My Book Sold for $656 in AI Royalties

Opening

Reader, a slightly odd payment landed in my account not long ago. It was a licensing fee for AI training use of a book I’d written — $300 to $656 per book. At today’s exchange rate, that’s roughly ₩900,000 at most. A publisher brokered the deal, and it went through a US company called Mentat (Mentat).

The first thought I had, looking at that number, wasn’t “that’s cheap” or “that’s a lot.” It was closer to I genuinely don’t know what this number is supposed to represent.

Then, on August 12, I came across a Wall Street Journal report. Apple is negotiating with news publishers to overhaul Siri, and the terms are unfamiliar. Instead of a flat fee, Apple wants to pay each time content is actually used.

Let me tell you where this is going. What’s unfolding right now isn’t a story about the price of creative work going up or down. The very sentence used to assign value is shifting — from “this book exists” to “this book was summoned.” And there’s an industry that has already lived through this exact rewrite once. We already know how it ends.

So Far, Price Has Been Attached to ‘Existence’

To see where $656 sits, we need to lay out the price sheets that have gone public over the last two years.

The most frequently cited benchmark is the HarperCollins–Microsoft deal. It was disclosed in November 2024, and the terms are quite specific. $5,000 per book, split 50-50 between author and publisher, meaning the author’s share is $2,500. The term runs three years, with restrictions preventing model output from reproducing more than 200 consecutive words of the original text or over 5% of the entire book. It was an opt-in arrangement — inclusion required the author’s consent.

The numbers on the media side are a different order of magnitude. News Corp’s deal with OpenAI, announced in May 2024, was reported at up to $250 million over five years. That’s a flat, company-to-company contract.

The third benchmark isn’t a deal at all — it’s a settlement. Anthropic’s $1.5 billion agreement with authors received final approval on July 21, 2026, covering roughly 500,000 works at around $3,000 per work. What matters here is that this wasn’t a price set through normal transaction — it was a price attached to unauthorized use.

Set my $656 alongside these three, and its position becomes clear. Comparing author shares alone, it’s a little over a quarter of the HarperCollins figure, and roughly a fifth of the per-work rate in the Anthropic settlement.

You might be tempted to jump straight to “Korean authors got sold cheap” — but I think that reading misses something important. Language, genre, length, term, and exclusivity all differ, which makes simple comparison difficult. The HarperCollins deal was English-language nonfiction limited to three years; my case has different terms entirely.

What I’m focused on isn’t the amount — it’s the fact that all three are using the same grammar. So much per book, so much per work. In other words, the price was attached to the simple fact that the book exists and entered the dataset. However often that book gets cited afterward, or even if it’s never called up again, the amount stays the same.

This is, in fact, a fairly favorable grammar for creators. It’s predictable.

What Apple is really changing isn’t the amount, it’s the sentence structure

Let’s circle back to the reporting that kicked this off.

According to the WSJ, Apple has been in touch with several news organizations over the past few months, offering multi-year deals. The budget under discussion runs nine figures — at least $100 million. But the payment structure is variable. Apple pays out in proportion to how much a partner’s content actually gets used.

The WSJ flagged this point specifically. The standard contract between major AI companies and news publishers bundles a guaranteed flat fee with broad access rights. What Apple is proposing isn’t that standard.

And Apple isn’t alone here. The same direction is already showing up in multiple places at once.

Microsoft is piloting its Publisher Content Marketplace (PCM). As of February 2026, AP, Business Insider, Condé Nast, Hearst, USA Today, and Vox Media are on board, and the pitch is usage-based reporting that pays out “according to value delivered.”

Perplexity built a $42.5 million distribution pool for Comet Plus. Publishers get 80% of subscription revenue, Perplexity keeps 20%, and payout is tied to three usage signals: direct visits through the browser, citations in search answers, and use by an AI assistant completing a task.

Cloudflare is doing this at the infrastructure layer outright. The agent wallets and toll booths I covered in the last issue were exactly this story.

But there’s a distinction that needs to be drawn here. I think it’s the crux of the whole matter.

What Apple needs for Siri isn’t training. It’s closer to grounding ¹ — pulling in the latest news and current information on the fly, in real time, to answer a question. Training happens once and it’s done. Grounding happens fresh every time a query comes in. That means it’s an act whose occurrences get counted.

So a usage-based model is a natural fit for news. New articles come out every day, and each one gets referenced anew, every time.

The problem is that once this contract template becomes the standard, it spreads to the next cubicle over. What happens when content that gets trained on once and rarely called again — books, papers, lecture notes — gets pulled into the same template? The value contributed to training was never meant to be measured by “when it’s used,” but the measuring device only counts calls.

Mentat’s business model shows this shift is already underway. The company brokers rights for 38 publishers and over 21,000 books, and it presents publisher revenue in two forms: a lump sum for dataset licensing, and recurring revenue based on citation frequency. The $656 I received was the former. Which means the latter column is already printed into the contract.

