Executive Summary

On 21 October 2026 the music publisher Reservoir filed two complaints in the Northern District of California. The plaintiffs are Reservoir Media Management, its UK entity and Shapiro, Bernstein & Co.; the defendants are the music generation service Suno and Anthropic. The case numbers are 3:26-cv-12305 for Suno and 5:26-cv-12308 for Anthropic. Plaintiffs' counsel is Richard Busch, who ran the 'Blurred Lines' case for the Marvin Gaye estate.

The two complaints aim at different objects. The Suno complaint identifies 1,050 musical compositions — the work that the music and the lyrics make together, not the sound recording. The Anthropic complaint identifies 853 sets of lyrics, which is to say text. The defendants produce different things, so the rights at issue differ too, and adding the two figures into "1,903 songs" mixes the units. The schedules include Joni Mitchell's 'Both Sides Now' and 'Big Yellow Taxi', Johnny Cash's 'Ring of Fire', Madonna's 'Papa Don't Preach', Miles Davis's 'So What' and Glenn Miller's 'In the Mood'.

The money does not arrive as one lump. Each complaint builds three counts on three different provisions, and for each count it multiplies the per-work statutory maximum by the number of works and writes the product down separately. The table below is those three layers. We redid the multiplication and the addition ourselves, and the results match the complaint to the last digit.

How the two Reservoir complaints are costed. The Suno complaint states the three amounts in its damages paragraphs (¶83, ¶88 and ¶94); the Anthropic complaint's PDF is behind a wall, so the figures there are the ones four independent outlets reported, matching each other down to the dollar. The two defendants together are being asked for $337,782,500.

The row this article is here for is not the middle one. It is the second from the bottom, in orange. Pull out the money attached to the single act of taking the copyright management information off, and it is $26.25 million from Suno and $21.33 million from Anthropic, roughly $47.58 million together. That figure is charged for erasing the label the work carried, independently of whether the work was used without permission.

Two amounts get mixed up here constantly. Summaries that went around right after the filing paired the "$25,000 a song" unit price with "about $28.9 million from Suno," and those two do not belong in the same equation. The $28.9 million is the whole DMCA figure, access-control circumvention and information removal added together; removal on its own is $26.25 million. Put an equals sign between them and the work count stops reconciling when a reader divides back out.

Dividing the two amounts back out shows which one the "$25,000 a song" unit price belongs to. From here on this article uses only the information-removal figure.

Building a claim by multiplying the statutory ceiling by a work count is not something these complaints invented. Nine months earlier, on 29 January 2026, music publishers including Concord Music Group and Universal Music sued the same Anthropic over more than 20,000 scores, lyrics and compositions, seeking more than $3 billion. Divide the two reported figures and you land around $150,000 a work, which is the ceiling the infringement provision sets. What is new in the Reservoir complaints is not the multiplication but running it separately against three provisions, which is what makes the information-removal layer visible as a number.

The scale bears pointing out as well. Reservoir holds more than 150,000 publishing copyrights and 36,000 master recordings. These two complaints identify 1,050 compositions and 853 sets of lyrics; set aside the unit mismatch and any overlap, add them anyway, and you are under 1.3% of the catalog. This is not a suit claiming the whole catalog was infringed. It is a suit that pulled out only what can be proved, and why that choice matters comes back in section 4.

One caveat covers everything below. This is the pleading stage, and no court has found a single fact. The filing date, the court, the case numbers, the amounts claimed and the work counts are on the face of the complaints and can be stated flatly, but whether there was infringement, and whether anyone knowingly took the information off, is something nobody has decided. Where this article carries over an allegation from the complaints, it says so.

Section 1202 of the US Copyright Act bars anyone from casually removing or altering the tag attached to a work. The copyright management information the statute describes covers the title, the name of the author, the copyright owner, the terms of use and the identifying number assigned to the work. Translated into music, that is the song title, the composer and lyricist names, the publisher, and the international standard numbers assigned separately to the recording and to the composition.

The question this provision asks is not the question the infringement provision asks. Infringement asks whether the work was used without permission. Section 1202 asks whether the record of who made it was erased. That is why the two claims sit side by side in one complaint and get costed separately. The table below puts the two questions next to each other.

The two claims live in one complaint but ask about different facts and are costed separately. They are not fully independent, though. Section 1202 requires that the removal was done knowing it would induce or conceal infringement, so how the infringement side is decided bears on the fate of this claim.

