AI in production · Start here
You don't need to be technical. You do need to stop guessing.
No prior knowledge assumed: the most exposed jobs are entry tasks such as logging and transcribing, so we teach those first.
What are these tools?
A language model predicts the most plausible next words, learned from vast amounts of text. Image and video models do the same with pixels. Plausible is often correct, which is why the wrong answers are dangerous.
What can't they do?
- They don't know what's true. A model will state a rate, a law or a credit with total confidence and be wrong.
- They don't know what happened recently, unless the tool is searching, and even then check the source.
- They don't repeat themselves reliably. Ask twice, get two answers: a problem for anything that must match across shots or days.
- They don't understand consequence: that a rearranged schedule breaks turnaround, or a generated face belongs to a living person.
- They aren't private by default. What you paste in may be stored and, on some plans, used to improve the product.
Ten words you'll hear
| Word | What it means |
|---|---|
| Model | The trained system itself. The app you type into is a wrapper around one. |
| Prompt | What you ask for. The biggest lever you have over the result. |
| Context window | How much the model holds at once. Past it, the earliest part is silently dropped. |
| Token | The chunks text is broken into, roughly a short word each. Limits are counted in these. |
| Training data | What the model learned from. Who owned it, and whether they agreed, is the central legal argument. |
| Hallucination | A confident, fluent, wrong answer. Built into how models work; you can't switch it off. |
| Diffusion | The method behind most image and video generators: start from noise and refine towards the description. |
| Fine-tuning | Further training on a specific set of material, to make a general model behave a particular way. |
| On-device or open weights | A model that runs on your own machine, so your material doesn't go to somebody else's server. |
| Provenance | A record of how a file was made and whether a model touched it. C2PA Content Credentials is the standard people mean. |
What's in use
Department by department.
- Editorial. Transcription and searchable rushes, slate reading, auto-sync and first-pass selects. The most adopted use in UK post, and the biggest time saving.
- Sound. Dialogue isolation, noise and reverb removal, source separation and speech-to-text logging. Mostly reads what you recorded rather than inventing anything.
- VFX and finishing. Roto and paint assist, matchmove and depth, denoise and upscaling, and models trained on a shot's own frames. The most mature use.
- Production office. Script breakdown, schedule drafts, movement orders, reformatting. High value and high data-protection exposure.
- Design, look and camera. Previs, look development, reference and lookbooks before the shoot; reframing, relighting and face work after it. On the day, very little touches the camera beyond logging and metadata.
- Virtual production. Machine learning already sits inside the real-time pipeline. Generated backings hold for locked-off or slow moves and need baking before they go on a wall.
- Performers. Scanning, digital doubles, de-ageing and voice. The hardest consent questions in the industry: read consent and credit before anything is captured.
The worked examples are on on set and in post.
What are your settings already doing?
Most leaks come from a default, not a decision. No technical knowledge needed.
- Check whether your inputs train the model. On many tools it depends on plan type (consumer or business) and a setting, not on whether you pay. Check before you paste anything that isn't yours.
- Check whether your images are public. Some image tools publish what you make by default. Under a non-disclosure agreement, that default can itself be the breach.
- Check retention. How long the provider keeps what you uploaded, and whether you can delete it.
- Know the commissioner's rule. Netflix's guidance for production partners (August 2025) says tools used on a production shouldn't store, reuse or train on production inputs or outputs. Many consumer plans don't meet that bar.
Where to learn it properly
The ScreenSkills-accredited Responsible AI training from AIMICI starts with a level 1 course of about an hour, with no AI background needed (Pact, 15 June 2026). Our own course, AI in production, puts the same habits to work on a unit, and it's fully funded.
Last reviewed 9 November 2026