The Feeling Before the Explanation
There is a moment when you watch AI-generated footage for the first time. The first two seconds read as real. Then something at the edge of the frame pulls at your attention: a hand with an extra knuckle, a reflection that does not match the room, a crowd that seems to breathe in unison. You cannot name what is wrong. You just feel watched.
Production teams call this the haunting effect. It is not a flaw in your perception. It is the predictable result of systems that learn the surface statistics of the visible world without learning the causal rules underneath. Clips generated by Sora and its contemporaries look convincing in stills and unsettling in motion for the same reason: the model knows what reality tends to look like, not what reality tends to do.
That distinction is a production variable, not an academic curiosity. It shows up in client reviews as 'I cannot explain it, I just do not like it,' and it costs revision rounds. What follows is a breakdown of the mechanisms behind the effect, then the practical half: diagnostic checks, prompt structure, a repeatable workflow, model selection criteria, and the ethical questions that come with footage that looks real.
Where the Unease Comes From: Six Mechanisms
Across a few hundred generated clips, discomfort almost always traces back to one of six sources. Naming them turns a vague feeling into a fixable defect.
Physics That Almost Holds
Video models internalize an approximation of physics. Objects fall, fabric folds, liquids pour, and all of it looks plausible until you track one object across time. Conservation rules are statistical, not enforced. A glass refills as the camera passes. A ball changes spin mid-air. Your brain runs a lifetime of physical predictions on autopilot, and a failed prediction registers as wrongness long before it registers as an error you can point to.
Identity Drift Across Time
Character consistency remains the hardest problem in generated video. A face shifts subtly across shots: hairline, eye spacing, jaw shape, the number of visible teeth. Inside one short clip the drift is nearly invisible. Across a sequence it becomes unsettling, because viewers experience it as a person being quietly replaced frame by frame. Any project featuring the same character in more than two shots has to plan for it.
Faces Doing Something Human-Adjacent
Real faces move with tiny asynchronies. Eyes lead the mouth by milliseconds, blinks arrive irregularly, pupils respond to light, and micro-expressions flicker for a fifth of a second. Generated faces hit the big beats — smile, frown, surprise — while smearing the micro-timing. This is the classic uncanny valley wearing a photoreal costume, and it explains why a still frame can look flawless while the moving version makes viewers lean back.
Texture, Light, and Lens Logic
Synthetic footage often ignores the optical rules of a real camera. Depth of field does not match any plausible aperture. Shadows fall without a consistent light source. Skin is glossy in a matte environment. Fast pans arrive without motion blur. Viewers rarely say 'the lens physics is wrong'; they say 'it looks like a video game,' which is the same observation in different words.
Sound That Does Not Belong to the Image
Most generations ship silent, so creators drop in stock music and call it finished. Footfalls do not sync with steps. The room has no tone. Dialogue does not match lip movement. Audio is also the cheapest fix available: room tone, a few foley hits, and careful level matching can ground a shaky clip almost instantly, and it is the highest-return repair in the whole workflow.
Narrative Gaps
Even a technically perfect clip feels wrong when the action refuses to resolve. A character reaches for a door and the clip ends. A car approaches and the cut arrives before impact. Humans are prediction engines that want closure: a beginning, a movement, a small resolution. Generations containing motion without consequence read as fragments of a dream — precisely the sensation audiences describe as haunting.
A Diagnostic Checklist for Suspicious Clips
Before regenerating anything, run a ninety-second pass. It prevents the most common waste of time: discarding a usable clip over one fixable detail.
- Compare the first and last frame side by side and describe out loud what changed.
- Examine hands, teeth, ears, and jewelry, the classic failure zones.
- Check every reflection and shadow for a consistent light source.
- Pick one object and trace its trajectory across all frames for continuity.
- Watch once muted, then listen once with your eyes closed.
- Count the story beats and ask whether anything resolves.
Log every failure in a running document with the prompt attached. After twenty clips you will have a personal list of what your prompt style tends to break, and that list is worth more than any generic prompting guide.
Prompt Architecture That Reduces the Uncanny
Most prompts describe a scene. Better prompts describe a scene plus the rules the scene must obey. A five-slot structure covers almost everything:
- Subject — who or what, with one or two identifying details.
- Action — one clear motion, not three.
- Camera — shot size, height, lens, movement.
- Light — the source, direction, and quality.
- Constraints — physical facts that must remain true for the entire clip.
A quick comparison. 'A woman walking through a rainy city street at night' gives a model enormous freedom to improvise, and improvisation is where errors live. A tighter version: 'Medium shot, a woman in her thirties wearing a soaked olive raincoat walks toward the camera along a narrow alley; steady handheld camera at chest height, 35mm lens, shallow depth of field; single overhead sodium lamp plus wet pavement reflections; rain falls at a constant angle, her coat stays on, puddle ripples spread outward only.'
The second prompt works because its constraints are checkable: constant rain angle, coat stays on, ripples spread outward. Note that all of them are phrased positively. Models handle negations weakly; 'no extra fingers' is far less effective than 'left hand relaxed at her side, right hand holding a phone.' Keep one action per clip, avoid rapid camera moves, and be cautious with crowds, mirrors, and hands as focal points, since those subjects concentrate the model's weakest predictions.
A Practical Production Workflow
The gap between a fun demo and reliable output is process, not prompting talent. This sequence works for commercials, narrative shorts, and social content alike.
