AI video limitations in 2026 are mostly about accuracy and control, not picture quality. Generated footage can look superb, yet it still struggles to show an exact product, readable text, a precise movement, a real person or one continuous minute of action. Below are ten limits as they stand in September 2026, with why each happens and what a studio does about it, including the cases where the honest answer is to film it.
Why does AI video get details wrong?
A video model does not build a scene from objects, the way 3D software does. It predicts plausible pixels from the enormous amount of footage it has seen. Plausible is the key word. The model knows what a perfume bottle generally looks like and has no idea what yours looks like. Almost every limit below follows from that.
The ten limits, and what a studio does about each
1. Exact products and readable text
Labels drift, proportions shift, logos turn into near-letters. Large, short text is often correct now, but small print and ingredient lists still break, and we would not trust any model with Cyrillic lettering. The workaround: start every product shot from your real photographs or 3D files, keep the product movement simple, check each frame, and add text, logos and the final packshot in the edit. That is how we approach product videos, and our guide to keeping the product accurate in AI product videos goes through the method step by step. If the viewer has to read the label while the product turns in someone’s hand, film it.
2. Hands and fine interactions
Hands have improved a great deal when they are still. They fail when they do something precise: opening a clasp, tying a knot, typing, applying mascara, slicing an onion. The model has no idea of contact, so fingers pass through objects or objects move on their own. Workarounds are to cut around the moment of contact, shoot wide, or film a real hand insert on a phone and cut it in. Tutorials that depend on technique should be filmed.
3. Long continuous takes
Single generations are short. Google’s Veo documentation lists 4, 6 or 8 seconds per clip, with extension in 7-second steps up to 148 seconds at 720p only. Kling 3.0 generates 3 to 15 seconds. Seedance 2.5, released on 31 July 2026, reaches 30 seconds in one pass. Extension continues from the last second of a clip, so faces, light and props drift a little further with each step. The workaround is ordinary film grammar: short shots, cut together. A two-minute unbroken walk-and-talk should be filmed.
4. The same character across many shots
One consistent face in three shots is easy. The same person in thirty shots, in different light, outfits and angles, is work. Reference inputs help: Veo 3.1 accepts up to 3 reference images, and ByteDance says Seedance 2.5 takes up to 30. Studios build a character sheet first, generate every key frame as a still from those references, and reject any take where the face slips. We describe that workflow in full in how to keep an AI character consistent across 20 shots. Expect small differences in teeth, ears and hairline that only a frame-by-frame check will catch.
5. Precise choreography and physics
Dance to a specific count, a golf swing, liquid poured to an exact level, a gymnast, two people shaking hands and then hugging: movement that must be correct, not merely convincing, is unreliable. ByteDance says as much in its own Seedance 2.5 announcement, noting room for improvement in the physical plausibility of complex motions and in scenes where several subjects interact. Workarounds are a filmed reference video to drive the motion, slow motion, and cutting before physics gives up. Sport and dance that experts will watch should be filmed.
6. Real people and real places
A model can produce a street that feels like Soho. It cannot produce your shopfront, your dining room or your founder as they really are, and a near miss is worse than nothing, because customers will visit and compare. Anything that works as evidence, such as premises, staff, customers or results, should be filmed. A phone is enough. We explain the split in AI video or a real shoot.
7. Exact brand colours
Models think in light, not in Pantone references. Your particular teal will come out warmer in one shot and greener in the next, because the generated lighting changes it, exactly as real lighting would. The workaround is post-production: isolate the product or the brand element in the grade and correct it shot by shot against a reference. Flat graphic elements in brand colours should be added in the edit and never generated.
8. Lip-sync in close-up
Native dialogue arrived in 2025 and is now common, but it holds up best in medium shots and short lines. In a tight close-up viewers notice soft consonants, teeth that change and a mouth slightly out of time. Language matters too. Google states that English is fully supported in Veo 3.1 and other languages have not been evaluated, while Kling lists Chinese, English, Japanese, Korean and Spanish. For Russian speech we test before promising anything. Workarounds: frame wider, keep lines short, use voiceover over action, or use a dedicated AI presenter pipeline built around a consenting real person. Our guide to sound design for AI video compares native audio, AI voices and voice actors.
