Bad data always sends the bill eventually

And most of the time, you can get away with not doing it.

The books still ship. Invoices still go out. Royalty statements are produced. Somebody knows which spreadsheet contains the real answer.

Until suddenly the business asks a question where the answer matters.

The $1.5 billion Anthropic copyright settlement is giving us a very good example.

The settlement covers books Anthropic downloaded from pirate libraries including LibGen and Pirate Library Mirror. Eligible works are expected to receive roughly $3,000 each before the final impact of fees, costs and adjustments. For many trade and university press books, the settlement provides a default 50/50 allocation between authors and publishers, although contracts can support a different split.

Fair enough.

Then the reconciliation notices went out.

Authors started finding publishers claiming percentages they did not expect. Old rights relationships resurfaced. Several parties appeared against the same work. Agents showed up. Some publishers had claimed 100 percent.

At first glance, this looked terrible.

In some cases, it was apparently much less dramatic. Publishers had entered the percentages incorrectly, particularly when filing claims in bulk, and started fixing them. The Authors Guild has been fairly careful about distinguishing administrative mistakes from actual disputes.

So no, I do not think the useful story here is “publishers are stealing authors' money.”

That would also be a little too easy.

The more interesting problem is that in quite a few cases somebody now has to answer a seemingly basic question:

Who actually held the rights to this book when Anthropic downloaded it?

And apparently, that answer is not always sitting neatly in a database waiting to be retrieved.

Rights reversion is one of the biggest issues emerging from the settlement.

Suppose a publisher held the rights to a book for twenty years and they eventually reverted to the author. Who gets the Anthropic money?

Well, when did they revert?

The Authors Guild points to August 10, 2022, the later of the relevant Anthropic download dates. If the rights had already reverted by then, the author may have a claim to the full settlement payment depending on the contract. If they reverted afterward, the publisher may still have a legitimate claim because it held the relevant rights when the infringement happened.

Now we have a problem.

A database entry saying:

Rights owner: Author

is not enough.

You need to know that the publisher held certain rights until, say, April 17, 2021. You need to know what those rights actually were. You need to know whether the contract was amended. You need the reversion date. And preferably you need something more reliable than “I think Susan handled that before she retired.”

Publishers Weekly reported that rights reversion has become the most common issue the Authors Guild is hearing about in the reconciliation process. According to the Guild, incomplete or missing publisher records are contributing to the problem.

There is really no elegant way to dress that up.

If nobody can reliably establish whether a publisher still owned a right, you have a rights-management problem.

Not an AI problem.

Not an Anthropic problem.

A rights-management problem.

I do understand how we get here.

Publishing contracts can live for decades. Companies get acquired. Imprints disappear. Systems get replaced. Data gets migrated, sometimes several times. People leave. Rights revert. Contracts are amended. The operational information gets updated in one place and not another.

I have spent enough time around publishing systems to know that these things do not become messy because somebody woke up one morning and decided to build a terrible rights database.

They become messy slowly.

One exception here. One workaround there. One spreadsheet that was only meant to survive until the new system was implemented in 2014.

And suddenly it is 2026 and the spreadsheet has tenure.

That part is completely understandable.

Pretending the result is good enough because it has worked so far is where I have a problem.

This is probably why the Anthropic story makes me smile a little.

I have been having versions of this conversation for years. The reason for fixing the data keeps changing. The problem underneath it does not.

We saw it with analytics.

Everyone wanted better dashboards, title profitability, forecasting and customer insight. Great. Then people started looking at the data feeding those dashboards.

Different customer records. Inconsistent categories. Product relationships that were never maintained properly. Historical information that nobody completely trusted.

There is only so much Power BI can do for you at that point.

A beautiful dashboard does not make bad data correct. It just gives the bad data nicer fonts.

Then AI arrived.

Suddenly publishers wanted semantic search, automated metadata, recommendation engines, internal assistants and tools that could understand the backlist.

And again we discovered something shocking:

The AI needs decent data too.

If your metadata contradicts itself, your contracts are sitting in PDFs nobody has structured, and three systems disagree on which record is correct, the AI does not float above all of that and discover the truth.

It processes the mess faster.

Congratulations.

Now rights are getting their turn.

