AI in video — categories of tools, workflow and responsibility
AI tools in video aren’t one thing. An upscaler that fills in detail and a model that generates a whole shot carry different risks and different obligations. The key is to know where AI is guessing — and to say so.
Five categories, five different questions
‘AI video’ covers very different techniques. In practice it helps to distinguish what the tool does to the recording: whether it only analyses it, alters existing pixels or creates new content.
| Category | What it does | Relationship to the original recording | Main risk |
|---|---|---|---|
| AI-assisted editing and organisation | Speech transcription, searching within shots, tagging objects, suggesting selects, syncing | Doesn’t change the recording, only analyses it | Transcription and tagging errors; confidentiality of material processed in the cloud |
| Enhancement | Noise reduction, sharpening, stabilisation, audio clean-up and dialogue isolation | Changes existing pixels and samples | A plastic, ‘waxy’ look on faces, artificial detail, artefacts in motion |
| Restoration | Removing scratches and dust, stabilising archive footage, colourisation | Fills in missing parts by estimation | Colourisation and filled-in detail are not historical fact |
| Upscaling and interpolation | Higher resolution, higher frame rate, slow motion | Invents detail and in-between frames that weren’t in the recording | Hallucinated textures, distortion with fast or overlapping motion |
| Generation | A new shot from text, an image or other video; face or voice replacement | Content is created without a corresponding recording | Misleading content, rights to likeness and voice, copyright, the duty to label |
Where AI fits in the workflow
Preparation and editing. The least risky and often the most useful: automatic transcription of interviews, searching for ‘where the budget is mentioned’, a rough assembly. The output is still your recording; AI just shortens the search. Check names, numbers and quotes — a transcription error easily ends up in the subtitles.
Rescuing material. Denoising an underexposed shot or separating dialogue from noise can save the only take you have. Work on a copy, compare with the original on a big screen and be careful with faces — smoothed skin and ‘painted-in’ eyes are the most common artefacts.
Upscaling and frame interpolation. With SD archive material, or a shot that has to be enlarged, an upscaler can give a better result than classic interpolation. But you need to know that the detail is the model’s estimate. In documentary, news or evidential material, that changes the meaning of the shot. Frame interpolation (e.g. from 25 to 50 fps, or slow motion) works with an estimate of the motion; it fails where objects overlap, with fine structures and with fast movement. If you need smooth slow motion, it is better to shoot at a higher frame rate.
Generation. Useful for previsualisation (storyboards, animatics), backgrounds, abstract visuals and concepts. In documentary and news production a generated shot is a fundamental intervention in credibility and needs clear labelling.
We don’t review specific commercial products here — their capabilities and terms change from month to month. Before using any tool, check the licence terms for its output, whether and where your data is processed, and whether it will be used to train the model.
What the EU AI regulation says
Under Article 113, Regulation (EU) 2024/1689 (the Artificial Intelligence Act) applies in principle from 2 August 2026. The transparency obligations in Article 50 were not postponed by the ‘Digital Omnibus’ package of amendments (which mainly postponed the rules for high-risk systems). Two paragraphs matter for video creators:
Article 50(2) — providers of AI systems that generate synthetic audio, image, video or text must ensure that the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. This doesn’t apply where the system only performs an assistive function for standard editing or doesn’t substantially alter the input. This obligation falls on the makers of tools, not on ordinary users.
Article 50(4) — deployers, i.e. those who use AI to generate or manipulate image, audio or video content constituting a deepfake, must disclose that the content has been artificially generated or manipulated. The regulation defines a deepfake as AI-generated or manipulated content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful.
Where the content forms part of an evidently artistic, creative, satirical, fictional or analogous work, the obligation is limited to disclosing the existence of such content in an appropriate manner that does not hamper the display or enjoyment of the work. Under paragraph 5 the information must be clear and distinguishable, provided at the latest at the time of the first exposure to the content, and in line with accessibility requirements.
This is a brief overview, not legal advice. For a specific project — especially advertising, political content, or using the likeness or voice of a real person — consult a lawyer. Besides the AI regulation, the rules on personality rights, personal data and copyright also apply.
Ethics beyond the obligations
- Consent. Only generate or alter the likeness and voice of a real person with their explicit consent — even if it is ‘just a test’.
- Truthfulness. In documentary and news, distinguish what was recorded from what was calculated. An upscaled archive shot is a different document from the original.
- Keep the originals. Always archive the unaltered source and record which AI steps were applied (tool, version, settings). It helps with corrections and with questions about provenance.
- Credit and transparency. If AI makes up a substantial part of a work, say so — in the credits, the description or accompanying text.
- Client data. Don’t upload a client’s material to cloud services without their consent and without knowing the processing terms.