- What's happening when a model makes a video
- The places the weirdness hides
- A real example from a client shoot that never happened
- Step by step: how I check a batch of AI clips before anyone else sees them
- The uncomfortable bit nobody selling you a tool wants to say
- What's improving, and what isn't
- What to do about it right now
- Frequently asked questions
- Official documentation
AI-generated videos look weird because the models predict what plausible pixels should look like, based on patterns in the video they were trained on, rather than simulating a real world with real objects in it. Small errors in hands, faces, liquids and shapes creep in and drift from one moment of the clip to the next, which is why fingers warp, blinks stutter and backgrounds quietly change.
Some of that improves with each new model release. Whether all of it can be fixed is still argued over, so the practical answer for anyone publishing today isn't waiting for better AI. It's knowing where the flaws hide, prompting around them and cutting the clips that fail.
What's happening when a model makes a video
An AI video tool isn't filming anything. It's generating a sequence of images that are statistically likely to follow one another, based on billions of frames of real video it was trained on. There's no camera, no set, no actor who understands that a coffee cup has to stay the same size when you pick it up. The model is predicting what pixels plausibly come next. A widely held view among people who work with these tools is that the models don't understand physics, gravity or object permanence the way a person does; they reproduce what those things usually look like, which works until a scene asks for something less common.
Motion is the other giveaway. AI clips often have a faint gliding quality, like the footage is floating slightly off from where your eye expects it, and it's worst on walking, pouring and anything that should have weight. Vendors don't explain how their models build motion between frames, so the cause isn't something you can look up, but the fix is the same either way: slower movement, shorter shots. You're not imagining that glide; it's one of the most common giveaways. For a wider look at generators like these, see AI tools for generating video content.
The places the weirdness hides
After running hundreds of generations for client work, the flaws cluster in the same few spots almost every time:
- Hands and fingers. Still the weakest point across every major model. Six fingers, fused knuckles, a thumb that vanishes mid-gesture.
- Teeth and eyes. Blinking that doesn't sync with speech, teeth that shift number or shape between frames.
- Object permanence. A logo on a mug that changes text halfway through a shot. A background chair that quietly disappears.
- Physics. Liquid pouring wrong, cloth that moves like it's underwater, shadows pointing two directions at once.
- Audio and lip sync. Mouths that move slightly out of time, or a voice that has that flat, slightly reverby quality even in a supposedly "quiet room."
If you want the deeper mechanics on why the sound side is often the giveaway before the visuals even are, I go into that in what makes AI videos sound weird and how to fix it, because audio artefacts are often easier to catch on a phone speaker than the visuals are on a laptop screen.
A real example from a client shoot that never happened
Last year I was helping a small coffee roastery client build a product teaser without booking a studio day. We generated 40 separate clips of "a hand pouring coffee beans into a grinder" using Runway. Out of 40, three were usable without obvious flaws. Two more were usable with careful cropping. The rest had at least one of: a hand with the wrong number of fingers, beans that fell upward for a frame, or a grinder that changed shape between the start and end of the clip.
That's 3 clean clips out of 40, or 7.5 percent, on a fairly simple, low-motion prompt. Count the two that worked after cropping and it's 5 out of 40, or 12.5 percent: 8 generations for every clip that made the final cut. Anything with faster movement, multiple people, or dialogue drops that rate further, sometimes to one usable clip in twenty. The client still saved money versus a studio shoot, but the time spent generating, reviewing, and discarding clips ate up almost as many hours as a short real shoot would have. Budget for that ratio in credits and review time before you promise anyone a deadline: if you need 6 usable clips at 8 generations each, plan for around 48 generations, not 6. To put that in real terms, take Google's Veo: Wikipedia's article on Veo says users can create a maximum of eight seconds per clip, and that Veo is sold through subscription tiers and Google "AI credits". So 48 generations at eight seconds each is 384 seconds, more than six minutes of generated footage, to end up with six clips and under a minute of finished video. Look up what one generation costs on your own plan and multiply it by 48 before you quote a price for the job.
Step by step: how I check a batch of AI clips before anyone else sees them
This is the exact process I run through before a client ever sees a generated clip:
- Watch it once at full speed, no sound. Your eye catches motion errors faster without dialogue distracting you.
- Watch it a second time at half speed on hands, mouths, and any text or logos. This is where most of the visible glitches show up.
- Check the first and last frame side by side. Anything that changed shape, colour, or position between the start and end of the clip is a red flag.
- Play the audio alone with your eyes closed. If it sounds flat or slightly robotic on its own, it'll sound worse once viewers are watching lips move too.
- Ask someone who hasn't seen the prompt to watch it cold. They'll spot the "something's off" feeling faster than you will, because you already know what to expect.
If a clip fails any of the first three checks, I don't try to fix it in the edit. I regenerate it. Trying to patch a warped hand in post costs more time than a fresh generation almost every time.
Here's the same process as a scorecard you can keep next to the edit. Stepping through a clip one frame at a time in your editor is the quickest way to catch a single bad frame.
| Check | What fails it | What to do |
|---|---|---|
| Full speed, no sound | Gliding, jerky or floating motion | Regenerate |
| Half speed, frame by frame on hands and mouths | Extra or fused fingers, teeth or eyes that change | Regenerate, or crop if the flaw sits at the edge of the frame |
| Text and logos | Letters that change or melt | Regenerate without text and add the text in the edit |
| First and last frame side by side | Objects that change shape, colour or position | Cut to the stable part of the clip, or regenerate |
| Audio alone | Flat, robotic or reverby voice | Replace with real voiceover or music |
| Cold viewer | "Something's off" with no clear reason | Cut it; if they feel it, your audience will |
The uncomfortable bit nobody selling you a tool wants to say
Is it fixable? The weirdness will keep shrinking, but don't plan on it disappearing in the tools you can use this year. Models like Kling, Runway and Google's Veo are trained to predict plausible pixels, not to simulate a physical world, so every release narrows the gap without closing it, and anyone who gives you a date for "fixed" is guessing. Treat the flaws as permanent for planning purposes and let any improvement be a bonus.