9. Direction by precise camera instruction
“Slow push in” works. “Start at knee height, crane up over six seconds, rack focus to the bottle on the fourth second, end on a 35 mm frame” does not. Models follow camera language loosely and often trade one instruction for another. Studios get control by designing the first frame, and sometimes the last frame, as still images, then generating several takes and choosing. If a shot has to match a storyboard frame for frame, plan for many attempts, 3D previsualisation, or a camera.
10. Legal and ethical limits on real people’s likeness
This limit is not technical and will not be fixed by the next model. In the UK, section 6 of the CAP Code urges marketers to get written permission before portraying a member of the public, referring to someone with a public profile or implying a personal endorsement. TikTok’s rules on AI-generated content prohibit AI content showing public figures making endorsements and the likeness of private adults used without permission, and require labels on realistic AI content. The EU AI Act’s transparency rules in Article 50, which the European Commission says apply from 2 August 2026, require deepfakes to be disclosed. In practice: no celebrities, no lookalikes, no real voices without signed consent. This is general information, not legal advice.
The ten limits at a glance
| Limit | Studio workaround | Film it when |
|---|---|---|
| Exact product, readable text | Real photos as the base, text added in the edit | The label must be read in motion |
| Hands, fine interactions | Cut around contact, real hand inserts | Technique is the point |
| Long takes | Short shots, edited | One unbroken take matters |
| Same character in many shots | Character sheet, stills first, strict checking | A real person exists |
| Choreography, physics | Reference video, slow motion | Experts will watch |
| Real people and places | None | Almost always |
| Exact brand colour | Shot-by-shot colour correction | Rarely needed |
| Close-up lip-sync | Wider framing, voiceover | Long speech to camera |
| Exact camera moves | First and last frames, many takes | Frame-accurate storyboard |
| Real likeness | Written consent or an invented character | Consent is missing: do not make it |
How do you decide whether an idea suits AI?
At Zero Shot, an AI video studio in London, we ask four questions before quoting.
- Does anything in the frame have to be exactly true? A product, a place, a person, a result. If yes, that element comes from a camera or a photograph, and AI builds the rest around it.
- Can the idea be told in shots of a few seconds? If yes, generation length is no obstacle.
- Is the difficult part spectacle or precision? Spectacle, meaning impossible settings, scale, weather and fantasy, is where AI is strongest. Precision is where it is weakest.
- Would filming it be cheap? A person talking to a phone in your own shop costs almost nothing to film. Do not generate what you can shoot in an hour.
Most briefs end up mixed, and a studio that never says “film that part” deserves some suspicion. Our checklist for choosing an AI video production company covers that in more detail.
If you are not sure which side of the line your idea falls on, send us a short brief. We will tell you what we would generate, what we would film, and give you a fixed price.
Questions people ask
How long can an AI-generated video be?
A finished film can be any length, because it is edited from many short clips. A single generation is short: as of September 2026 Google documents 4, 6 or 8 seconds for Veo 3.1, Kling 3.0 offers 3 to 15 seconds, and ByteDance's Seedance 2.5 reaches 30 seconds.
Can AI video show my exact product?
Not reliably from a text description. Studios start from your own photographs or 3D files, generate around them, check every frame, and often composite the real packshot and label back in during the edit.
Can I put a celebrity or a real person in an AI video?
Not without their written permission. UK advertising rules urge marketers to get permission before portraying anyone or implying their endorsement, and TikTok bans AI content that shows public figures making endorsements. This is general information, not legal advice.
Will these limits disappear soon?
Some are shrinking fast: clip length, native sound and character references all improved during 2026. Exact control over products, text, colour and movement improves more slowly, and the limits around real people are legal, not technical, so they will stay.