Licensing Multimodal Records Assembled from Several Sources and Rightsholders

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

Why one dataset-level license rarely covers a joined record

A single grant is insufficient because the licensor usually owns the container, not every work and likeness inside it. An audit of widely used public datasets found that datasets assembled from several sources carry obligations that differ by component, so a top-level license can overstate what a commercial user may do [6]. The same failure shows up in enterprise records, where the system of record and the content it stores answer to different parties.

Typical joins that create layered rights:

• Call plus screen plus CRM. A Gong- or Zoom-style recording, its ASR transcript, the screen share showing a vendor's SaaS UI, and the Salesforce opportunity it links to. See multimodal meeting recordings.
• Robot log plus facility video. ROS bag trajectories owned by an operator, overhead cameras owned by the facility, and workers in frame. See who owns robot data.
• CAD plus customer specification. A supplier's STEP or SolidWorks files paired with a customer's drawing package and email thread.
• Support ticket plus attachments. End-user screenshots and PDFs inside a Zendesk or Jira record. See support tickets with screenshots.

For a single clip's layers (footage, people, music, brands), the video clip rights layers guide goes deeper; this page covers records joined across systems and owners. The multimodal training data hub lists related sourcing guides.

Who holds which rights in a multimodal record

Each component has its own rights holder, and the buyer's job is to name every one before price is discussed. Four layers recur across nearly every joined record.

1. System and data owner. The company that operates the recording platform, CRM, MES or robot fleet usually controls the export, but its vendor contracts may restrict data reuse.
2. Recorded people. Employees, customers, contractors and bystanders whose faces, voices or names appear. Recognizability is broader than a face: stock-platform policy treats voice, tattoos, clothing and even distinctive surroundings as grounds for requiring a release [1].
3. Embedded third-party works. Software UIs, slide decks, licensed stock imagery, background music and brand marks captured on screen or in frame. See screen recordings and third-party content and third-party content inside licensed corpora.
4. Customer-owned material. Specifications, contracts and files a supplier's own customers sent in. Whether the supplier may license these depends on its customer agreements, and a supplier that quietly expands its terms to permit AI training can draw scrutiny; FTC staff have warned that adopting more permissive data practices this way may be unfair or deceptive [12].

Releases and consents: what language actually covers AI training

A release covers AI training only if it says so, and older releases often do not. Some current model releases name artificial intelligence and machine-learning training as a covered use explicitly [2]; a 2015 marketing-video release that grants "advertising and promotional use" does not obviously extend to training a generative model.

Voice and face data add statutory layers. Illinois BIPA defines biometric identifiers to include voiceprints and scans of face geometry and requires written releases and a published retention schedule, while excluding photographs themselves [7]. Texas requires notice and consent before capturing a biometric identifier for a commercial purpose [8], and California Penal Code 632 requires all-party consent to record a confidential communication [9]. For the Illinois specifics, see the BIPA guide for AI data licensing.

When releases are missing or ambiguous, the practical choices are to drop the modality, redact it, or obtain new consent. Redaction across faces, voices and on-screen text is its own discipline; see de-identifying multimodal records.

Per-asset AI-use signals and how to check them on delivery

Per-asset signals exist for some media and should be read on every delivered file that carries them. The Creator Assertions Working Group defines a cawg.training-mining assertion that sits in a C2PA manifest and marks ai_training, ai_generative_training, ai_inference and data_mining as allowed, notAllowed or constrained [3]. C2PA has clarified that this assertion is referenced from the CAWG specification rather than defined in the core label set [4].

These signals express a preference; they do not technically prevent use, and their force depends on recipients honoring them [5]. For EU-facing general-purpose model providers, Article 53(1)(c) of the AI Act requires a copyright policy that identifies and complies with rights reservations under Article 4(3) of the DSM Directive [10], and the GPAI Code of Practice copyright chapter sets out how signatories implement that policy [11]. As of October 2026, Article 53 obligations apply and AI Office enforcement powers for new models began on 2 August 2026.

A delivery-time check should parse manifests with a C2PA validator (for example c2patool), log the assertion per asset ID, and quarantine files marked notAllowed or constrained until counsel reviews them. Absence of a manifest is not permission; it only means the rights question falls back to the license and releases.

Component rights map: the artifact to request before you sign

A component rights map is a table, one row per modality in the joined record, that states owner, people, embedded works, legal basis and permitted use. Ask suppliers to fill it in during diligence and attach it to the license schedule so that permitted uses and takedown handling are stated per modality.

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

Map every row to a stable join key (call ID, Opportunity ID, file hash) so that a withdrawal by one rightsholder can be executed without dropping the whole record. The record-level takedown guide covers the mechanics, and time synchronization checks explain why segment timestamps must survive redaction.

License terms that should change for multi-rightsholder records

The license schedule should describe records, modalities and uses at component level rather than "the dataset." Clauses that buyers commonly negotiate for joined records include:

• Per-modality use grants. Training, evaluation, retrieval and synthetic-data generation listed separately for each modality, because a screen share cleared for evaluation may not be cleared for generative training.
• Representation scope. Which rows of the rights map the supplier stands behind, and which (third-party UI, customer files) it explicitly does not.
• Withdrawal and takedown. The unit of withdrawal, the join key, and how derived artifacts such as transcripts and embeddings are handled.
• Downstream disclosure. Whether the buyer may describe the data in an EU training-content summary or a California AB 2013 disclosure.
• Model-provider routing. Whether licensed content may be sent to third-party model APIs for labeling or retrieval; see licensed content and third-party model API terms.

For a field-by-field view of general terms, read AI data license terms explained, and the dataset licensing glossary entry for vocabulary. The broader training data due diligence checklist covers non-rights diligence.

How SourceX approaches layered-rights requests

SourceX sources operational datasets from US companies on request and manages the commercial process, including licensing agreements and ongoing purchases. Every dataset is rights-reviewed for ownership and consents and delivered under a license that defines records, uses, term and delivery, with diligence materials on source, rights, preparation and allowed use prepared per dataset.

Personal details such as names, emails, phones and account numbers are removed or replaced before delivery, the method is recorded and a sample is checked, though no method is perfect. Every release is approved by the supplying company, delivery runs through private, access-controlled workflows only after an executed agreement, and nothing is contracted until a supplier agrees. Buyers can describe the multimodal records they need on the SourceX buyer page; a request does not guarantee a match.

Clearing multi-rightsholder multimodal data with SourceX

SourceX looks for US businesses that hold the data you describe, assesses data and licensing permissions, and agrees pricing and allowed uses in a license before any transfer. Prices are not published; terms are agreed per deal. Start by describing the joined records and modalities you need at sourcex.si/buyers.