AI video in 2026 is good enough that people rarely spot it from the picture quality. They spot it from five specific tells, and four of the five are fixable with prompting and shot choice rather than with a better model. The fifth, physical causality, is the one that still needs you to shoot it.
1. Physics that almost works
Liquids that pour slightly wrong, fabric that settles too fast, objects that make contact without transferring force. Viewers cannot name it and they register it, because everyone has watched real things fall their whole life.
Fix: avoid shots whose subject is a physical interaction. Generate the room, the mood, the product sitting still. Film the pouring, the clicking, the sealing. This is the same line that holds for product demos.
2. Hands and small articulated things
Improving fast, still the most-checked area. Hands doing something specific are harder than hands existing.
Fix: frame them out, keep them still, or keep them small in frame. If the hands are the point, film them.
3. Eye contact that never breaks
Generated people look at the lens with an unwavering, slightly glassy steadiness. Real people blink irregularly, glance away, refocus.
Fix: prompt for the glance away, the blink, the look down and back. Or frame in profile or three-quarter. Sustained direct eye contact is the strongest uncanny trigger in the format.
4. Motion that is too smooth
Generated camera moves are frictionless. Real camera moves have weight: a tiny overshoot at the end of a pan, a settle after a push.
Fix: prompt for handheld, for the imperfection, for a slight drift. Counterintuitively, asking for a worse camera operator produces a more believable shot. This is why UGC-style generation needs prompts that fight the model's cinematic default.
5. Lighting that never commits
The most common and least discussed. Models default to pleasant, even, sourceless illumination. Real light comes from somewhere and leaves shadows.
Fix: name the source and the direction. "Single window, camera left, hard afternoon light, deep shadow on the right side of the face" produces a more convincing frame than any quality adjective. Committed lighting is the single highest-leverage change available in a prompt.
The finishing pass that helps
Grade it. Untouched model output has a characteristic flat, slightly plasticky look. A real grade, even a light one, moves it a long way.
Add grain. Small amount. Digital cleanliness reads as synthetic, and grain is the cheapest correction there is.
Cut it properly. A generated clip shown full length feels generated. The same clip trimmed to its best two seconds and cut against something else stops announcing itself.
Mix real and generated. One real shot in a sequence raises the perceived reality of everything around it. This is the most underused technique in the category.
What a better model does not fix
Physical causality, hands doing precise work, and a recognisable person's identity across a long take. If your shot depends on any of those, the honest answer is to film that shot and generate everything around it. A different model changes how often it fails, not whether.
FAQ
Which model looks the most realistic? They cluster closer than marketing suggests, and the tells above appear across all of them. Prompt specificity and finishing move the result more than model choice at the top end. Compare on identical prompts in best AI video generator 2026.
Does higher resolution make it look more real? No, and sometimes the reverse: more resolution shows more of the artefacts. Lighting and motion matter far more.
Will viewers care if they can tell? Depends on the format. Social and advertising audiences increasingly do not, provided the work is good. Documentary, testimonial, and anything evidentiary is a different contract entirely.
Commit the lighting, break the eye contact, and film the physics. The 8frame canvas is free and unlimited, and generation is paid from $19/month.