Licensing Unpublished and Raw Footage from Creators and Production Companies for AI Training

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Why unpublished creator footage is worth the extra diligence

Unpublished footage is worth the diligence because it can add visual material your model has probably not already seen, and it usually comes with the camera-original quality that generation models need. That novelty is a working hypothesis rather than a measured fact: material that never left a hard drive is simply less likely to have been crawled, but some of it may have been posted later under a different edit. Treat novelty as something to test with near-duplicate checks against your existing corpus, not something to pay a premium for on faith.

The practical advantages are concrete. Camera originals in ProRes, DNxHR, BRAW or high-bitrate H.264/H.265 keep detail that platform re-encodes destroy, and full-length takes give long, uncut temporal context with natural camera motion. Outtakes and alternate takes are useful for text-to-video training because they show the same scene with controlled variation.

The trade-off is fragmentation. Instead of one stock library with one contributor agreement, you face hundreds of owners, each with different contracts, release habits and folder structures. The rest of this guide is about controlling that fragmentation. For the broader map of sources, start at the video data hub.

Who actually owns a creator's raw footage

Ownership of raw footage often does not sit with the person who shot it, so chain of title has to be established clip by clip. Under US law copyright vests in the author, but for a work made for hire the employer or other person for whom the work was prepared is considered the author, and ownership can be transferred in whole or in part [1]. A freelance shooter who delivered a brand film may hold the drive while the client holds the copyright, or the reverse.

Four ownership patterns recur when buying from small owners:

• Self-initiated work. Personal B-roll, travel and nature footage shot on spec. The creator is usually the owner, subject to releases.
• Client commissions. Corporate, wedding, event and branded content. Production contracts often assign footage, including unused takes, to the client, or grant the client exclusive rights. Ask for the contract, not a summary of it.
• Employee work. Footage shot by staff of a production company within the scope of their employment belongs to the company, not the camera operator.
• Co-productions and agency work. Rights may be split, or held by an agency or network that commissioned the series.

The client-ownership risk is the one most often missed (a hypothesis drawn from how commissioning contracts are typically written). Many commissioning agreements cover "all materials created in connection with the project," which reaches outtakes, not just the delivered cut. If a supplier cannot show the governing contract, or the project file is labeled with a client name, keep that clip out of the licensed set until ownership is resolved.

Production companies with feature or episodic material raise a different set of questions, including union and talent agreements and plate ownership. Those are covered in film and TV production footage licensing. Owner-side questions are discussed in whether film and post-production studios can sell data to AI companies.

People, property and other layers inside each clip

Owning the footage does not clear what appears in it, so each clip needs a separate check of people, property, music and on-screen content. Stock platform guidance treats a person as recognizable not only by face but by features such as voice, tattoos, clothing or distinctive surroundings, and requires a model release in those cases [2]. Raw creator footage is often shot run-and-gun, which means bystanders, clients' staff and family members with no release on file.

The layers to log per clip:

• Recognizable people. Release on file, release scope (does it cover uses beyond the original project?), minors present.
• Private property and interiors. Property releases for identifiable homes, venues and private businesses.
• Brands, artwork and screens. Logos, murals, posters and monitors showing third-party content.
• Audio. Location audio can capture conversations and licensed music playing in the background; production music laid on during editing is usually licensed only for the original project.
• Embedded metadata. Camera files and proxies carry timestamps, GPS coordinates, device serials and sometimes operator names. An audit of a large web-scraped image dataset found Exif tags with timestamps, geolocation and personal names, and noted that the dataset's download tool extracts and stores those tags [8]. Video sidecars and QuickTime atoms raise the same issue.

A fuller breakdown of how these layers interact sits in rights layers in a video clip. Where footage shows employees at work, see recording workers on video for AI datasets.

Signals to check at intake: do-not-train flags and machine-readable licenses

Check every clip for prior licensing signals at intake, because footage that was ever published may already carry terms or opt-outs that conflict with your license. Content credentials can attach a creator's "do not train" preference directly to media files [4], and a creator who set that flag on a published edit may not remember doing so when offering the raw takes. Machine-readable licensing standards such as Really Simple Licensing now let publishers state AI licensing terms for online content, including video [3].

If you train or supply general-purpose models with an EU nexus, these signals matter beyond good practice. Article 53(1)(c) of the EU AI Act requires GPAI providers to maintain a policy to comply with Union copyright law, including identifying and honoring rights reservations under Article 4(3) of the DSM Directive; these obligations have applied since 2 August 2025, with AI Office enforcement for new models from 2 August 2026 and for existing models from 2 August 2027 [5]. As of October 2026, a direct license from the owner is the cleanest answer, but your intake record should still show that you looked for reservations and what you found.

