agree

In a real photo lit by the sun or one lamp, a line drawn from any shadow through the object that casts it will meet the others at a single point: the light. This is a well-established forensic test from image-forensics researcher Hany Farid. AI images often fail it, because each shadow is painted to look plausible locally, not computed from one light.

How to check: pick two or three clear shadows, draw lines from shadow to object, and see whether they converge.

3. The reflections show a different room

The same geometry applies to reflections. Lines joining objects to their reflections in a mirror or window should meet at one point. In AI images the reflection is often a plausible-looking room, just not this room.

The sharpest version is the eyes. Both corneas are curved mirrors about six centimetres apart in the same space, so they should reflect the same window or light. AI portraits often show each eye reflecting a different scene.

4. The background blur is a cliff, not a slope

A real lens softens things gradually: the further something is from the focus distance, the blurrier it gets. AI "portrait" images often cut the subject out and blur everything behind it by the same amount. Look at the ground plane or a wall receding into the distance. Real blur ramps. Fake blur jumps.

5. The fine detail melts when you zoom in

Real suiting is woven: a twill shows a steady diagonal at about 45 degrees all the way across. Zoom into an AI jacket and the weave slides into smooth paint. Hands are the famous case. They are small in most training photos, they fold many bones in many ways and hide their own fingers, and most image models work on a heavily shrunken version of the picture, so finger counts and joints go wrong.

6. Everyone looks like the same cousin

AI faces drift toward one attractive average. Models blend faces toward the mean of what they learned, learn from photos already filtered for looks, and are then tuned toward what people rate as appealing. The result is a family resemblance across unrelated AI people, the same glow and the same grade.

How to tell if a video is AI: six tells

7. The camera moves, but the world doesn't hold

An AI video clip is not a camera moving through a place. It is a sequence of guessed pictures. When the camera moves, nothing holds the wall in place, so architecture slides, bends or quietly rebuilds itself. When we need a camera to move through a generated world in our own productions, we rebuild the scene in 3D and put a real virtual camera in it.

8. Anything off-screen is forgotten

An AI video has no world outside the frame. Whatever leaves the model's working memory is gone. Pan away and pan back, and the room has changed: a door moves, a painting disappears, a person is wearing something else.

9. Physics is guessed, not solved

A simulation solves gravity, mass, friction and collision every frame. AI video guesses what the next picture should look like from patterns. That is why, in a storm, the trees bend while the character's hair stays perfectly still, and why car wheels drift, freeze or spin backwards: the spokes are painted, not linked to the distance the car travels.

10. Nobody breathes

Real people breathe, and it shows: chests rise, speech pauses carry an inhale, shoulders move. AI voices are largely learned from text that never marks a breath, so pauses come out silent, and AI characters' chests stay still. Watch the lips on P, B and M too. Those sounds need the lips to fully seal; generated mouths often fake the shape without the seal.

11. The film grain is fake

AI video often hides plastic-looking skin under a uniform layer of filter grain. Real film and sensor grain change with exposure: shadows and highlights fizz differently. Grain that is equally strong everywhere is a costume.

12. Upscaling to 4K makes it worse, not better

Generated video has small frame-to-frame wobble at edges, because fine details are re-guessed each frame. The human visual system is extraordinarily sensitive to this kind of misalignment (vision science calls it hyperacuity). Real footage holds still, so higher resolution only adds detail. Upscaled AI footage adds evidence.

Why these tells exist

All twelve tells come from the same few facts about how today's image and video models work:

• It starts as static. A diffusion model begins with pure noise and cleans it up over many passes, often around 50. Fine details are invented during that process, and in video they are invented fresh for each frame, so they don't always agree.
• The model works on a stamp, not the picture. Most video models never touch full-size frames. An encoder shrinks the video roughly eight times in each direction and folds several frames into one; the model works on that compressed version, and a decoder blows it back up. Small things, like fingers, text and weave, get lost in the squeeze.
• It prints the average. Models reproduce the centre of what they learned, then get polished toward what people scored as pretty. That is why so much AI work shares one look.
• It has no world. There is no scene, no light source and no camera behind the image, only a picture that resembles one. Every tell above is a place where a real world would have forced consistency.

This is also why the tells are a moving target. Each one is an engineering problem, and engineering problems get worked on.

What doesn't work as proof

Some popular "AI detectors" mislead more than they help.

