Executive Summary
When a company acquires an AI startup, training data usually lands on the diligence sheet as an asset. If it is large, labeled, and running through a working pipeline, it looks on paper like an intangible asset you can put a value on. Lately, though, legal and technical practice has started to look at that line item again. Unverified data may not be an asset at all. It may be a liability.
The international law firm Mayer Brown frames it this way: synthetic data is now treated as its own asset class in a deal, but it carries three diligence risks along with it, in ownership, provenance, and quality. Ownership is hard to establish because copyright rarely attaches. Provenance inherits the infringement risk of the underlying source. Quality hides a performance decline that never shows up on the books. None of the three can be confirmed from financial documents. They surface only under technical diligence.
This piece works through those three axes one at a time to show why the data quality and lineage management Pebblous has long talked about now has to be translated into the language of valuation and contract terms as well. It leaves the reader with a single question. The data we call an asset — would it still be one after diligence?
When the Balance Sheet Flips
There is a reason we call data an asset. Good data produces model performance, and performance becomes the value of the product. Pebblous has written many times that data is an asset. But sit down at the diligence table for an acquisition or an investment and the same sentence reads a little differently. A label that says "asset" is no guarantee that the data stays an asset after the deal closes.
In a July 2026 analysis, Mayer Brown treated synthetic data as its own asset class in an M&A deal. The backdrop is simple. High-quality public data is running dry, and licensing someone else's data or scraping it without permission raises the risk of litigation. So companies are shifting toward generating data directly with models. Ari Morcos of the data-curation firm Datology points out that the internet is only a tiny slice of the world, and that most real data lives in private form on corporate servers.
Once synthetic data became part of what a deal is worth, diligence had to follow. Mayer Brown's conclusion is clear: synthetic data is not a shortcut around diligence, risk allocation, or operational risk. It creates three new questions instead. Whose data is this (ownership)? Where did it come from (provenance)? Is it actually usable (quality)? Each question demands a different method of diligence, and none of them can be answered from the financial statements alone.
The point is that all three risks sit off the books. A data catalog can record however many terabytes or hundreds of millions of records, but those numbers say nothing about whether ownership holds, whether the provenance is lawful, or whether the quality is being maintained. The line between asset and liability is drawn by technical diligence, not by accounting documents.
Drawn out, it looks like this. Above the waterline sits the visible asset on the books, but below it three layers of risk stack up, each larger than the one above. Diligence is the procedure that looks beneath that line.
▲ Original Pebblous diagram — the asset on the books and three layers of hidden diligence risk beneath it
Ownership Diligence: Whose Data Is This?
The first question is ownership. It is easy to assume that acquiring a synthetic dataset means the rights to that data come with it. But the recent direction of U.S. copyright law unsettles that premise. In January 2025, the U.S. Copyright Office stated that a generative-AI output must carry sufficient human expressive contribution to qualify for copyright protection, and that entering a prompt does not, on its own, meet the bar.
A ruling in the same direction landed in the courts. In the March 2025 Thaler v. Perlmutter decision, the federal appeals court reaffirmed the principle that a work must have been authored by a human from the outset. That means it is structurally difficult to claim copyright in a synthetic dataset the model produced on its own. And where copyright does not attach, the exclusive rights an acquirer was counting on grow just as thin.
The alternative is trade-secret protection. But that path comes with conditions. To be recognized as a trade secret, real safeguards — confidentiality procedures, access controls, handling rules — have to actually be in operation. If the data can be viewed from anywhere inside the company and there is no record of who took it out, it is hard to call it a trade secret in legal terms. This is why ownership diligence has to descend past document review into a technical check on how data access controls are actually run.
Provenance Diligence: Inherited Infringement Risk
The second question is provenance. When data carries a "synthetic" label, it is tempting to feel it is unrelated to any problem in the source data. But if the generative model that made the synthetic data was trained on data acquired without a license, that risk stays behind the label and rides all the way through to the output.
The 2025 Bartz v. Anthropic case drew this distinction sharply. The court found that digitizing lawfully purchased print books for internal training was so highly transformative that it counted as fair use. Holding a central library of pirated books, on the other hand, was found not to be fair use. Even for the same data, where and how it was acquired separates the lawful from the infringing. The case was later resolved through a large class settlement.
From the standpoint of acquisition diligence, the decisive fact is this: the infringement risk in source data does not disappear when the transaction closes. The risk transfers, intact, to the owner of the model. The acquirer is the new owner. Whether the deal is large or small, an unverified source remains a potential claim after the acquisition.
So provenance diligence asks not what the data is but where the data came from. Can you prove the generative model's training data was lawfully acquired? Do the terms of any third-party foundation model restrict using its output to train a competing model? Have you secured indemnity from the model provider against IP infringement? These questions can only be answered when the data's provenance can be traced all the way through.
Quality Diligence: The Model Collapse the Books Never Show
The third question is quality. If the first two risks live in the domain of law, quality is where the idea of hidden data debt shows up most literally. The books record vast training data as an asset, but whether that data is already eroding the model's performance never shows up without a technical evaluation.
The evidence comes from the research literature. The model-collapse study Shumailov and colleagues published in Nature in 2024 shows how performance breaks down when AI-generated data is reused, without separate verification, to train the next generation of models. In the early stages, distributional errors accumulate little by little and drift away from the original distribution; in the later stages, rare events vanish altogether. Because the model over-predicts common patterns and misses rare ones, the diversity of the data shrinks with each generation.