---

This Experiment Already Ran Its Course Once

What I want to talk about here isn’t a prediction. It’s the results report from an industry that already went through the same sentence swap.

Music.

As the recording industry moved from sales to streaming, it changed the sentence of value from “the album sold” to “the track got played.” And on March 18, 2024, Spotify announced a policy: tracks that fall short of 1,000 streams over the trailing 12 months no longer earn recording royalties.

Spotify’s explanation is reasonably sound. The share of the total royalty pool taken up by sub-threshold tracks is about 0.5%, and per track, that works out to an average of $0.02 a month — roughly ₩30. But since intermediary distributors and banks set minimum withdrawal amounts anywhere from $2 to $50, with fees running $1 to $20, that money never even reaches the artist — it evaporates inside the pipeline. The logic is: better to pool it and send it to artists who are actually active.

The logic is hard to refute. But look at the outcome. Metered pricing created a zone where the amount converges toward zero — and then simply cut that zone off. And the tracks in that zone are the overwhelming majority. If 99.5% of all streams come from tracks that clear the threshold, flip that around and it means the vast majority of tracks were splitting just 0.5% of streams among themselves. The long tail² doesn’t survive metered pricing. More precisely: even if it survives, it gets cut away because it shrinks below the cost of settling it.

But this experiment is actually older than that — and it ran in a far gentler form.

There’s a system called Public Lending Right³, which compensates authors based on how many times their books are borrowed from libraries. It started in Denmark in 1946, is now run by 33 countries, and the UK adopted it in 1979. According to World Intellectual Property Organization data, roughly 24,000 creators in the UK receive compensation under it every year.

Here’s the interesting part. The UK system caps annual payouts at £6,600. Why put a ceiling on it at all? Because if you divide the pool purely by loan count, a small handful of bestselling authors sweep up the funds. A system that has run metered pricing for nearly 80 years is using an artificial mechanism to correct for exactly the part that pure metering can’t handle.

To sum up: pricing by usage isn’t a new experiment. And every place that’s tried it before has, without exception, discovered that pure metered pricing doesn’t work on its own — and has bolted on some form of correction. Spotify drew a floor and discarded what fell below it; the UK’s PLR drew a ceiling and pressed down what rose above it. The directions are opposite, but both are, in the end, the result of giving up on “pay exactly what’s earned.”

Right now, AI content contracts don’t have that correction mechanism yet.

---

Oswald’s Lens

I’ve worked on several projects where the task was converting flat-rate subscriptions into usage-based pricing as part of a GTM strategy. Every single time, one pattern kept resurfacing. A shift to usage-based billing looks like a pricing change, but what it actually does is move power.

Under a flat-rate model, the seller sits at the negotiating table. The fight is over price. Once you move to usage-based billing, the object of negotiation shifts from price to measurement: what counts as usage, who counts it, and who owns the meter. And the answer to all three questions almost always sits with the buyer.

When Perplexity says it has defined three usage signals, Perplexity is the one who defined those three. When Microsoft says it will pay “based on value delivered,” Microsoft is the one who calculates that value. When Apple says it will pay when content is used, only Apple knows whether it was used.

Working with data for years has trained a reflex into me: whenever I look at a metric, the first thing I check is who owns the meter producing that number. Whoever holds the meter ends up holding pricing power over that entire market. In usage-based content deals, the meter sits 100% inside the AI company. I have yet to see a single case where a usage report was audited by a third party.

Last week I wrote about a Chinese delivery platform handing out free camera-equipped helmets to its riders. The line I wrote there applies just as much here. Whether it’s labor or creative work, the danger point isn’t when compensation gets low — it’s when the transaction stops looking like a transaction.

$656 might be a small number. But at least it was a transaction with a receipt. Because the amount was written down, you could ask for a raise next time, compare it against other authors, or push the publisher on the revenue split. The act of negotiation could actually take place.

Once you move to call-based settlement, that sentence becomes hard to write. I have no way of checking how many times my book was referenced this month. When the number reads $0, I can’t tell whether that’s because nobody looked for it or because the meter simply didn’t count it.

So I don’t think what creators should be demanding right now is a higher per-unit rate. It’s two lines: a guaranteed minimum and access to the usage report. Not the kind of floor Spotify draws for you, but the kind of floor creators demand for themselves, first.

---

Closing

The core comes down to three lines.

• Until now, AI training-data contracts have priced “existence.” HarperCollins’s $5,000 per book, Anthropic’s settlement of $3,000 per work, and my own $656 — all follow the same grammar.
• The metered pricing that Apple, Microsoft, and Perplexity are shifting toward prices “calls” instead. That structure fits news content, but once it spreads to books and papers, most of them fall into the $0 bracket.
• The results of this experiment are already in. Spotify drew a floor and cut off the long tail; the UK’s Public Lending Right held a ceiling for 80 years. Today’s AI contracts have no such corrective mechanism.