Practical intake checks:

• Read C2PA manifests and XMP on any file that has them, and record any training-related assertion.
• Ask whether any edit from the same shoot was published, where, and under what platform or stock terms.
• Search the supplier's site for RSL or robots-based AI terms, and keep a snapshot.
• Record conflicts rather than silently dropping clips, so you can explain exclusions later.

Aggregating footage from many small owners

Aggregation works only if every supplier delivers into one schema and signs one license structure, because inconsistent metadata turns into unprovable rights at scale. Dataset-level audits show how quickly this goes wrong: an audit of more than 1,800 text datasets found license information missing for more than 70% of datasets on popular hosting sites and license error rates above 50% [7]. The fix is to capture rights at the clip level, at ingest, and never reconstruct it later.

Agree three things before any footage moves:

1. A per-owner agreement. Each owner signs the same core terms covering the clips listed in a schedule, the permitted uses, term and delivery. Avoid side letters that create one-off exceptions you cannot track.
2. A clip manifest. One row per clip with a stable ID, checksum, owner, rights basis and release references. Croissant, a schema.org-based JSON-LD format for ML datasets, is a reasonable container for describing files and record structure [6].
3. A payment and audit trail. Map payments to owners and to the clip IDs they cover, so a later takedown or dispute can be traced to exact files.

Expect operational friction with small suppliers: inconsistent folder naming, proxies mixed with camera originals, mislabeled frame rates and dead drives. Budget time for ingest QA, and run shot detection, deduplication and quality scoring as described in curating raw video into training clips.

Clip-level rights manifest

A clip-level manifest is the single artifact that makes many-owner footage defensible, because it ties each file to its owner, rights basis and checks. The record below shows the fields worth requiring from every supplier.

Illustrative example: invented to show structure; it does not describe an available dataset.

{
  "clip_id": "own042-shoot2019-0137",
  "sha256": "9f2c...e41a",
  "owner_id": "own042",
  "owner_type": "production_company",
  "rights_basis": "self_initiated",
  "client_contract_ref": null,
  "capture": {
    "codec": "ProRes 422 HQ",
    "resolution": "3840x2160",
    "fps": 23.976,
    "duration_s": 48.2,
    "audio": "location_stereo"
  },
  "people": { "recognizable": 2, "minors": false, "release_refs": ["MR-0412", "MR-0413"] },
  "property": { "identifiable": true, "release_refs": ["PR-0091"] },
  "third_party_content": ["background_music_unknown"],
  "previously_published": { "status": true, "where": "creator_site", "ai_terms_found": "none" },
  "do_not_train_flag": false,
  "metadata_scrubbed": ["gps", "device_serial", "operator_name"],
  "decision": "include_video_mute_audio"
}

Note the decision field: the background music means the clip is licensed with the audio track removed, not dropped. Recording partial decisions this way keeps usable footage in the set.

Due diligence checklist for raw footage purchases

Run this checklist per supplier, then sample clips against it, before signing a license for creator or production-company footage.

Illustrative example: invented to show structure; it does not describe an available dataset.

The AI training data due diligence checklist covers general items that apply to any category, such as data provenance documentation and security review. Compare creator footage with the narrower terms covered in stock footage licensing for AI training before deciding which channel fits each part of your requirement.

How SourceX approaches creator and production footage requests

SourceX sources operational datasets from US companies on request and manages the commercial process, including licensing agreements and ongoing purchases; nothing is held in stock, and a request does not guarantee a match. Buyers describe the data they need rather than naming businesses, and SourceX looks for US companies that hold it, with every release approved by the supplying company. SourceX does not source scraped web content or generic CCTV or photos.

Each dataset is rights-reviewed for ownership and consents and delivered under a license that defines records, uses, term and delivery. Diligence materials covering source, rights, preparation and allowed use are prepared per dataset. Delivery runs through private, access-controlled workflows only after an executed agreement and supplier approval. If your lab needs footage from owners such as production companies, you can describe the requirement to the SourceX buyer team. Media buyers may also find the media and publishing buyer page useful.

Describe the unused footage you want to license

Tell SourceX what footage you need for training: subjects, capture formats, durations, people and audio requirements, and allowed uses. SourceX looks for US companies that hold matching data, assesses data and licensing permissions, and nothing is contracted until a supplier agrees. Start a buyer request at sourcex.si/buyers.