▲ Original Pebblous diagram (reinterpreting the model-collapse concept from Shumailov et al., Nature 2024)
The same study offers a mitigation. Mixing a fixed share of real, human-made data — roughly 10 percent — into each generation of training slows the collapse noticeably. Put the other way around, synthetic data is safe only when it is used as a supplement to real data, not a replacement for it. That is why the ratio of synthetic to real data in a target's data, and whether it has been run through performance-degradation testing, becomes a diligence item.
Quality debt hides best in the financial documents. An ownership dispute or an infringement claim eventually surfaces on paper, but a performance decline that creeps in gradually is marked by no number in any catalog. Until you open up the data composition and run the collapse tests, the asset on the books and the liability in reality wear the same face.
When AI-Ready Data Becomes a Diligence Item
Lay out the diligence checklist Mayer Brown proposes and familiar words appear: data composition (the ratio of synthetic to real to scraped), source tracing (proving the legality of training data), quality verification (performance-degradation testing), and, alongside them, copyright claims and privacy risk. These are the items that get carried into the representations and warranties (R&W) of the deal agreement.
Carrying them into representations and warranties means, concretely, writing sentences like these into the contract. The seller represents that it holds commercial usage rights to the synthetic dataset, that it has not violated the terms of the third-party model used to create that data, and that it acquired the source data lawfully. Added to that are statements that it actually performed quality-verification procedures and found no material defects, and that it has never received an IP-infringement claim to date. Because a false representation comes back as damages after the acquisition, the seller is better off running diligence on its own data before signing. The acquirer's diligence, in the end, is a procedure for confirming the diligence the seller did in advance.
This list overlaps almost entirely with what Pebblous has called AI-Ready Data. Ownership diligence asks whether the data's lineage can be traced; provenance diligence asks whether the source has been verified; quality diligence asks whether the quality is being managed. Lineage, provenance, and quality have long been discussed as principles of data engineering for regulatory readiness or model performance. At the M&A diligence table, the same principles reappear in the language of valuation and contract terms.
▲ Original Pebblous diagram — how AI-Ready Data principles map onto M&A diligence and R&W clauses
So this piece does not deny the old Pebblous line that data is an asset. It only adds a condition. Verified data is an asset; unverified data is a liability that has inflated the balance sheet. Actually doing that verification is what the practice of AI-Ready Data amounts to, and it has now become a diligence item that protects enterprise value beyond regulation and performance.
Return to the opening question. The data we call an asset — would it still be one after diligence? The answer depends not on the size of the data but on how well it has been traced, verified, and managed. Diligence is a procedure the acquirer runs, but the preparation for it is work the party creating the data does in advance.
Editor's note. Prompted by Mayer Brown's July 2026 legal analysis, this piece looks at the other side of the data asset from a Pebblous reader's point of view. If you are interested in how to put data-lineage tracing and quality verification into practice, you can look into Pebblous's data-diagnostics service DataClinic.
Frequently Asked Questions
What is data debt?
It refers to data that is booked as an asset in the accounts or the data catalog but, because ownership, provenance, and quality have not been verified, actually comes back to the acquirer as legal claims or performance decline. Size does not make data an asset — it stays an asset only when it has been verified.
Why has synthetic data started being treated as an M&A deal asset?
As high-quality public data runs dry and licensing or scraping others' data raises the risk of litigation, companies have shifted toward generating data directly with models. Once synthetic data became an asset that makes up product value, it entered the scope of diligence too.
Does copyright attach to a synthetic dataset?
Generally it is difficult. The U.S. Copyright Office (January 2025) and Thaler v. Perlmutter hold that a work must carry sufficient human expressive contribution. Output made from a prompt alone struggles to meet the bar, so trade-secret protection is raised as an alternative — but it is recognized only when real safeguards, such as access controls, are in place.
Does a "synthetic" label make the copyright risk of the source data disappear?
No. If the generative model that made the synthetic data was trained on data acquired without a license, that infringement risk is inherited all the way to the output. Bartz v. Anthropic showed this by separating lawful purchase from holding pirated copies. The risk does not vanish when the transaction closes — it transfers to the owner of the model.
What does model collapse have to do with acquisition diligence?
Reusing AI-generated data to train the next generation without verification lets distributional errors accumulate and rare events disappear, so performance breaks down (Shumailov et al., Nature 2024). Because this decline does not show up in a catalog number, technical diligence checking data composition and running performance-degradation tests is required.
Is there a safe way to use synthetic data?
The Nature study reports that mixing roughly 10 percent real, human-made data into each generation of training significantly slows the collapse. In other words, synthetic data is safe when used as a supplement to real data rather than a replacement. How a target manages this ratio is the heart of quality diligence.
How does AI-Ready Data connect to acquisition diligence?
The AI-Ready Data principles of lineage tracing, provenance verification, and quality management map directly onto ownership, provenance, and quality diligence. What had been treated as data-engineering principles for regulatory readiness and performance reappears in M&A as the language of valuation and representations and warranties (R&W).
References
Primary Sources · Legal Analysis
- 1.Mayer Brown. (2026). "Synthetic Data as a Deal Asset: Ownership, Provenance, and Diligence Considerations in AI Acquisitions." Mayer Brown Insights.
- 2.U.S. Copyright Office. (2025). "Copyright and Artificial Intelligence, Part 2: Copyrightability."
Case Law
- 3.Thaler v. Perlmutter (2025). U.S. Court of Appeals for the D.C. Circuit.
- 4.Bartz v. Anthropic PBC (2025). U.S. District Court for the Northern District of California.
Academic Research
- 5.Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., & Gal, Y. (2024). "AI models collapse when trained on recursively generated data." Nature, 631, 755-759.