Ted Entertainment, Inc. v. OpenAI, Inc.

Where the case stands

OpenAI moved to dismiss the single count with prejudice on 27 July 2026 (ECF 24). The plaintiffs opposed on 24 August (ECF 30) and OpenAI replied on 14 September (ECF 35). The motion was noticed for hearing before Judge Labson Freeman on 17 December 2026 (ECF 24 at 1). The sources reviewed show no ruling.

The motion targets the original class action complaint (ECF 1). OpenAI’s motion and reply both cite it as “CAC ECF 1” (ECF 24 at 1; ECF 35 at 1), and the opposition calls it the operative complaint (ECF 30 at 1). One footnote in the opposition refers to “the FAC” (ECF 30 at 23 n.6). We found no amended complaint in this case in the sources reviewed.

On 30 August, ChatGPT Is Eating the World reported that Judge Labson Freeman struck the plaintiffs’ opposition in this case (ECF 30) because its footnotes broke her standing order. By its account the footnotes were single-spaced, cited legal authority, and numbered six against a guideline of five. OpenAI’s reply cites the opposition as ECF 34 (ECF 35 at 1), which suggests a re-filed version. This page quotes ECF 30, the version we have. We have not read the strike order or ECF 34.

Ted Entertainment withdraws

On 29 October 2026, ChatGPT Is Eating the World reported that Ted Entertainment had voluntarily dismissed this suit without prejudice. By its account Ted Entertainment also dropped its suits against NVIDIA, Apple, Amazon, ByteDance and Meta, kept only its suit against Snap, and gave no reason. The post says Matt Fisher is now the sole plaintiff here. We have not seen the dismissal notice.

The complaint and every brief reviewed also name a third plaintiff, Golfholics, Inc. The sources reviewed do not show when or how Golfholics left this case. In the parallel ByteDance case, Judge Corley’s order records that Golfholics voluntarily dismissed its claim there after briefing (ByteDance ECF 55 at 2 n.2).

What the complaint alleges

The plaintiffs did not register their videos and do not sue for copyright infringement. They rely on the rule that registration is not needed to sue for circumvention of an access control (¶ 9). OpenAI notes the same choice (ECF 24 at 2). The theory runs in four steps.

1. YouTube lets people stream, not take the file. The complaint says YouTube keeps the underlying video files “locked away,” that even Premium downloads stay inside the app, and that its terms grant no rights to use content “independent of the Service” (¶¶ 3, 37, 44).
2. Five technical measures guard the files. A “rolling cipher” on the file URL, IP blocking and rate limiting, short-lived session-bound URLs, CAPTCHA challenges, and proof-of-origin tokens (¶¶ 49-62).
3. OpenAI trained Sora on YouTube-derived datasets. The link is an inference. OpenAI’s Sora System Card says Sora was trained partly on data “mostly collected from industry-standard machine learning datasets and web crawls.” The complaint says, on information and belief, that those included HD-VG-130M, HD-VILA-100M, Panda-70M and HowTo100M, which list YouTube video IDs and timestamps but hold no video files (¶¶ 67-68, 74, 93). It points to two open-source projects that tried to reproduce Sora using Panda-70M and HD-VG-130M (¶¶ 71-72). It also says OpenAI did its own scraping (¶ 98).
4. Getting the files required breaking the measures. On information and belief, OpenAI used yt-dlp “or a substantially similar” tool to defeat the cipher, and virtual machines or cloud infrastructure to rotate IP addresses when YouTube blocked one (¶¶ 99-106). It is said to have renewed session URLs, avoided CAPTCHAs and reused proof-of-origin tokens outside YouTube’s player (¶¶ 108-110).

The complaint treats each clip as a separate act of circumvention (¶¶ 74, 146). On that basis it puts HD-VG-130M at “roughly 130 million” circumvention events and HD-VILA-100M and HowTo100M at about 100 million each (¶¶ 78, 87, 90). It pleads willfulness (¶ 147). Statutory damages for circumvention under 17 U.S.C. § 1203(c)(3)(A) run from $200 to $2,500 per act.

The plaintiffs’ videos in the datasets

The complaint counts each plaintiff’s videos in each dataset. It does not say how many of the listings are the same video appearing in more than one dataset.

Ted Entertainment’s figures are IPLYR’s sums of the per-channel counts in ¶ 15. The specific videos are listed in Exhibits A, B and C (¶¶ 16, 19, 21), which are not in the copy we reviewed. For Fisher and Golfholics, the complaint says every listed video was “included in datasets used by Defendant to train its generative AI models” (¶¶ 19, 21).

