--- title: Data Has a Chain of Custody Now author: Hilal Agil date: 2026-04-15 source: https://hilalagil.com/essays/data-has-a-chain-of-custody-now/ topics: Artificial Intelligence, Data, AI Governance, Music --- # Data Has a Chain of Custody Now In late January, three music publishers — Universal Music Publishing Group, Concord, and ABKCO — filed a $3.1 billion lawsuit against Anthropic. I read it with more than professional interest. I spent the first chapter of my working life as a musician, and the catalogs named in that suit are the kind of work people spend decades of their lives making. But the detail worth paying attention to isn't the number. It's the legal theory. The suit doesn't primarily argue that training an AI model on songs is illegitimate. It argues that the data was taken through piracy — that the collection itself was the violation. And that framing is showing up everywhere now. Across the AI lawsuits moving through courts this year, the question is drifting from whether training is fair use to a much older and sharper question: where did you get this, and can you prove it? In other words, data has acquired a chain of custody. I think that's one of the most consequential shifts happening in AI right now, and it's being underestimated because it looks like legal housekeeping. ## From free harvest to accounted supply For the first decade of modern machine learning, data was treated like weather — something that was simply there. You scraped the web, you trained, and the provenance of any individual piece of text or audio was nobody's concern, least of all the model's. That era is visibly closing. Reddit now licenses live access to its data on usage-based terms measured in the tens of millions of dollars per year, and treats that access as a product with pricing tiers, not a favor. Publisher content marketplaces have moved from experiment to ordinary business. The going assumption among researchers — that the supply of quality public text is on a path to exhaustion sometime between now and the early 2030s — has stopped being a provocative forecast and become planning input. Put those together and you get a new default: data is supplied, not harvested. Supplied things have owners, terms, prices, and records. Harvested things don't. The entire apparatus of the last decade was built for harvesting, and it is being retrofitted, lawsuit by lawsuit and license by license, into a supply chain. ## The web was never the world Here's why I think this matters beyond the courtroom. The web — the thing all of this litigation is fighting over — was always a thin and biased sample of reality. It's what people happened to type, photograph, and upload. It contains very little of what the world actually knows: what happens on factory floors and coral reefs, inside instruments and hospitals and power grids, in the lived experience of people who never post. AI is now capable enough to make use of all of that. The models are not data-starved because the world lacks information; they're data-starved because almost none of the world's information was ever designed to be reachable, and its owners — rightly — won't hand it over into a system with no chain of custody, no rules of access, and no way to get paid. So the two problems are actually one problem. The reason the web's data is ending up in court and the reason the world's data stays locked away are the same: we built systems for taking data, not for governing it. ## Provenance is the product This is the conviction underneath most of what I'm building now. If the next decade of AI depends on data the web never had, then the infrastructure that matters isn't another scraper — it's the layer that lets data move with its history and its rules attached: who it belongs to, what it may be used for, on what terms, with a record that survives the transaction. That's the thesis behind [Ipnops](/about): making the world's data — digital and physical — something AI can reach, understand, and act on safely, with clear rules for how it's accessed and governed. And it's the thesis behind the verifiable-data side of [Vestigim](/essays/vestigim/): musicians shouldn't have to discover in a court filing, years later, how their work entered a training run. The record should exist from the start, and the terms should be enforceable from the start. None of this requires believing the lawsuits will all succeed, or that scraping will disappear. It requires only noticing which direction every incentive now points. Rights holders want records. Regulators want records. Model builders, frankly, want records too — provenance is the only durable defense they'll ever have. For most of a decade, the honest answer to "where did your model's knowledge come from?" was: everywhere, and don't ask. That answer is dying in courtrooms right now. What replaces it — data that carries its chain of custody with it — is a better foundation for everyone, including the machines.