The element is what matters. Section 1202 reaches removal or alteration done "knowingly" or "having reasonable grounds to know." It is not a provision that punishes metadata lost by accident, which means a plaintiff cannot stop at "the tag is gone" and has to plead circumstances supporting "they took it off knowing." That is why the Reservoir complaints spend so long on the details of the collection process, in a place that has nothing to do with the fair use fight.

The complaints plead three strands of such circumstances. Everything in this paragraph is an allegation no court has ruled on. First, the complaints assert that leaked internal Suno source code still contains workflows for scraping YouTube Music and the lyrics site Genius. Second, that commercial proxy services and CAPTCHA-defeating tools were used to get past gated sections. Third, that tracks were run through stem separation — splitting a recording into components such as vocals and backing — and that covers were trained alongside originals. None of the three is about what the model learned. All three are about how the material was taken and what got dropped on the way.

The defendants' side is recorded here as well. Anthropic issued a statement the day after the filing.

Suno did not respond to requests for comment from several outlets. On the plaintiffs' side, Reservoir chief executive Golnar Khosrowshahi framed the filings as support for technological progress coupled with a commitment to keep protecting the industry's core creative work. Both sides are quoted here for a simple reason. What this article is trying to read is not who wins, but which item got a price attached to it.

Put the conduct the complaints describe into data terms and it narrows to one question. At which point in a training pipeline does the credit fall off? Walking through music preprocessing stage by stage to answer that produces an uncomfortable picture. It is less that the pipeline loses the metadata than that it never accepts it in the first place.

Start with decoding. The moment a compressed track is unpacked into an array of numbers by a tool like ffmpeg, the tags riding on the file container have nowhere in the output to go. What the model eats is a waveform, not a file. Nobody erased anything; the format changed.

Stem separation cuts the thread once more. Demucs, the widely used source separation model, takes a single stereo mix and returns four files: drums, bass, vocals and other. The only information that travels with them is the filename. What the original track was, who wrote it, on what terms it can be used — none of that is written anywhere. Once the song is no longer a unit, there is no tag left to take off.

There is one more symptom of how far this field sits from metadata. Demucs's own training data was labeled by filename heuristics. A track called 'Lead Guitar.wav' goes into the 'other' bucket, and the ambiguous cases were settled by a person listening to them. With no structured metadata available, file names had to do the job.

And here a counterexample turns up, from inside the same industry. Stable Audio 3 preserves the metadata it receives from AudioSparx and Freesound and builds its training prompts out of it. Fields such as description, title, keywords, beats per minute, genre, mood and instrumentation are sampled at random and strung together into a text condition. In that design the metadata is not a cost but the signal that makes the model useful. Build it a place to ride along and performance can go up.

That counterexample carries legal weight. The mainstream proposal in the law review literature on applying section 1202 to generative AI is that liability should not extend to cases where keeping the information is technically impossible or unreasonably burdensome. For that proposal to persuade, "keeping it was impossible" has to be true. In music, another model in the same industry carries the metadata and trains on it. If cutting the thread was a design choice rather than a limit of the technology, it starts to meet the "knowingly" element that section 1202 requires. Nowhere else in this article do the data question and the legal question sit this close together.

This is the point where it becomes tempting to blame cost for the broken chain. If attaching a label and carrying it all the way through were expensive, it would be hard to fault whoever dropped it. But music happens to be the data where that excuse works least well. The industry has run permanent per-work identifier schemes for decades, and the cost of issuing one more number rounds to zero.

What it costs to issue a music identifier. The price of attaching one is near zero, and the markings still vanished from the training data. The expense is not in issuing the number but in designing a process that preserves and hands it on across every stage of the pipeline.

A complication particular to music sits on top of that. Rights in a single song come in three layers. The composition that the music and lyrics make together, the lyrics as text, and the recording of one particular performance can each belong to a different owner. That structure is why one Reservoir complaint identifies 1,050 compositions and the other 853 sets of lyrics. Suno produces sound and Anthropic produces text, so the layer each one targets differs. Tracking provenance in music means tracking three layers separately, which is why preserving only one of them still leaves the other two cut.

If the markings vanished even from data where identifiers are nearly free, there is no need to ask about data with no identifier scheme at all. Text and images scraped off the web carry no work number, no rights-holder registry and no terms-of-use field. What happened in music should be assumed to be happening, more severely, across the much larger territory outside it.