• Error Level Analysis glow is not proof. The bright patches people screenshot as evidence of editing are usually JPEG compression rounding on resave, in the format's 8×8 blocks, not a pasted-in face.
• AI labels don't survive screenshots. Provenance labels such as Content Credentials (the C2PA standard) live in the file's metadata beside the pixels. A screenshot creates a new file from the pixels alone, so the label is gone.
• Single biological signals are being learned. Skin subtly changes colour with every heartbeat, and cameras can measure it, which made pulse detection a strong deepfake test. By 2025, fakes had begun to reproduce a pulse.
• Your gut is not enough. In a 2022 study by Nightingale and Farid, people identified real versus AI-generated faces correctly only 48.2% of the time, worse than a coin toss, and rated the AI faces as more trustworthy.

There is a second-order risk too. Legal scholars Chesney and Citron call it the liar's dividend: once convincing fakes exist, genuine footage can be dismissed as fake. Media literacy is not only about catching fakes. It is about keeping the ability to trust what is real.

What 1.4 million views taught us about teaching media literacy

The campaign was also an experiment, and the results are worth sharing for anyone communicating about AI.

Type of film	Example	Instagram views
Physical proof you can't argue with	AI leaves no fingerprint	963,000
Your own body as the test	Your face has a pulse	110,000
Spot-the-mistake	Why AI gave hands six fingers, Fake blur is a cliff	2,000 to 50,000 per film
Cryptic, curiosity-led titles	Check the mirror, Hand it back to the AI	2,000 to 3,000 per film
AI news and research stories	AI trained on AI collapses	under 5,000 per film

Three findings stood out:

1. Proof beats opinion. The films that showed a physical fact, a sensor pattern or a heartbeat, outperformed everything else by an order of magnitude. People share what they can verify.
2. Make it personal. The second-best film turned the viewer's own face into the evidence. Literacy lands when it is about you, not about the technology.
3. Clarity beats mystery. Cryptic titles underperformed plain, specific claims. "AI leaves no fingerprint" tells you exactly what you will learn.

What this means for brands and filmmakers

If you publish images or video, AI literacy now sits on both sides of the screen.

• Your audience is checking. The tells above are spreading fast; a campaign that fails them reads as cheap, whatever it cost.
• Platforms are checking. AI tools write provenance metadata into files, and platforms such as Instagram can read it on upload and add an "AI info" label to your post or ad.
• The prompt is step one. In our own productions the prompt is the first of seven stages. Selection, boards, 3D, animation fixes, compositing and the final grade are the other six, and those six are where the tells get caught.

This is the same position we argue in AI Slop Is a Craft Problem, Not a Tech Problem: the tools are not the risk; work made without judgement is.

Frequently asked questions

Can you tell if an image is AI-generated with 100% certainty?

No. No single visual check is conclusive, and the best fakes pass many of them. Combine several tells (shadows, reflections, blur, fine detail) with source checks: who posted it first, and is there an original file with intact metadata?

What is the most reliable way to spot an AI image?

For an ordinary viewer, light and geometry: shadows that point to different lights and reflections that show a different room are hard for current models to get right. For forensic work, sensor fingerprint analysis (PRNU) is far stronger, because AI images have no sensor.

Why does AI get hands wrong?

Hands are small in most photos, take a huge range of poses and often hide their own fingers. Image models also work on a shrunken version of the picture, so that small, complicated detail gets lost and then re-invented.

Do AI labels and watermarks protect audiences?

Partly. Metadata labels such as Content Credentials disappear when someone screenshots the image. Researchers are developing watermarks planted in the noise before the image exists, which are designed to survive crops and screenshots, but no single method is universal yet.

Will these tells stop working?

Some already are. Pulse-based detection weakened once fakes learned a heartbeat. The underlying principle lasts longer: a real image records one consistent world, and a generated one has to fake that consistency everywhere at once.

Watch the series

The full campaign is on YouTube as AI Media Literacy by Plattipus, with new films added as our research continues.

If you are a filmmaker, brand or research partner working on AI provenance, literacy, or work that has to hold up to a close look, collaborate with us.

Sources

• Lukáš, Fridrich and Goljan, Digital Camera Identification From Sensor Pattern Noise, IEEE Transactions on Information Forensics and Security, 2006.
• Hany Farid, work on shadow and reflection consistency in photo forensics.
• Westheimer, 1975, on visual hyperacuity.
• Nightingale and Farid, AI-synthesized faces are indistinguishable from real faces and more trustworthy, PNAS, 2022.
• Chesney and Citron, Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security, California Law Review, 2019.
• C2PA, Content Credentials technical specification.
• Instagram view counts as logged by the Plattipus team, 15 to 23 October 2026.