You can type a sentence describing a scene and, a minute or two later, watch a short clip that didn't exist before, complete with camera movement, lighting and convincing physics. Not long ago, AI video was mostly a novelty, shared for its melting faces and impossible physics. Today, the best clips could pass for real footage, at least for a few seconds.
But video is also where AI's limits are easiest to spot. Hands do strange things. Objects appear and vanish between frames. People move in ways people can't. The clips are short, and the longer you ask the AI to hold a scene together, the more likely it is to drift into the uncanny.
Learning to notice those signs is useful for more than judging your own clips. The same technology is used to make deepfakes: video of real people saying and doing things they never did. So this Thing also looks at deepfakes from the maker's side: what makes AI video convincing, how to label what you make and why you should never make video of a real person without their permission.
In the activity, you'll create a few short clips of your own and look at them with a critical eye. Enjoy the creative side. Then keep that critical eye switched on, because you'll need it again in Thing 16, which is all about scams and checking what's real.
How AI video generation works
The basic process will feel familiar from image generation. You write a text prompt describing a scene, and the AI generates a short video clip. Some tools also accept a still image as a starting point, animating it into motion. This image-to-video approach often produces more predictable results because the AI has a visual reference to work from rather than building everything from text alone.
Under the surface, video generation models are trained on enormous datasets of video footage. They've learned patterns: how water flows, how fabric moves in wind, how a camera pans across a landscape. When you give them a prompt, they draw on those learned patterns to generate a sequence of frames that, ideally, look like a coherent piece of video.
"Ideally" is doing a lot of work in that sentence, though. Video is far more complex than a still image. An image generator produces one consistent frame. A video generator produces dozens of frames per second, and each one has to be consistent with the ones before and after it. Every object needs to move in a physically plausible way. Lighting needs to stay consistent. If there's a person in the scene, their face, body, and clothing need to look the same in every frame. This temporal consistency (maintaining coherence over time) is the central challenge of video generation, and it's why video AI lags behind image AI in reliability.
This is also why the clips are short. Most free-tier generations produce five to ten seconds of video. That's enough to showcase a mood, demonstrate a concept, or create a social media clip, but it's a long way from generating a full scene or a narrative sequence. Some paid tiers allow extensions up to several minutes, but quality tends to degrade with longer generation, and the costs climb steeply.
The current tool landscape
Video generation is the most expensive AI creative category, which means free tiers are more limited than what you've encountered with image or music tools. That said, several platforms offer enough free access to give you a proper feel for the technology.
Ethics and copyright
The ethical questions around AI video are the same ones you've met with images, voice and music, but amplified.
Copyright and training data. Like image and music generators, video models are trained on existing video content. The legal status of this training is still being worked out, and the same tensions between AI companies and content creators apply here.
Environmental cost. Video generation requires significantly more computing power than image generation, which translates to higher energy consumption. This isn't a reason to avoid using the tools, but it's worth being aware that generating dozens of video clips has a real environmental footprint. That's another reason to be thoughtful rather than indiscriminate about what you generate.
The biggest ethical question of all, deepfakes, gets its own section next.
Deepfakes from the creator's side
If you can make a convincing clip of a scene, you can make one of a person. That's why consent and labelling matter.
A deepfake is AI-generated or AI-altered video, audio or imagery that shows a real person saying or doing something they never said or did. Now that you've made clips yourself, it's worth looking at deepfakes from the maker's side.
What makes AI video convincing
The clips that fool people tend to play to AI's strengths. They're short. The camera is fairly still, the lighting is soft and the background is simple. There's little hand movement or handling of objects. Faces are often seen front-on, at phone-video quality, where small flaws are hard to see. And a cloned voice laid over the top (Thing 8) does a lot of the persuading.
The tell-tale signs
The flaws you'll look for in the activity are the same ones that can give a deepfake away: hands and fingers that warp, teeth or jewellery that shift, blinking or lip movements that don't quite match the words, skin that looks too smooth, lighting that doesn't match the room and text that turns to nonsense. These signs get harder to spot with every new model, though, so treat them as clues, not proof. Thing 16 covers how to check something properly.