Work to fill that gap outside music has already started, though. On the image and video side there is a standard that signs provenance information into the file at the moment of capture or creation, and it is genuinely in use. More than 6,000 companies and institutions have joined the Content Authenticity Initiative behind it, TikTok has applied the label to more than a billion videos, and OpenAI applies it to everything its image generators produce. Sony, Canon, Nikon and Leica have built in-camera signing at the point of capture. Against that, this article's research turned up no case of the same standard being applied to music or audio. The medium that has run per-work numbering longest stands furthest back in the move to embed provenance into what gets generated.

4.1Strip the Tags and the Training Use Still Shows

Read this far and the conclusion tilts toward resignation. The tags fell off and only waveforms remain, so what is there left to prove? A study published in May 2026 overturns that premise. Without looking at model parameters or training logs, it decides by querying alone whether a given track was used in training. Across several recent music generation models it reports accuracy as high as 98.6%, a 1.9% rate of wrongly flagging unused tracks, and a 1.0% rate of missing used ones.

One thing must not get mixed in while reading those numbers. What the method measures is whether the track was used in training, not how closely the model reproduces it. Figures measuring memorization and reproduction are something this article could not obtain, so it does not fold the two questions into one.

It does answer the question deferred in section 1. Why did Reservoir identify under 1.3% of the 150,000 works it holds? Because this is not a suit over everything but a sample suit over what can be proved. Lay those two facts over each other and the section's conclusion appears. Provenance does not disappear because someone erased it. The cost of reconstructing it merely shifts from the party that erased it to the party it was taken from, and that cost is falling fast.

The two sides can now be set next to each other. On one side is the cost of attaching provenance and carrying it all the way through. On the other is the price attached when that is not done. One caution applies to the right-hand side. The $25,000 a work written into the complaints is not a settled figure but a ceiling.

The statutory band runs from $2,500 to $25,000 a work, and Reservoir wrote in the top of it. If a court picks the floor, Suno's share is not $26.25 million but $2.625 million, a tenfold drop, and a violator who establishes it did not know can have the court reduce the award or remit it entirely. On top of that the provision itself turns on acting knowingly, so reaching the ceiling means carrying that element all the way to the end.

Amounts that have actually been realized sit an order of magnitude away. Bartz v. Anthropic, the largest AI copyright settlement disclosed to date, came to $1.5 billion over 482,460 works. Per work, that converts to somewhere between $2,200 and $3,100 depending on which stage of the calculation is used: lowest on the first distribution, higher once attorneys' fees are set aside, highest if the gross settlement is simply divided. Whichever stage is counted, it stays around 10% of the section 1202 ceiling. That case, though, was an infringement settlement in the book market rather than an information-removal case, and it is a realized settlement rather than an amount claimed. Both qualifiers travel with the figure every time.

So what has a court set per song in music? As of October 2026 that number does not exist. The Munich Regional Court did order Suno to pay damages in July of this year, but the amount was deferred to a follow-on proceeding and the case covered only six songs. Section 6 returns to that ruling. There is therefore not a single realized music figure to place on the right of the chart below, and the one realized figure available for comparison is borrowed from the book market.

The chart below puts the per-work amounts in this article on a single logarithmic scale. The scale is logarithmic because the left side and the right side are separated by orders of magnitude.

The left side comes from an earlier Pebblous report. Auditing the license and provenance of a single public dataset takes 3.5 to 15 hours, and a program bundling five to ten datasets takes 40 to 160 hours. In money, the cheapest case is under $1,000 and the most expensive, with legal review attached, runs to about $30,000. Set those figures beside the right side and the contrast sharpens. Documenting ten datasets end to end costs about what the information-removal ceiling comes to for one or two songs of music.

Then why was nothing written down? Because not writing it down was not a choice these two companies alone made. In an audit of 1,858 public datasets, missing license information ran above 72% and mislabeling above 50%. What is worth noticing is that the audit itself pulled the "unknown" share down from the low 70s to 30%. The information had not vanished from the world; nobody had written it down. An audit looking at audio datasets alone found 36% sourced from web scraping and 35% likely to contain material that requires a purchased license. Documentation debt is closer to the default setting than to an exception.