If you happen to be negotiating an AI-related contract with a publisher or agency right now, I’d suggest checking just one sentence this week. Does it include a clause stating that “compensation shall be calculated based on usage,” and do you have the right to verify that usage? Right now, a lot of contracts still leave that back section blank.

💬 Have you ever faced, in contract form, the terms under which your writing, photos, lectures, music, or code get used for AI training or services? I’m less curious about the dollar amount than about exactly what sentence appeared in the contract. Was it a flat fee or usage-based? Was there any way to verify usage? Tell me in the comments. If enough cases come in, I’ll organize them into types from the perspective of Korean contract practice in the next issue.

---

📨 If you know someone who creates and sells writing, photos, music, or code, please pass this piece along to them.

---

The English draft matches the Korean source well. Here is the corrected version with minor surgical fixes:

References & Further Reading

Primary sources

• “Apple in Talks to Pay Publishers to Improve AI-Powered Siri,” The Wall Street Journal, August 12, 2026. ··· This is where today’s piece starts. The key line: “the standard contract guarantees a fixed sum — Apple’s doesn’t.” Paywalled.
• MacRumors, “Apple in Talks to Pay Publishers for News Content to Power Siri AI”, August 12, 2026. ··· I’d recommend this if you can’t get past the WSJ paywall. It summarizes the nine-figure budget and the variable-payment structure.
• 9to5Mac, “Report: Apple seeks publisher deals to give Siri AI better access to current events”, August 12, 2026. ··· The headline alone makes clear that what Apple wants isn’t training data but “access to current information.” It ties directly into today’s discussion of grounding.
• Transparency Coalition, “HarperCollins AI deal with Microsoft sets first public price for training data”, November 2024. ··· Lays out every term — $5,000 per title, a 50-50 split, a three-year term, and a 200-word output cap. Still the most useful price benchmark out there.
• Microsoft Advertising, “Building Toward a Sustainable Content Economy for the Agentic Web”, February 2026. ··· Microsoft’s own explanation of the usage-based payment logic behind PCM. Worth checking directly how far the phrase “value delivered” is actually defined.
• Digiday, “How Perplexity’s new revenue model works, according to its head of publisher partnerships” ··· Covers the $42.5 million pool, the 80-20 split, and the definitions of the three usage signals.
• Spotify Loud & Clear, “Why don’t songs with fewer than 1,000 annual streams earn recording royalties anymore?”, March 18, 2024. ··· The source for the $0.02 monthly figure and the 0.5% pool. The clearest document I know of showing where the floor of usage-based pay actually gets drawn.
• Publishing Perspectives, “Court Grants Final Approval to Landmark $1.5 Billion Anthropic Settlement”, July 2026. ··· Explains how the roughly $3,000-per-work figure was calculated, and notes that the settlement carries no precedential weight.

Background

• Interview with Jim Parker, “The public lending right and what it does”, WIPO Magazine. ··· Traces the practice from its 1946 origins in Denmark to its spread across 33 countries, including the UK’s £6,600 cap. Read it as an 80-year clinical record of usage-based pay, and it gets a lot more interesting.
• Authors Guild, “AI Licensing for Authors: Who Owns the Rights and What’s a Fair Split?” ··· Lays out the revenue-split debate between authors and publishers from the authors’ side. If you’ve got a contract in front of you, start here.
• eMarketer, “Usage-based deals could reshape AI content monetization models” ··· Makes the counterargument that usage-based pay can disadvantage smaller publishers. Useful if you want to see the other side of today’s argument.

Related past issues worth reading

• The Bot Cloudflare Gave a Name and a Wallet ··· A look at what usage-based deals look like at the infrastructure layer. The contract and the tollbooth are two sides of the same problem.
• Why JD.com Gave Away Free Helmets ··· Tackles the problem of “a transaction that stops looking like a transaction” — from the labor side. Today’s piece is the creator-side version of that same question.
• Copy-Paste Gets Caught, Copilot Doesn’t ··· Covers how copyright loses its teeth once source-tracking breaks down. When usage can’t be counted, usage-based pay ends up on the same path.

---

📝 Glossary

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis, including business data management and business analytics, while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master's from Korea University's Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

1. Grounding: A method by which an AI fetches up-to-date external material on the fly and uses it as the basis for its answer. Unlike training, where knowledge is baked into the model ahead of time, grounding pulls in fresh references every time a question comes in — which is why “how many times it was used” can actually be counted. ↩

2. Long tail: The large number of items that individually generate little but collectively add up to a lot, trailing behind a small number of hits. In the book market, think of the hundreds of thousands of titles behind the bestsellers. ↩

3. Public Lending Right (PLR): A system that compensates authors a set amount each time their book is borrowed from a library. Because it prices usage rather than sales, it’s the oldest form of the usage-based pay we’ve been discussing today. ↩