Industries / Technology / Social media / Creator economy
Rights and credit
Rights and credit is who owns a creator's work, who may reuse it, and who is named and paid when it is copied, matched by software or used to train AI.
7
Milestones achieved, from 2007 to 2024.
2
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2007
Uploads matched to owners' files
On 15 October 2007 YouTube launched, in beta, Video Identification, the system now called Content ID. Rights holders supply reference files, and the system finds uploads that match them and lets each owner choose "whether to block, promote, or even—if a copyright holder chooses to license their content to appear on the site—monetize" them. YouTube said the tool went "above and beyond our legal responsibilities."
Credit
YouTube (product manager David King)
2015
Weigh fair use before takedown
In June 2007 YouTube removed Stephanie Lenz's 29-second home video of her young children dancing in the kitchen to Prince's "Let's Go Crazy", after Universal Music sent a takedown notice. On 14 September 2015 the US Court of Appeals for the Ninth Circuit held that the Digital Millennium Copyright Act (DMCA), the 1998 US law behind takedown notices, "requires copyright holders to consider fair use before sending a takedown notification". Fair use is the US rule that allows some unlicensed uses, such as commentary and criticism.
Credit
Stephanie Lenz; US Court of Appeals for the Ninth Circuit
2020
The takedown balance, tilted
Section 512 of the DMCA is a safe harbor: it shields platforms from liability for users' uploads if they remove infringing ones when notified. In May 2020, after a study that drew more than 92,000 written comments, the US Copyright Office reported that "Congress' original intended balance has been tilted askew." Rightsholders told it individual creators struggled to use the system and faced a "whack-a-mole" problem of removed uploads reappearing.
Credit
US Copyright Office
2022
Upload filters upheld, with limits
Article 17 of the EU's 2019 copyright directive makes video platforms directly liable for unlicensed uploads unless they meet set conditions, which in practice means reviewing uploads in advance, often with automatic recognition and filtering tools. Poland asked the EU's Court of Justice to annul it as a threat to free expression. On 26 April 2022 the court dismissed the case, holding that the article carries the needed safeguards: a filter that cannot tell lawful from unlawful content is not allowed, and users may upload their own content for purposes such as parody or pastiche.
Credit
Court of Justice of the European Union (case C-401/19)
2023
Beethoven claimed on two platforms
Adam Eric Berkowitz uploaded 102 new recordings of Beethoven's piano sonatas, music long in the public domain, to YouTube and Facebook. Content ID claimed 29 of them and Facebook's Rights Manager 29, only 12 of them the same; on YouTube the claims covered 2.67 of just over 10 hours of music, and Facebook muted 20 seconds of each claimed video. He published the test in January 2023 in Transactions of the International Society for Music Information Retrieval.
LessonBoth systems named the right composition every time, which shows that recognizing a piece of music is not the same as knowing who owns a given recording of it.
Credit
Adam Eric Berkowitz (Tampa-Hillsborough County Public Library)
2023
Short dance steps protected
Choreographer Kyle Hanagami posted his routine to Charlie Puth's "How Long" on YouTube in November 2017, where it drew nearly 36 million views by May 2022, and registered its copyright in 2021. He sued Epic Games, saying its Fortnite emote "It's Complicated", a dance players buy for their avatars for 500 V-Bucks of in-game currency, copied four counts of it. A lower court dismissed the claim as copying "individual poses". On 1 November 2023 the Ninth Circuit reversed and sent the case back, holding that he had plausibly alleged substantial similarity.
Credit
Kyle Hanagami; US Court of Appeals for the Ninth Circuit
2024
Video captions in AI training
On 16 July 2024 Proof News reported that subtitles from 173,536 YouTube videos, taken from more than 48,000 channels, made up a dataset called YouTube Subtitles inside the Pile, a training collection released by the nonprofit EleutherAI. It reported that Apple, Nvidia, Salesforce and Anthropic had trained AI models on the Pile despite YouTube's rules against harvesting material without permission; Anthropic confirmed using it.
LessonCreators found out from reporters, who traced the use through the companies' own research papers.
Credit
Annie Gilbertson and Alex Reisner (Proof News)
Watch for
Face matching for creators
YouTube's likeness detection scans uploads for AI-altered or AI-made copies of an enrolled person's face and shows the matches to that person, who can ask for removal. It launched in 2025 for roughly 4 million creators in YouTube's Partner Program and from May 2026 was opening to every creator over 18 who verifies with a government ID and a selfie video. YouTube's help page calls it experimental; YouTube has published no rates of misses or false matches and told TechCrunch removals so far were "very small".
Achieved whenan outside test publishes detection and false-match rates, or YouTube reports removal counts that an outside party can check.
Claimed by
YouTube, March and May 2026
Rights to your own face
Denmark's culture minister, Zenia Stampe, introduced a bill on 8 October 2026 that would amend Danish copyright law to ban sharing realistic deepfakes of people's faces and voices without their consent, covering ordinary people as well as performers, with exceptions for satire, parody, caricature and criticism. She hopes it takes effect at the start of 2027.
Achieved whenthe Folketing, Denmark's parliament, passes it, it takes effect, and a platform removes content under it.