5.1Provenance Is Becoming a Condition Outside the Courts Too

Courts are not the only channel through which a price gets set. Contracts moved first. The deal Universal Music struck with Udio in October 2025 added, on top of artist opt-in and a walled-garden output model, an obligation on Udio to run content identification — to filter outputs substantially similar to Universal's catalog. Royalties were reported to be tied to how closely an output resembles the training data. It amounts to a contract whose accounting does not run at all unless provenance can be traced. Suno in turn reached the first major-label licensing deal, with Warner Music, in November 2025.

No per-song rate can be read out of these contracts, though. Every music generation AI licensing deal concluded in 2025 and 2026 kept its financial terms confidential. So the comparison "what would a license have cost per song" cannot be made from public material, and the impossibility of making it is itself the current state of this market. It also means the only public channel where a price gets set is still the court.

The claim that a price has been set needs discounting here, by this article itself. Being charged and collecting are different things, and section 1202 claims have a poor record in AI litigation so far. But stringing those dismissals into "this provision does not work against AI" misses where these complaints stand. What got dismissed was mostly claims aimed at the output. What Reservoir aims at is the training stage.

The ruling that drew that line came five weeks before the filing. On 16 September 2026 the Ninth Circuit declined to allow a section 1202 claim against output in the GitHub Copilot case. Its reasoning rests on the words of the statute: "remove" and "alter" presuppose an existing copy that already carried the information, so producing a new output without that information in it is not removal. In the same ruling, the identicality the district court had treated as an element was not recognized as an independent requirement. Identicality is merely circumstantial support for removal having occurred.

The part that matters for this case is the part the ruling did not reach. It did not decide the input-side theory — the claim that the information was taken off during training — because the plaintiffs there did not press it to the end. The opinion went a step further and indicated that a claim asserting the stored copies had been served back without the information could have been the stronger one. Thirty-five days later Reservoir walked through the door left open, in the very court the ruling binds.

The table below sorts the public record of section 1202 claims by what each one aimed at. The empty cell at the far right is the message of the table.

The leading cases in which a section 1202 claim has been litigated. Dismissals dominate, but the places where they landed are mostly on the output side, and no US court has yet ruled on a section 1202 claim that squarely contests the training stage.

One case sits outside the table. It is not a US matter, so it is not listed, but a ruling three months ago in Germany covered the same defendant and the same collection method. On 31 July 2026 the 42nd Civil Chamber of the Munich Regional Court found Suno largely liable in a case the German collecting society GEMA brought over six songs. Two findings overlap with this one. The court treated the copying that happens during training as an infringement of the reproduction right and held that the text and data mining exception does not stretch that far, and it found that Suno circumvented stream protection measures when pulling audio from YouTube. Conduct that the Reservoir complaints plead as allegation has already been written down as fact by a court in another country.

That ruling should still not be read as a trailer for this case. The law applied was German copyright law, and removal of copyright management information was not at issue there. The judgment is not final, and Suno said it is considering an appeal, calling the decision a fundamental misunderstanding of its technology. Damages were ordered but the amount was pushed into a follow-on proceeding. So, as section 5 put it, no court has yet set a per-song figure in music generation AI.

A fact the table does not show belongs here too. US AI copyright suits had piled up to somewhere around 150 by October 2026, and eight new ones arrived in the week of this filing, the most in any single week — but nobody has published a count of how many of them include a section 1202 claim. So this article uses no percentage. What can be confirmed is that section 1202 claims against large language models numbered at least seven as of October 2025, and that federal court opinions citing the provision between 2020 and 2024 came to 330, more than 1.5 times the preceding five years. How often the provision gets litigated is going up.

Rather than predict how this case ends, then, it is more accurate to put it this way. The door Reservoir walked through has never been closed, and nobody has ever passed all the way through it either. This case is the next test of whether section 1202 reaches training data, and until there is a result, $25,000 a work stays a ceiling on negotiating leverage rather than an award. For the record, the Elvis Presley estate filed a separate suit against Anthropic over 274 songs in the same week; it is built on different issues, so this article does not take it further.

There is a reason Pebblous spent this long on two complaints filed in a US court. What happens across the preceding six sections is not a product of circumstances peculiar to music. It takes the same shape anywhere data is gathered and fed to a model.

