Seven checks before you share

Baby clips are the perfect viral format: short, silent, emotionally loaded and impossible to verify without context. They are also the perfect target for a face filter, because viewers are already primed to find a newborn’s face fascinating and slightly strange.

None of the checks below require software. They require about ninety seconds and a willingness to be the person who does not share something.

The seven checks

1. 01

   Find the earliest upload, not the loudest one

   Search the caption text and reverse-search a still frame. If the first post is an aggregator with a news-style name, the trail has already gone cold. A clip that appears simultaneously on five accounts is a repost, not a source.

   Looks fine

   You found a single, older upload with a normal person attached to it.

   Suspicious

   Every version credits a different country.

2. 02

   Watch the hairline, not the hair

   Overlays fail at boundaries. Look at where the hair meets skin and where it meets the background. Real hair has flyaway strands crossing the edge; a mask has a clean, slightly vibrating cut.

   Looks fine

   Individual strands cross the outline and bend with the skin.

   Suspicious

   The outline slides a few pixels when the head turns.

3. 03

   Compare the light on the hair to the light on the skin

   The most reliable single check. A composited element carries its own illumination. If the cheek has a warm directional shadow and the hair is evenly, flatly lit from nowhere, they were not photographed together.

   Looks fine

   Highlights and shadows run in the same direction on both.

   Suspicious

   Hair is uniformly bright while the face is half in shadow.

4. 04

   Look at the mouth, teeth and tongue

   Mouth interiors are still the hardest region for generative video: teeth merge, tongues lose their shape, and the boundary between lips and gums softens. A closed mouth is easier to fake than an open one.

   Looks fine

   Teeth are distinct and the tongue holds a shape.

   Suspicious

   A white smear where the teeth should be.

5. 05

   Check hands — but do not stop there

   The old advice was 'count the fingers'. Modern tools usually get hands right, so clean fingers mean little. What still fails is contact: a hand touching a cheek should compress the skin and cast a shadow on it.

   Looks fine

   Fingers dimple the skin and the shadow lands on the face.

   Suspicious

   The hand floats above the cheek with no contact shadow.

6. 06

   Read the background like a document

   Medical equipment, signage, wristbands and monitor text are where generated scenes quietly fall apart. Lettering becomes decorative squiggles; tubes end in mid-air; two objects occupy the same space.

   Looks fine

   Text is legible and consistent when the camera moves.

   Suspicious

   Hospital signage reads like a font rather than a language.

7. 07

   Ask what the original file would look like

   The step almost nobody takes. A viral claim should be able to produce an unedited file, a date, a place and a person who filmed it. If the answer to all four is 'no', the correct label is unverified — not real, and not fake.

   Looks fine

   The uploader can supply the original and a location.

   Suspicious

   Ten million views, zero provenance.

Take the five-question check

No score is published, nothing is stored, and there is no leaderboard. Answer all five to see the explanations.

Why “an AI detector said so” is not a check

Detectors and chatbots estimate plausibility from pixels. They do not have the file, the metadata, or the account history. Two general models looked at the baby with mustache clip and did not flag it — which tells you something about the models, not about the clip.

The failure mode runs both ways. A detector that says “87% AI” can be wrong about a real video of a real child, and that accusation follows a family around. Provenance is slower, duller and far more reliable.

When “fake” is not the right word

There are four honest labels, and most arguments collapse because people skip straight to the last one:

“Unverified” feels like a cop-out and is actually the most accurate thing most of us can say. It also leaves room to update when evidence arrives, which is more than a confident verdict usually does.

A thirty-second routine

1. Pause. Do not share on the first watch.
2. Search the caption. Find the oldest post you can.
3. Watch the hairline and the shadows once, deliberately.
4. Ask who filmed it, where, and when. If nobody knows, say so.
5. If you do share it, add one sentence of context. That sentence is the entire intervention.