Claimed by
Danish Ministry of Culture, October 2026
Open
01
Public-domain test, every platform
A repeatable audit of how platforms' matching systems treat new recordings of public-domain music. Consenting performers record at least 300 public-domain works on several instruments; the testers upload them to their own accounts on YouTube, Facebook, Instagram, TikTok and X, and log every claim, mute or block, who made the claim, what share of the audio it covers, and how long each dispute takes. The whole run is repeated six months later.
The briefCredit · unclaimed02
A match service for dances
An opt-in matching service for choreography. Choreographers who consent register a reference video of their own routine; the service turns it into a sequence of body poses and finds re-performances in a test pool, including mirrored, faster or slower versions and excerpts as short as four counts, the length at issue in Hanagami's case. It is tested only on consenting dancers and licensed dance data.
The briefCredit · unclaimed03
Test face matching on yourself
An outside test of YouTube's likeness detection that uses only the testers' own faces. Each tester enrolls a channel. A second channel the tester controls then uploads AI-altered clips of the tester's own face, made from footage they filmed themselves and marked as AI-made in the title and description: face swaps, lip-sync edits and fully generated clips from several tools, plus unaltered clips as controls. The test logs which clips appear in the tester's Likeness tab and how fast.
The briefCredit · unclaimed
Open 01 · the brief
Public-domain test, every platform
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A repeatable audit of how platforms' matching systems treat new recordings of public-domain music. Consenting performers record at least 300 public-domain works on several instruments; the testers upload them to their own accounts on YouTube, Facebook, Instagram, TikTok and X, and log every claim, mute or block, who made the claim, what share of the audio it covers, and how long each dispute takes. The whole run is repeated six months later.
Why it is still open
Berkowitz's 2023 test covered one composer, one instrument and two platforms, and Ulrich Kaiser's 2018 test covered YouTube alone. Meta and Pinterest refused the French regulator Arcom's request to measure their content recognition tools, and rightsholders told Arcom that TikTok's tool arrived late and was ineffective. Searches found no published audit of TikTok or Instagram.
Built with what exists
public-domain scores from IMSLP, the Petrucci Music Library; a home recording setup; the testers' own accounts; each platform's dispute form; a shared spreadsheet.
First step
Record 20 public-domain piano pieces, upload them to one test account on each platform on the same day, and log what happens within 48 hours.
Done looks like
at least 300 recordings by at least 5 performers on 5 platforms, two rounds six months apart, with claim rates, claimants, seconds claimed and dispute times published alongside the recordings under an open license.
Open 02 · the brief
A match service for dances
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An opt-in matching service for choreography. Choreographers who consent register a reference video of their own routine; the service turns it into a sequence of body poses and finds re-performances in a test pool, including mirrored, faster or slower versions and excerpts as short as four counts, the length at issue in Hanagami's case. It is tested only on consenting dancers and licensed dance data.
Why it is still open
In Hanagami v. Epic Games the choreographer found the copying and went to court himself, and the Ninth Circuit held in 2023 that a short sequence can be protected. DanceMatch, an April 2026 research system, retrieves similar choreographies from video and releases a benchmark, but it is research code, not a service choreographers can register with, and searches found no platform that matches movement the way Content ID matches audio and video.
Built with what exists
an open pose estimator such as Google's MediaPipe Pose; DanceMatch's published method and its DanceTypesBenchmark; the AIST++ dance motion dataset from Google, whose software tools are released under the Apache 2.0 license, for unrelated clips; a dynamic time warping baseline, which lines up two motion sequences that run at different speeds.
First step
Film 5 routines and 25 re-performances by friends who sign consent forms, extract poses with MediaPipe, and match them with dynamic time warping.
Done looks like
at least 50 registered routines from consenting choreographers, at least 500 re-performances by at least 20 consenting dancers, and 2,000 unrelated clips; published recall at four and eight counts and the false-match rate, with the code.
Open 03 · the brief
Test face matching on yourself
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An outside test of YouTube's likeness detection that uses only the testers' own faces. Each tester enrolls a channel. A second channel the tester controls then uploads AI-altered clips of the tester's own face, made from footage they filmed themselves and marked as AI-made in the title and description: face swaps, lip-sync edits and fully generated clips from several tools, plus unaltered clips as controls. The test logs which clips appear in the tester's Likeness tab and how fast.
Why it is still open
YouTube's help page calls the feature experimental. YouTube has published no rates of misses or false matches and told TechCrunch that removals so far were "very small". A search for an independent test found none.
Built with what exists
likeness detection in YouTube Studio (age 18 or over, government ID and selfie video); open tools such as LivePortrait, which animates a face from a photo, and FaceFusion, which swaps faces; one or two commercial video generators; a spreadsheet.
First step
Enroll one tester, upload 10 marked clips from a second channel, 5 public and 5 unlisted, and check after a week which ones the tool found.
Done looks like
at least 3 testers and at least 200 altered clips each, across at least 4 tools and 3 kinds of edit, plus 100 unaltered controls; detection rates by tool and kind of edit, median hours to detection and any false matches, published with the method.