7.1An Item of Good Practice Just Got a Number

The definition of AI-Ready Data that Pebblous has been using always included "provenance can be traced." Until now that was a recommendation, not a line item with money on it. These complaints attached a band of $2,500 to $25,000 a work to it. Just as DataClinic counts missing values, duplicates and label errors when it diagnoses a dataset, there is now a basis for making the presence of a provenance field, and whether it survives to the end, diagnostic items of their own. For products that generate data and ship it out, embedding provenance in what gets shipped moves from an optional feature toward a delivery requirement.

7.2Quality Metrics and Provenance Metrics Run on Different Axes

Copyright management information is metadata, not content. The accuracy, completeness and consistency that international standards measure are all metrics on the content axis, and "has the record of where this data came from been preserved" lies outside that axis. As section 3 showed, the heart of the problem is not that the pipeline discards provenance but that it was never given a place to carry it. Decode and the tags fly off; split the stems and the song stops being a unit. Preserving provenance is therefore only possible by designing each stage to lift a record out and move it alongside, which is work on a different axis from raising a quality score. This is exactly the place to say that a provenance metric has to be stood up next to the quality metrics.

7.3Neither the Buyer Nor the Supplier Knows

Contracts for buying and selling training data rarely contain a clause saying provenance metadata will be preserved and handed over with the data. The receiving side takes delivery of data with a broken chain and does not know it; the supplying side drops the chain in preprocessing and does not know that counts as a defect. The 98.6% from section 4 adds urgency. "We erased it, so nobody will know" no longer holds. Here is what can be picked up and used now.

• When bringing in a dataset, ask whether the original identifier, the rights holder and the terms of use survive at the record level. "There is provenance" and "it is attached to every record" are different answers.
• Put a clause in the supply contract requiring identifying information to be preserved and passed on after preprocessing. Where a stage makes preservation impossible, get that stage and the reason in writing.
• Stand up a provenance field in the AI bill of materials, not just the model and the dataset. Record which transforms ran and what dropped at which stage, and the question can be answered afterward.
• For a pipeline already in production, measure once what disappears on either side of the decoding and splitting stages. Usually it was not recorded rather than impossible to record.

7.4The Point Where a Norm Turns Into a Price

Several articles on data provenance have accumulated on the Pebblous blog: Suno's Copyright Case Is Auditing the Provenance of 60,000 Training Songs, Music Publishers Ask a Court to Order an Accounting of Claude's Training Data, 21 Million Songs Trained AI — and a Newsroom Named the Source, and The Real Invoice for Free Data. This one is the next in that line and also shifts what the line is doing. The earlier pieces argued that provenance ought to be traceable. This one reports that somebody has calculated what erasing provenance is worth, filed the bill in court, and done so through a door that opened five weeks earlier. Getting to the point where a norm turns into a price before anyone else does is what this article is for.

The sources are of different grades, so they are grouped. The first group is court documents and statutes, the origin of the amounts, case numbers and holdings in this article. The second is academic literature, covering provenance loss and recovery in the pipeline and the documentation-debt figures. The third is legal analysis and industry reporting, and the fourth is earlier Pebblous articles on the same problem.

Court documents and statutes (primary)

• 1.Reservoir Media Management, Inc. et al. v. Suno, Inc., No. 3:26-cv-12305 (N.D. Cal., filed 21 October 2026), Complaint ¶83, ¶88, ¶94 and Prayer (g)–(h) — the source of the Suno amounts and the 1,050 compositions in the section 1 table, and of the collection-process allegations in section 2. Original PDF read directly.
• 2.Reservoir Media Management, Inc. et al. v. Anthropic PBC, No. 5:26-cv-12308 (N.D. Cal., filed 21 October 2026) — the Anthropic amounts and the 853 sets of lyrics in the section 1 table. ⚠️ The original PDF was inaccessible, so the figures used are those four independent outlets reported, matching each other down to the dollar.
• 3.Doe 1 v. GitHub, Inc., No. 24-7700, 2026 WL 2728464 (9th Cir. 16 September 2026) — section 6's dismissal of the output theory, the holding that identicality is not an independent element, and the input theory left undecided.
• 4.The New York Times Co. et al. v. Microsoft Corp. et al., No. 1:23-cv-11195 (S.D.N.Y.), order of 26 March 2025 — the split-by-plaintiff outcome in the section 6 table. Only the Daily News and the Center for Investigative Reporting claims survived; the Times's own §1202(b)(1) claim was dismissed.
• 5.17 U.S.C. §1202 (integrity of copyright management information) and §1203(c)(3)(B), (c)(5) (statutory damages, reduction and remission) — the definition and elements in section 2 and the $2,500–25,000 per-work band in section 5. law.cornell.edu
• 6.Reservoir Media, Inc., FY2026 Form 10-K and investor fact sheet — the catalog scale cited in sections 1 and 4 (more than 150,000 publishing copyrights, 36,000 masters). sec.gov