OpenAI’s motion: a copy control, not an access control

OpenAI’s core argument is about the structure of § 1201. Subsection (a) bans circumventing measures that control access to a work. Subsection (b) addresses measures that protect copying and other exclusive rights, and bans only trafficking in tools, not the act of circumventing. OpenAI says Congress left that gap so that copying would stay governed by infringement law, including fair use (ECF 24 at 1, 8-10). Because anyone can watch the plaintiffs’ videos for free (¶ 35), OpenAI says the measures restrict downloading, not access, and the claim fails (ECF 24 at 11-14). It relies on MDY v. Blizzard, Lexmark and Hattler v. Ashton, a 2017 C.D. Cal. decision that dismissed a § 1201(a) claim over downloading of videos streamed on Vimeo (ECF 24 at 7-9).

It also says the plaintiffs are using § 1201 to avoid fair use. It cites a June 2026 ruling against Ted Entertainment on fair use in Ted Entertainment v. Saber, and Ted Entertainment’s own successful fair-use defense in Hosseinzadeh v. Klein (ECF 24 at 1-2). The plaintiffs answer that fair use is not a defense to circumvention and that registration is not an element (ECF 30 at 17-18).

OpenAI’s last point is timing: the complaint never says when YouTube put any of the five measures in place, so it does not show they existed when the downloads happened (ECF 24 at 19). The plaintiffs point to ¶ 139, which says YouTube applied the measures when each video was published, and say the request-level details are in OpenAI’s hands (ECF 30 at 24-25). OpenAI replies that deployment dates are facts about YouTube, a third party, not about OpenAI (ECF 35 at 10-11).

In reply, OpenAI adds that an access control needs a lock and an authorized key that screens some requesters out. It says the complaint describes “an automated handshake” that every visitor passes (ECF 35 at 4-5). It argues that VidAngel is different because DVD encryption leaves no unencrypted route to the film, while YouTube streams to everyone (ECF 35 at 3). It asks for dismissal with prejudice: “No amendment can change the fact that anyone can watch Plaintiffs’ videos on YouTube” (ECF 35 at 11). The plaintiffs ask for leave to amend if any part is found deficient (ECF 30 at 25).

How the sister cases have gone

Ted Entertainment filed parallel suits against other AI developers. OpenAI counts complaints in four jurisdictions against eight companies (ECF 24 at 2). Three courts had ruled on the same five-measure theory by mid-October, and none dismissed it. Each ruled on the pleadings only.

• Ted Entertainment v. Snap (C.D. Cal., 19 Aug 2026). As described in the opposition, the court denied dismissal and held all five measures plausibly alleged as access controls. It said “[n]othing in § 1201(a) requires an access control to take the form of a paywall, password prompt, or other measure requiring affirmative user input” (ECF 30 at 1, 10). OpenAI says Snap collapses the line MDY draws (ECF 35 at 7-8). We have not read the Snap order.
• Ted Entertainment v. ByteDance (N.D. Cal., 2 Oct 2026). Judge Jacqueline Scott Corley denied ByteDance’s motion against an amended complaint (ByteDance ECF 55 at 1, 12). She held the plaintiffs had standing on a lost-advertising-revenue theory: retrievals that bypass the player generate no ad impressions (at 5). She held that copyright owners can sue under § 1203 even though YouTube, not they, runs the measures (at 6-8). She found that proof-of-origin tokens, as alleged, operate whenever anyone views a video (at 9-10). She found Hattler unpersuasive because it predates VidAngel (at 11). ChatGPT Is Eating the World reported the ruling.
• Ted Entertainment v. Apple (N.D. Cal., 9 Oct 2026). Judge Richard Seeborg denied Apple’s motion (Apple ECF 61 at 1, 12). He held that the question is whether a measure provides “some degree of control over access,” not whether it is a use control or an access control (at 5). He held that an access control does not need affirmative user action (at 7). He also held that the alleged use of yt-dlp to defeat the rolling cipher “alone” was enough to survive the motion (at 11). ChatGPT Is Eating the World reported the ruling.

These rulings bind no other judge, and the OpenAI motion is aimed at an original complaint, not an amended one. The opposition also cites two D. Mass. decisions involving Suno, and OpenAI says the court in UMG v. Suno “declined to decide whether yt-dlp circumvented the rolling cipher” (ECF 30 at 1; ECF 35 at 8). ChatGPT Is Eating the World reports that Ted Entertainment has since dismissed its ByteDance and Apple suits as well.