---
title: "Rights-cleared vs. scraped video: the legal exposure in training data"
description: Scraped video brings copyright, anti-circumvention and privacy risk into model training. What legal and procurement teams should ask a data source.
---

[Insights | Syntropi](https://insights.syntropiai.com)

# [Rights-cleared vs. scraped video: the legal exposure in training data](https://insights.syntropiai.com/rights-cleared-vs-scraped-video)

 Written by [Syntropi](https://insights.syntropiai.com/author/syntropi) | Nov 2, 2026, 7:12:37 AM

Video is now a core training input for frontier models, from video generation to world models to agents that need to understand physical activity. Much of the video that has historically been easiest to obtain was never licensed for that purpose. It was collected from public platforms by crawlers, often in bulk and often in ways the hosting platform's terms do not permit.

For a long time that looked like a technical question. It increasingly looks like a legal and commercial one. This post sets out where the exposure comes from, what has changed recently, and what to ask of any video data source.

Where the exposure comes from

Scraped video can carry several distinct kinds of risk, and it helps to separate them, because they arise under different laws and are addressed by different controls.

**Copyright.** Most video on public platforms is protected by copyright held by its creator or a rights holder. Whether training on it without permission is lawful depends on the jurisdiction and on doctrines such as fair use in the United States or text-and-data-mining exceptions elsewhere, and courts are still working through how those apply to model training. Uncertainty is itself a cost: it is hard to price a dataset whose legal status may be decided by litigation after your model has shipped.

**Circumvention and terms of service.** Separately from copyright, collecting video may involve getting around technical access controls a platform uses, or breaching its terms of service. In the United States, the Digital Millennium Copyright Act has an anti-circumvention provision that is distinct from infringement and does not depend on whether the use itself would be fair.

**Privacy and personal data.** Video of real people contains faces, voices, homes, and sometimes location and other identifying information. In jurisdictions with data protection laws, such as those modeled on the GDPR, processing that information typically needs a lawful basis, and people captured incidentally in scraped footage have not consented to anything. Biometric privacy laws in some US states add further obligations.

**Provenance and documentation.** Even where a specific use might ultimately be defensible, a dataset without clear provenance makes it difficult to answer questions from customers, regulators or auditors about what a model was trained on. Emerging AI regulation, notably in the EU, puts increasing weight on training-data transparency and documentation.

## What has changed

Video-specific litigation has moved from hypothetical to active. In 2024 and 2025, YouTube creators filed proposed class actions in US federal court alleging that AI developers collected YouTube videos for model training without permission. One such suit against Nvidia was voluntarily dismissed without prejudice in March 2025. Later in 2025, a group of creators filed a further proposed class action against Nvidia in the Northern District of California concerning video allegedly used to train a world model, followed in December 2025 by similar suits against Meta and ByteDance. Reporting on the later complaints indicates they rely on the DMCA's anti-circumvention provision rather than copyright infringement claims.

These cases are allegations, and we draw no conclusions about their merits. The relevant point for procurement is structural: plaintiffs are testing theories, such as anti-circumvention, that do not turn on the fair use questions many teams had assumed would decide the matter. A data strategy that relies on a single legal theory holding up carries more risk than one that does not need it.

## What "rights-cleared at the source" means

The alternative to scraping is to obtain rights before the video exists. At Syntropi this is how every recording in our library is produced, and the specific controls are worth stating plainly because they are what a legal or procurement reviewer should check for in any vendor.

Every contributor signs a contract before recording, so rights are established at creation rather than inferred afterward. We do not scrape. For each capture assignment we deliver rights documentation, so a customer can trace a delivered dataset back to the agreements that cover it. Our library is 100% rights-cleared on this basis, with contributors from around the world.

Safety is part of the same chain of custody. Every clip passes through an ML pipeline that checks ingestion quality, classifies the scene against the dataset specification, and runs safety review, including classifiers for child sexual abuse material and other illegal content, which are always on. Human review sits on top of the automated gates. Contributors typically submit two and a half to three times the volume that is ultimately accepted, which gives a sense of how much is filtered before delivery.

## Questions to ask any video data source

Whether you are evaluating Syntropi or anyone else, the same questions apply. We suggest legal and procurement teams ask them in writing.

How was the video obtained, and was any of it collected from third-party platforms? Who holds the copyright in each recording, and what license or assignment covers training and evaluation use? Did the people who appear in the video consent, and how is incidental capture of bystanders handled? Is personal data removed or minimized before delivery, and what personal data, if any, does the delivered dataset contain? Can the vendor provide per-dataset or per-assignment rights documentation that would satisfy an audit? What content safety screening is applied, and is it always on? How are takedown or withdrawal requests handled after delivery? Which jurisdictions were contributors in, and does the documentation account for their local law?

Vague answers to the first two questions are usually a signal in themselves.

## Why researchers should care too

Rights clearance can sound like a procurement concern that researchers can leave to others. In practice, it determines what research can be used. A model trained on data that later needs to be removed may have to be retrained. Evaluation results on a benchmark built from scraped content may be hard to publish or commercialize. And collection designed around consent tends to produce better data for long-horizon work: contributors who are recording deliberately can be asked to capture complete tasks, narrate their reasoning and keep going through mistakes, none of which scraping can provide.

## Closing

There is no way to make legal risk zero, and this post is not a substitute for advice from counsel who know your circumstances. But the difference between video licensed at the source and video collected without permission is real, documentable and increasingly consequential. It is worth knowing which one is in your training set.

If your team needs rights-cleared video along with the documentation to support review, you can [request sample data](https://syntropiai.com/#request).

 

*This post is general information for research, procurement and legal teams. It is not legal advice. The law in this area is unsettled and differs by jurisdiction; consult your own counsel about your specific situation.*

## Sources

- [Nvidia Faces Class Action Over Scraping YouTube to Train AI (Bloomberg Law)](https://news.bloomberglaw.com/ip-law/nvidia-faces-class-action-over-scraping-youtube-to-train-ai)
- [YouTube Creators Drop AI-Training Copyright Suit Against Nvidia (Bloomberg Law)](https://news.bloomberglaw.com/ip-law/youtube-creators-drop-ai-training-copyright-suit-against-nvidia)
- [Meta, ByteDance Hit With YouTubers' AI Copyright Scraping Suits (Bloomberg Government)](https://news.bgov.com/ip-law/meta-bytedance-hit-with-youtubers-ai-copyright-scraping-suits)

[View full post](https://insights.syntropiai.com/rights-cleared-vs-scraped-video)

```json
{
  "@context" : "http://schema.org",
  "@type" : "BlogPosting",
  "author" : {
    "@type" : "Person",
    "name" : "Syntropi"
  },
  "dateModified" : "2026-11-02T07:12:37.621Z",
  "datePublished" : "2026-11-02T07:12:37Z",
  "headline" : "Rights-cleared vs. scraped video: the legal exposure in training data",
  "image" : {
    "@type" : "ImageObject",
    "height" : 60,
    "url" : "/hs/hsstatic/content_shared_assets/static-1.4092/img/default-amp-logo.png",
    "width" : 60
  },
  "mainEntityOfPage" : "https://insights.syntropiai.com/rights-cleared-vs-scraped-video",
  "publisher" : {
    "@type" : "Organization",
    "logo" : {
      "@type" : "ImageObject",
      "height" : 60,
      "url" : "/hs/hsstatic/content_shared_assets/static-1.4092/img/default-amp-logo.png",
      "width" : 60
    },
    "name" : "Insights"
  }
}
```