What you can plan around is what's dated and on the record. OpenAI's Sora was one of the best-known video generators, and Wikipedia's article on Sora records that the app was shut down on 26 April 2026, with the API discontinued in September 2026. A tool can disappear faster than its flaws get fixed, so don't build a content workflow that only works in one generator: keep your prompts, scripts and edit templates somewhere you can take to another tool. If you relied on Sora, Google's Veo, Runway and Kling are the obvious generators to move to; run the same five prompts through each, score the results with the checklist above, and pick on your own results rather than on demo reels, because every one of them shows its best clips on its home page. If you want a structured way to think about those risks, NIST's AI Risk Management Framework, listed below, is intended for voluntary use and is a sensible place to start.
What's improving, and what isn't
Hand rendering has visibly improved over recent model releases, though no vendor publishes an error rate for hands, so treat that as an observation rather than a measured figure. Camera movement and consistency across longer clips have also improved, and models handle a static shot or a slow pan far better than fast motion. What hasn't moved much: multi-person scenes with interaction (a handshake, two people passing an object), fine facial expression during speech, and anything involving text that needs to stay legible and consistent across frames.
If your content suits a slightly surreal, floaty quality, a mood piece or a music visual for example, working with that look can beat fighting it, and there's a good breakdown of why that dreamlike effect happens and when it works in your favour in why AI generated videos sometimes look like dreams.
What to do about it right now
If you're publishing AI generated video for a business, three practical habits cut the "weird" feeling down a lot:
- Set the clip length to the shortest option your generator offers before you hit generate, and build longer scenes by cutting several short shots together in your editor. Longer clips give the model more chance to drift, and a cut hides the join far better than a 10 second take that changes halfway through.
- Avoid close-ups on hands, mouths, and text unless you've generated enough variations to be picky.
- Add real audio (voiceover, music, sound effects) rather than relying on generated speech, since that's usually the fastest tell.
Two technical habits sit alongside those. First, generate at the highest resolution your plan allows instead of upscaling afterwards. An upscaler sharpens whatever is in the frame, including a fused finger or a melting edge, so a clean low-resolution clip upscaled is fine but a flawed one just becomes a sharper flaw. Second, export once. Bring the generated file straight into your editor, keep its original frame rate, and export the final edit a single time; every extra download, re-upload and re-export adds compression that makes AI texture look muddier.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
The biggest single win is in the prompt. Take the roastery shot. A vague prompt like "a hand pouring coffee beans into a grinder" asks the model to get fingers, a pouring motion and falling beans right all at once. A tighter one, "close-up of coffee beans falling into a steel grinder, fixed camera, no hands in frame", removes the hands entirely, which removes the weakest area, and the fixed camera gives the model less to keep consistent. You lose nothing the viewer needed: they still see beans and a grinder. Write prompts that leave out what the model is worst at, and add people with real footage when you need them.
One more step before you publish: disclosure. YouTube's help page on disclosing GenAI content says creators must disclose AI-generated or altered content that seems realistic, and that you make the disclosure during the upload process. Instagram and Facebook can add their own AI label to a post too, so for a business account the firm rule is simple: if a clip could pass for real footage, label it as AI yourself rather than waiting for the platform to.
If you're already editing the footage afterward, there's a proper walkthrough of the enhancement side, colour grading, stabilising, and layering fixes without making the final cut look more obviously fake, in how to enhance a video using AI tools without making it look fake. And if you're weighing up whether the visible flaws matter for where you're planning to publish, I've also covered how platforms treat this content in whether uploading AI videos to YouTube affects monetisation, because the rules there have shifted more than most creators realise.
For a full technical rundown on the visual and audio side together, my longer piece on why AI generated videos can look and sound off goes into the render pipeline in more depth than I've covered here.
If you're posting the finished clip to Instagram and it looks different once it's up, see instagram why are some videos brighter.
And because AI content is now labelled on some platforms, here's why Instagram posts show an AI info label now.
I run AI training and workshops for teams, and help businesses with AI enablement and workflow automation, so the busywork runs itself and people get time back. Most of that work is with SaaS marketing teams and small businesses. Here's how we can work together.
Frequently asked questions
Is Sora still available?
No. According to Wikipedia's article on the model, the Sora app was shut down on 26 April 2026 and the API was discontinued in September 2026.
Why do AI video hands still look wrong in 2026?
Hands have many small moving joints that change shape constantly, and the model has to predict each finger's position frame by frame across the whole clip, so a small error shows up as an extra or fused finger. That's widely observed rather than a figure any vendor publishes, and hands have improved with newer models without disappearing as a weak point.
Is it worth using AI video for a small business instead of filming real footage?
For simple, short, low-motion shots (product close-ups, background b-roll, social teasers) it can save real money and time. For anything involving hands, dialogue, or multiple people interacting, expect to generate far more clips than you keep, sometimes as few as one usable clip in every eight to twenty attempts.
How many clips should I generate to get one I can use?
In the roastery example above, 5 of 40 clips were usable on a simple, low-motion prompt, which works out at 8 generations for every clip kept. Faster movement, several people or dialogue usually push that ratio higher, so plan credits and review time around those ratios, not around one try per shot.
For a place to publish your own article, try AI blogs open to guest writers.
Official documentation
Run a business, a brand or a blog? I feature tools, products and expert guest posts for an audience of 15,000. See how it works