Stage 1: Shot List Before Generation
Write the shot list first, with four columns: action, duration, continuity anchors such as wardrobe, props, and light direction, and an accepted-variant count. Planning continuity before generating is what prevents identity drift later, because you can group shots by wardrobe and lighting.
Stage 2: Generate Short, Select Aggressively
Generate five to ten variants of a four-to-six second beat rather than one long take. Keep the best three and save the prompts that produced them in a project file. Shot-length discipline matters: the longer a model runs unsupervised, the more errors accumulate, and errors compound rather than cancel out.
Stage 3: Repair Before You Regenerate
Extension, inpainting, frame interpolation, and a color pass solve most problems for a fraction of the effort of a fresh generation. Regenerate only when the motion logic itself is wrong. Learn to separate 'this looks slightly off' from 'the fundamental action is confused,' and let only the second category earn a new generation.
Stage 4: Cut for Rhythm, Then Build Sound
Do not let clips run to their natural length. Cut ten to twenty percent earlier than the moment an error becomes visible — audiences rarely notice a slightly early cut, but they always notice a hand melting. Build room tone and foley before adding music, then let the score support the ambience instead of replacing it.
Stage 5: Final Polish and a Fresh-Eyes Pass
Add subtle grain, a grade, and a hint of lens imperfection. Slight imperfection is the antidote to footage that reads as too clean. Then have someone who did not work on the project watch it once and tell you where they stopped believing it. That single sentence is usually the most useful note you will receive.
Matching Models to Shot Types
Model capabilities shift quickly, so treat this as a decision framework rather than a ranking. Broad text-to-video models such as Sora and comparable systems handle complex scene simulation and long descriptive prompts. Faster lightweight models win on iteration speed for moodboards and style tests. Image-to-video gives you control when composition is the priority: lock the frame first, then animate it. Video-to-video is for restyling existing footage. Specialized utilities handle talking heads, upscaling, and cleanup, and they are often what rescues a shot a main model nearly got right.
When evaluating a tool for a real project, ask five questions: what is the maximum clip length before coherence breaks, how well does it obey physical constraints written into a prompt, does it support image-to-video, what resolutions and aspect ratios come out, and what do the usage terms allow for commercial work? Answer those honestly and the choice usually makes itself.
Ethics, Consent, and Disclosure
Footage that looks real creates obligations that footage clearly does not. Do not generate recognizable real people without consent, and be especially careful with public figures, because a convincing synthetic clip of a real person can cause harm that no later correction undoes. Avoid using generated footage for documentary-style reenactments of real violence or disasters as entertainment.
Keep provenance records — prompts, seeds, model versions, dates — so you can answer questions about how a shot was made. If you build a synthetic performer, document the consent and licensing behind the likeness and treat that documentation as a production asset. Disclosure is not a legal technicality; it is part of the craft. The same realism that makes generated footage powerful is what makes audiences uneasy when they suspect they are being fooled. A simple label protects the credibility of everything else you make.
Mistakes That Amplify the Effect
- Letting clips run long. Errors accumulate, so cut early.
- Stuffing prompts with style words but no physical constraints. 'Cinematic, hyper-real' describes a look, not a behavior.
- Centering hands, crowds, or mirrors. These are high-risk subjects and belong in the background, not the spotlight.
- Music-only audio. Ambience is what makes an image feel like a place.
- Regenerating instead of repairing. Most problems are cheaper to fix than to reshoot.
- Ignoring frame edges. A large share of glitches appear at the borders where the model has less training pressure.
- Using generated footage to imitate a specific living person or a protected brand identity.
- Skipping the fresh-eyes review. A creator's brain learns to ignore errors after the fifth viewing.
FAQ: Practical Questions Creators Ask
Why does a clip look fine on my phone but wrong on a big screen? Small screens hide micro-detail, and motion blur plus compression smooth over timing errors in faces and hands. Review at the largest size you can before final delivery.
Does higher resolution fix the uncanny feeling? No. Resolution increases detail; it does not add causality. A sharp clip with wrong physics is still wrong, and sharper detail can make the errors more visible.
How long should a generated clip be? Four to six seconds is the sweet spot for most models. Anything longer needs a reason, and that reason should usually be an edit point.
What is the fastest way to fix a bad face? Reframe or cut around it rather than regenerating the whole shot. If the face is essential, try an image-to-video pass that starts from a clean still.
Do negative prompts work? Weakly, and inconsistently between models. Positive constraints perform better: describe what should be happening instead of what should not.
How do I stop character drift across shots? Fix wardrobe and lighting in writing before generating, keep shots short, and reuse reference frames. Grouping all shots of one character into a single generation session also helps.
Can I use generated footage commercially? That depends entirely on the terms of the specific tool, and terms change. Read them per project, keep records, and avoid trademarks, likenesses, and music you have not licensed.
Will these problems disappear? Physics and consistency keep improving, but audiences also become more practiced at spotting synthetic motion. The practical answer is not to wait: build a workflow that assumes imperfection and designs around it.
The haunting quality of AI video is not a permanent curse, and it is not only a problem to eliminate. It is a signal that shows you where a clip stops being believable. Learn to read that signal quickly, fix what is cheap to fix, cut around what is not, and let sound and editing carry the rest. That is the whole craft in one sentence, and it stays true regardless of which model you open next.