Academic literature

• 7.Longpre, S. et al. The Data Provenance Initiative: A Large Scale Audit of Dataset Licensing & Attribution in AI. arXiv:2310.16787 (2023); Nature Machine Intelligence (2024) — section 5's missing license information above 72%, mislabeling above 50%, and the audit pulling the unknown share down to 30%. arxiv.org
• 8.Agnew, W. et al. Sound Check: Auditing Recent Audio Dataset Practices. Proc. AAAI/ACM AIES 8(1):26–40 (2025). arXiv:2410.13114 — section 5's 36% of audio datasets sourced by web scraping and 35% likely to contain infringing material. arxiv.org
• 9.Liu, Y.C. et al. Auditing Training Data in Generative Music Models via Black-Box Membership Inference. arXiv:2605.29202 (28 May 2026) — section 4's accuracy of up to 98.6%, 1.9% false positives and 1.0% false negatives. arxiv.org
• 10.Défossez, A. et al. Music Source Separation in the Waveform Domain and Hybrid Transformers for Music Source Separation (arXiv:2211.08553) — section 3's stem separation output structure and filename-heuristic labeling.
• 11.Stable Audio 3, arXiv:2605.17991 — the counterexample in section 3: a design that preserves metadata and uses it as the conditioning signal in training prompts.
• 12.Bersh, L. Beyond Infringement: Rethinking DMCA §1202 for Generative AI. 78 Stan. L. Rev. 667 (2026) — section 3's technical-impossibility safe harbor proposal and section 6's survey of seven section 1202 claims against language models. review.law.stanford.edu

Legal analysis and industry reporting

• 13.Sterne Kessler, Ninth Circuit Rejects DMCA Attribution Claims Against Generative AI Outputs(30 September 2026) and Haynes Boone, Ninth Circuit Rejects DMCA Section 1202(b) Theory Against Copilot Output in Doe v. GitHub — commentary on the section 6 ruling.
• 14.Digital Music News (22 October 2026), CelebrityAccess (25 October 2026) and LawCommentary — the fact of the filings, the amounts claimed, and both sides' statements in section 2. digitalmusicnews.com
• 15.Chartlex, Udio–UMG walled garden explained, and Digital Music News (25 November 2025) — section 5.1's content identification obligation, the similarity-linked royalty structure, and the Warner Music licensing deal with Suno.
• 16.AAP and Authors Alliance guidance on the Bartz v. Anthropic settlement (final approval 20 July 2026) — section 5's $1.5 billion settlement total, 482,460 works, and the per-work conversion. authorsalliance.org
• 17.ISRC registration guidance and Journal of the Copyright Society, Section 1202(b) and AI — section 4's identifier issuance costs and section 6's 330 opinions citing §1202(b) between 2020 and 2024. isrc.com
• 18.TechCrunch (29 January 2026), Music publishers sue Anthropic for $3B — section 1's suit by Concord Music Group, Universal Music and others against Anthropic, over 20,000 works, seeking more than $3 billion. The figure of around $150,000 a work is this article's own division of the two reported numbers. techcrunch.com
• 19.GEMA v. Suno, Munich Regional Court, 42nd Civil Chamber, 42 O 763/25 (31 July 2026) — the German ruling in sections 5 and 6: training-stage copying held infringing, the text and data mining exception excluded, circumvention of YouTube stream protection found, damages unquantified, judgment not final. Confirmed through Bird & Bird and JUVE Patent commentary. juve-patent.com
• 20.C2PA and Content Authenticity Initiative adoption status (2026) — section 4's more than 6,000 members, more than a billion TikTok labels, OpenAI's application across its image output, and camera-maker signing support. c2pa.org

Related Pebblous articles

• 21.Suno's Copyright Case Is Auditing the Provenance of 60,000 Training Songs
• 22.Music Publishers Ask a Court to Order an Accounting of Claude's Training Data
• 23.21 Million Songs Trained AI — and a Newsroom Named the Source
• 24.The Real Invoice for Free Data
• 25.The AI Bill of Materials (AI BOM) That Clears Out Shadow AI