Anthropic did not create the underlying problem. It simply put enough money behind one very specific question that suddenly everyone needs the answer.

The current fights about who gets which percentage were not actually the first warning.

Earlier, authors discovered that some books could not participate in the settlement because the copyright had apparently not been registered in time. In some of those cases, authors said the publishing contract required the publisher to handle registration.

The Authors Guild then surveyed more than 500 authors covering over 1,400 books.

For 864 books where respondents provided information about a contractual registration requirement, authors reported that publishers had failed to comply in 804 cases. That is about 93 percent.

Before anybody quotes that number in a keynote tomorrow: no, 93 percent of publishers are not failing to register copyrights.

The Authors Guild itself says the survey was self-selecting. Authors who had a problem had a pretty obvious reason to respond. It would be irresponsible to turn that into an industry-wide failure rate.

But 804 reported books is still 804 reported books.

And 321 of them were published between 2019 and 2025, so we cannot explain the whole thing away as dusty contracts from 1978 either.

If the contract says somebody is supposed to register the copyright and nobody does it, there are only so many sophisticated explanations available.

Sometimes a process just failed.

The reason anyone cares now is because there is money attached to the failure.

We talk a lot about technical debt. Publishing has plenty of data debt too.

The nice thing about data debt is that for years it can look completely harmless.

The title is published. The author gets paid. The sales report more or less works. Everyone knows the workaround.

So why spend money fixing it?

That argument works surprisingly well right up until somebody asks you to answer the question across 20,000 titles.

Then suddenly there are six people on a Teams call discussing a spreadsheet last updated by someone who left the company four years ago.

This is also why I struggle with the idea that every data-cleanup project should have an obvious short-term ROI before you start.

Sometimes you simply do not know what the future use case will be. 

Five years ago, nobody was going to walk into a budget meeting and say:

"I think we should clean up our historical rights data because there is a chance an AI company will download millions of books from pirate libraries and eventually agree to a $1.5 billion settlement."

That business case would probably not have made it past the first slide.

And quite rightly.

But that is also the problem.

The question that eventually makes your data valuable is often not the question you expected.

Now imagine somebody approaches a publisher tomorrow with a serious proposal to license 20,000 backlist titles for an AI application.

Great.

Can you license them?

Not “do you own the books?”

Can you license this specific use, in these territories, for this period, under these contracts, and do you know who else needs to get paid?

For ten titles, you can pull the contracts and figure it out.

For 20,000, “we will look at the contracts” is not really a strategy.

You need structured information you trust.

And I think we are going to discover this over and over again as AI licensing develops. Not because every publisher should already have designed the perfect rights architecture for use cases nobody could predict.

That would be ridiculous.

But if your answer to every new opportunity starts with three weeks of archaeology, that eventually becomes a commercial problem.

I would not start a “perfect data” initiative.

Nothing good has ever happened after someone puts Perfect Data Initiative on a PowerPoint slide. 

Instead, pick one business question that you believe will matter.

Maybe it is AI licensing.

Maybe it is title profitability.

Maybe it is rights availability.

Maybe it is understanding who your readers actually are.

Then take a meaningful piece of your catalog and try to answer the question using the systems you have today.

No manual cleanup first.

No special spreadsheet.

And no asking the person who has worked there for 23 years and somehow knows everything.

Because if the process only works when that person is in the room, you do not have a process. You have a very valuable employee.

See what happens.

You will probably find some things that work surprisingly well. You will probably find some things that do not.

Then fix those.

I will meanwhile continue repeating the same deeply unsexy advice I have been giving publishers for years:

Get your data in order.

Analytics found the problems.

AI found them again.

Now a $1.5 billion copyright settlement has found some more.

The next use case is already on its way. We just do not know what it is yet.

So here is my question: What is the one thing about your backlist that you would really prefer nobody asked you to prove tomorrow?

Sources and further reading

Anthropic Copyright Settlement: official FAQ and allocation rules

Anthropic Copyright Settlement: current settlement status

Authors Guild: guidance on settlement claims and rights reversion

Publishers Weekly: Authors Guild Calls on Publishers to Share Anthropic Wealth

Authors Guild: copyright registration survey

Jane Friedman: Disagreements in the Anthropic Settlement Were Inevitable, But It Feels Worse Than Expected