Broadcaster and streamer rules on generative AI in production: what they say about your tools and data
If you deliver to a broadcaster or streamer, its published AI rules decide whether you can use a cloud AI tool on your rushes and what you must tell the commissioner. This page collects the current public rules in one dated table, with the three questions every one of them asks: does the tool train on inputs, does it keep them, and is its use approved. It is written for producers and their compliance leads; it is not legal advice, and your commissioning agreement may say more.
| Broadcaster | Document and date | Approval | Data rule, in their words |
|---|---|---|---|
| Netflix | "Using Generative AI in Content Production", Partner Help Center, August 2025 (partner login; wording as reported by StreamTV Insider, 21 Aug 2025, and TV Technology, 27 Aug 2025) | If any guiding principle isn't met, escalate to your Netflix contact; written approval where output includes final deliverables, talent likeness, personal data or third-party IP | "The generative tools used do not store, reuse, or train on production data inputs or outputs." "Where possible, generative tools are used in an enterprise-secured environment to safeguard inputs." Don't input unreleased assets without approval. |
| ITV | "Use of AI in ITV Commissioned Content", v2, September 2026 | Low-risk uses (transcription, logging, post) need no pre-approval but must be "notified to ITV as part of the final deliverables"; medium-risk uses need the Commissioning Executive and Compliance Advisor before filming | Approved tools list (Gemini, Gemini Notebook, ChatGPT, Runway, ElevenLabs, Adobe Firefly, Quickture, GitHub Copilot). "We stick to vetted tools that we have approved so that we can protect our data." Medium risk: "any use of AI which involves personal data (such as names, faces, voices, images, or recordings)". |
| Channel 4 | "AI Editorial Guidelines for Commissioned Content" (undated; principles PDF May 2025) | "All use of AI in Commissioning needs to be discussed and approved with a Channel 4 Commissioning Editor in advance"; tools documented "including purpose, scope, and human oversight involved" | "No uploading of confidential, personal, or rights-sensitive information into external systems." |
| Sky | "SKY Framework on using AI in Commissioned Programming" (undated PDF) | "SKY must provide written approval before any moderate or high-risk use cases are implemented." Keep "clear auditable records of AI use." | "You must only use secured, self-vetted and 'enterprise' graded tools." "Never inputting scripts, personal data or any unreleased/owned materials into any public AI." Confirm the tool "will not reuse, retain, or train on production materials/assets or Sky data." |
| BBC | Editorial Guidelines, "Guidance: The use of Artificial Intelligence" (undated, in force 2026) | Indies: "A senior editorial figure … should be responsible and accountable for its use of AI"; raise material uses at commissioning. Staff tools are authorised through BBC procurement (list on the intranet). | "any information input into an AI may be used by the developer of that AI to train it further or be shared with third parties. This may raise significant data protection or information security concerns". |
| TV 2 Danmark | "TV 2s eksterne produktioner og kunstig intelligens", 4 Dec 2024 | Written approval and declaration for flagged categories | "AI-modeller må ikke trænes med TV 2s data." Keep prompts and versions as part of the raw material. |
| SRG SSR (Switzerland) | Regulations on the use of AI, in force 1 June 2026 (staff) | Procurement process for tools | "Confidential or classified internal information must not be entered into such tools." |
| ABC (Australia) | "Using AI tools in ABC content", revised 16 Oct 2025 | "Only AI tools that have been approved for use should be used in editorial content." Commissioned third parties bound by the policies. | Adhere to legal and security obligations; audiences generally informed of AI use. |
| SBS (Australia) | "Guiding Principles for Use of AI", v2, 17 Apr 2026 | Senior editorial approval for significant AI use; works with external producers on compliance | Protect "personal, sensitive and/or confidential audience and corporate data". |
| Disney, National Geographic, Apple TV+, Amazon MGM, Warner Bros. Discovery, Paramount, NBCUniversal | No public partner-facing AI rules found (Oct 2026) | Contract, security questionnaire, TPN | See the MPA best practices below. |
What this means for a metadata tool like Shotlore
- Free AI keys fail every rule above. Google's free Gemini tier uses content to improve products and allows human review; Shotlore's guidance is a paid key only, in the production's own Google Cloud project.
- Paid keys pass "no training" and give you a processor contract, but Google keeps prompts and responses for 55 days for abuse monitoring, so "do not store" (Netflix) and "stored or cached externally" (Sky) need stating to the commissioner. The details, with Google's wording, are on security and confidentiality.
- Faces and voices are the approval trigger at ITV, Netflix and Sky. Frames with contributors, and transcripts, fall in that bucket; plan the disclosure before the shoot.
- Channel 4's wording is the strictest. Use Shotlore's on-device keywords (Mac and Windows; nothing leaves the computer) unless the Commissioning Editor agrees otherwise.
- Everyone wants a record. Sky asks for auditable records; TV 2 asks you to keep prompts; the MPA best practices ask for an audit log of AI actions. Shotlore's per-job send log (next update) is built for that.
The industry baseline: MPA Content Security Best Practices v5.3.2
The Motion Picture Association's best practices, which the Trusted Partner Network assesses vendors against, added an AI Supplemental on 23 September 2026. It asks the production company to "obtain client approval for AI/ML application use" (OR-5.0), to "obtain written client approval before applying any AI tool or service to client content" and keep "an audit log of all AI actions including prompts submitted … outputs produced … and timestamp" (TS-7.2), to source "Application Configuration Guidelines provided by the licensor" and complete "a model card" for each AI application (TS-7.1), and to confirm from any third-party AI provider "that content cannot be used for model training" with "least retention or zero retention where possible" (TS-7.4).
Corrections and additions welcome at hello@shotlore.app; we'll keep the table current and dated. How Shotlore itself handles your material: security and confidentiality.