Consent comes first
The simplest rule is never to make video, or a voice, of a real person without their clear permission. That includes friends and family "just for a laugh", colleagues and public figures. Reputable tools ban it in their terms, and some kinds of deepfake are now crimes in the UK. Even a harmless joke can be screenshotted, shared and taken out of context.
Label what you make
If you share AI video, say so in the caption. EU rules that began to apply in August 2026 require AI-generated content to carry a machine-readable marking and deepfakes to be clearly labelled. They don't apply directly in the UK, but labelling is good practice wherever you are. Many tools also add a visible watermark, an invisible one (such as Google's SynthID) or Content Credentials, a tamper-evident record of how a file was made. Don't try to remove them. They're there so people can tell what's real.
Resources to explore
Your main tool for this activity. A daily allowance of free credits, with watermarked output. Text-to-video and image-to-video. Run by a Chinese company, so use made-up prompts only. Web-based.
Limited one-time free credits (125). Industry standard for professional AI video. Web-based. A good second tool for comparison.
Video models from several companies in one place. Paid only: the free plan has no monthly credits.
Paid plans only. Strong physics simulation and high-fidelity output.
A continuously updated leaderboard where people vote on clips from the major video models side by side. Useful for seeing which tools are currently leading.
A free checker from the Content Authenticity Initiative. Upload an image, video or audio file to see whether it carries Content Credentials showing how it was made or edited, including whether AI was involved.
Activity: your first AI videos
You're going to generate AI video clips using at least two different approaches, evaluate the results critically, and reflect on where this technology is useful today and where it falls short. Think of this as the video version of the image generation exercise from Thing 9: you're learning by making, comparing, and evaluating.
- Set up your account. Create a free account at kling.ai. If you'd like to try a second tool for comparison (recommended if you have the patience for two sign-ups), create a free account at runway.com as well. Runway gives you a one-time allowance of free credits rather than a daily one, so save it for the comparison step.
- Start with image-to-video. Find or create a still image to use as your starting point. Use one of the AI-generated images you created in Thing 9, or generate a new one from a made-up description. Don't use personal photos or images of real people. Upload it to Kling and add a short motion prompt describing how you'd like the scene to come alive.
- Try text-to-video. Generate a clip entirely from a text description. Write two different prompts: one naturalistic, one creative or abstract. If you run out of credits, Kling's free allowance tops up daily, so you can come back and finish another day.
- Evaluate what you've made. Watch each clip several times and assess the results against the evaluation criteria below.
- Compare tools (optional but recommended). If you signed up for a second tool, run the same prompt or upload the same image and compare the results.
Your output
A document containing:
- The prompts (and source images, if used) for each video you generated
- The video clips themselves (downloaded or screen-recorded), or screenshots from the clips if downloading isn't available on your tier
- Your evaluation of each clip, covering at least the points listed in the evaluation criteria above
- A short reflection (a few paragraphs) on your overall impression: what surprised you, what disappointed you, where you can see AI video being useful in its current state, and what would need to improve before you'd use it for something that mattered
Why this matters
Video generation rounds out your tour of AI's creative capabilities. You've now experienced AI working across text, images, speech, audio, music, and video, and the pattern is consistent: AI produces impressive results quickly, but the quality varies and human judgement is essential for evaluating the output.
More importantly, video is where the pace of improvement is fastest. The limitations you notice today are likely to be significantly reduced within months. By forming a clear, honest assessment of where things stand now, you're creating a baseline you can measure future progress against. That habit of critical evaluation is one of the most useful things you can take from this programme. And the critical eye you've practised here is the one you'll use in Thing 16 to judge whether a clip of a real person can be trusted.
Claim your Open Badge
Submit your prompts, generated video clips (or screenshots), evaluation notes, and written reflection as evidence for your Thing 11 badge via cred.scot.
Submit your prompts, video clips or screenshots, evaluation notes, and reflection as evidence to claim this badge via cred.scot.
Claim now