YouTubers Say TikTok's Parent Scraped Millions of Their Videos to Train AI. A Judge Just Let the Case Proceed.

On October 2, a federal judge in California refused to throw out a lawsuit with implications for every video on the internet. The case was brought by the company behind the h3h3 Productions YouTube channel and a golf creator known as MrShortGame Golf, and it accuses ByteDance, the parent company of TikTok, of bypassing YouTube's anti-scraping defenses to harvest millions of videos for training its text-to-video AI models.

The ruling from U.S. District Judge Jacqueline Scott Corley does not decide whether ByteDance did anything wrong. But it keeps alive one of the most consequential legal theories in the AI copyright wars: that scraping a platform against its technical guardrails can itself violate federal law, even when the videos were free for anyone to watch.

What the creators allege

The lawsuit, Ted Entertainment, Inc. v. ByteDance Inc. (case number 3:25-cv-10933 in the Northern District of California), was filed on December 23, 2025. Ted Entertainment runs the h3h3 Productions and H3 Podcast Highlights channels; Matt Fisher posts golf instruction under the name MrShortGame Golf. The two are suing on behalf of a proposed nationwide class of YouTube creators.

The complaint describes a deliberate pipeline. Rather than license footage or use YouTube's official APIs, ByteDance allegedly defeated five of YouTube's technological protection measures: an obfuscated signature scheme known as a rolling cipher, IP-based blocking and rate limiting, short-lived streaming URLs that expire when a viewing session ends, CAPTCHA human-verification challenges, and proof-of-origin tokens that verify a request comes from an authorized player.

The harvested videos, the filing says, went into training generative video systems including MagicVideo and Seedance, ByteDance models that turn written prompts into synthetic video clips. The complaint points to large video datasets referenced by ByteDance researchers (HD-VILA-100M, Panda-70M, and HowTo100M), which index references to millions of videos rather than holding the files, and quotes employees describing a practice of downloading videos with transcripts and audio from YouTube. Among the tools named in the filing is yt-dlp, the open-source video downloader.

Publicly viewable is not the same as freely copyable. A court just treated YouTube's anti-scraping machinery as an access control under federal law.

Why the case is not about copying

Here is the twist that makes the case a test for the whole industry: the plaintiffs did not sue for copyright infringement. Under U.S. law, an infringement claim generally requires the work to be registered with the Copyright Office, and most YouTube videos never are. That registration gap would have doomed a conventional infringement suit before it started.

Instead, the creators invoked the Digital Millennium Copyright Act's anti-circumvention rules, section 1201(a), which makes it unlawful to bypass a technological measure that controls access to a copyrighted work. The theory targets how the data was taken, not what the model eventually produced. It sidesteps the fair-use debate entirely.

ByteDance pushed back on two fronts. It argued that YouTube's measures govern downloading, not access, because the videos are free to watch in a browser; viewers never had to beat a paywall, a password, or encryption. It also challenged whether the creators had standing to sue at all.

Judge Corley rejected both arguments. She pointed to the proof-of-origin token, which the complaint says is generated during an active playback session and refused when missing, meaning even ordinary viewing flows through a process the platform controls. She relied on the Ninth Circuit's 2017 decision in Disney v. VidAngel, which held that a single measure can function as both an access control and a copy control, and that a defendant can still circumvent an access control even when authorized ways of reaching the work exist. At a hearing in late September, Corley had already signaled where she was headed, asking ByteDance's lawyers how the alleged conduct was not circumventing, at least at the pleading stage.

On standing, she accepted a concrete pocketbook injury: the creators plausibly lost per-view advertising revenue and YouTube Premium revenue, plus the views and watch hours that feed YouTube's recommendation engine and Partner Program eligibility. She found that injury traceable to the alleged circumvention.

A wave of suits, not a one-off

This is not an isolated fight. In August 2026, Judge Andre Birotte Jr. of the Central District of California let Ted Entertainment and Fisher bring nearly identical DMCA claims against Snap, ruling that YouTube's barriers may be modest, perhaps by design given YouTube's business model, but they are barriers nonetheless. Ted Entertainment has reportedly filed similar suits against Amazon, Apple, Meta, and OpenAI, all pending in California and Washington state. Meta and Nvidia were hit with parallel class actions from YouTubers over AI-training scraping earlier in the year.

The pattern is unmistakable. Creators, not record labels or film studios, are becoming the enforcement arm of AI data sourcing. Where book publishers and news outlets fought over text corpora, video creators are testing whether the DMCA offers a cleaner weapon: one focused on the taking, not the output.

What it means for the creator economy

For the millions of people who earn a living on YouTube, the case touches the central anxiety of the AI era. Creators upload hundreds of hours of footage to YouTube every minute, work that funds their livelihoods through advertising and sponsorships. If AI companies can vacuum up that footage at scale, defeat the platform's defenses, and train synthetic-video generators that compete for the same audiences, the pipeline sustaining the creator economy starts to leak from both ends: the work is taken, and the machine that replaces it is trained on the taking.

That is why the standing ruling matters beyond this one case. By recognizing lost views, watch hours, and recommendation placement as real, compensable injuries, the court treated a creator's presence on a platform as an economic asset that cannot be freely extracted. The order does not say ByteDance circumvented anything, and it does not decide whether training on the videos infringes copyright. Those questions now move to discovery.

But the practical message to AI builders is already legible. Models trained on data pulled through official APIs and licensed corpora sit outside this legal theory. Datasets assembled by defeating a platform's locks, even if the content was publicly watchable, carry a risk that no training label can hide.

What happens next

The case now enters discovery, where both sides will probe what ByteDance actually bypassed and what it actually took. Class certification will determine whether the suit represents a handful of channels or a nationwide class of creators. Settlements in similar disputes have been common: the industry would much rather negotiate licensing terms than litigate the boundaries of the DMCA in public.

Whatever the outcome, the direction of travel is set. The first wave of AI copyright cases asked whether models may learn from text scraped from the web. This case asks a sharper question: whether the locks on a platform's front door matter. For now, at least in one California courtroom, they do.