The Expertise Paradox: The Same Post That Builds Your Authority Also Trains Your Replacement

TL;DR. To be treated as an authority, you have to publish specific, hard-won expertise in public, because vague thought leadership builds nothing. But that same specificity is the most valuable material for AI systems that commoditize expertise, and they rarely send readers back. A 2025 Pew study found only 1% of visits to a page with an AI summary produced a click to the cited source. The real question is not whether to publish, but which layer of your knowledge you make reproducible.

Part 1 of a three-part series on what happens to expertise once it goes public in the AI era.

To be treated as an authority, you have to publish specific, hard-won expertise in public, because vague thought leadership builds nothing. But that same specificity is the most valuable material for AI systems that commoditize expertise, and they rarely send readers back. A 2025 Pew study found only 1% of visits to a page with an AI summary produced a click to the cited source. The real question is not whether to publish, but which layer of your knowledge you make reproducible.

The Argument That Keeps Happening on Our Team

The same debate resurfaces every time someone on our team drafts a genuinely good post.

One person has cracked something real. A method, a worked example, the actual reasoning behind a decision that took years to earn. They start writing it up, then stop. "Do we really want to hand this over?" they ask, half-joking about needing to protect what they know. The other half isn't joking at all.

Someone else pushes back immediately. Showing the work is the only way to be taken seriously. Nobody trusts an expert who won't demonstrate the expertise. If you hedge, you sound like everyone else.

Both people are right. That is the problem.

We've started calling this the expertise paradox, and we want to name it plainly before anyone pretends it away. The same post that builds your authority also trains your replacement. The specific, detailed, reproducible material that earns you a reputation is exactly the material that feeds the systems commoditizing that reputation. You publish to be recognized. In doing so, you also contribute a training example, usually without any say in the matter.

This is not a case for silence. Silence has its own cost, and we think it's the worse one. It is a structural tension, and anyone building a public voice right now has to work inside it consciously.

Why Vague Thought Leadership Builds Nothing

Start with how authority actually forms, because the mechanism matters.

A reference voice is not built on opinions. It's built on specificity. The post that gets cited, saved, and forwarded is the one that shows the exact steps, names the tradeoffs, and reveals the reasoning most people keep private. Vague thought leadership ("align your strategy," "embrace the shift") builds nothing because it's indistinguishable from a thousand other posts. Nobody hires the person who states the obvious with confidence.

Specificity is the whole game. The worked example. The number you measured yourself. The counterintuitive thing you learned the hard way.

Here's the trap. That same specificity is what transfers cleanly into a model.

A vague claim carries almost no reusable information. A precise method, with steps and conditions and edge cases, is a near-perfect training signal. The more citable your writing, the more extractable it is. Those two properties are not in tension inside the content. They are the same property, viewed from two sides.

And the transfer is neither rare nor visible. The Data Provenance Initiative ran a large-scale audit of AI training datasets, covering 44 data collections and more than 1,800 fine-tuning text datasets. The audit found that licensing information was omitted more than 70% of the time and mislabeled more than 50% of the time across popular dataset-hosting sites.

Sit with that. Most of the time, nobody can even reconstruct where the training material came from or under what terms. This isn't a fringe scenario. It's the default condition of the pipeline. When you publish your best specific work, you should assume it can enter these systems with no traceable consent and no attribution attached.

The Reciprocity Gap Is the Real Story

For two decades, the deal was implicit but reliable. You publish something useful, it gets found, and the finding sends people back to you. Traffic. Backlinks. A name people start to recognize. The value you gave away came back as reputation and reach.

That loop is breaking, and the data is blunt about it.

A 2025 Pew Research Center study looked at the browsing behavior of 900 U.S. adults. When an AI-generated summary appeared above the search results, only 8% of users clicked a traditional web link, compared with 15% when no summary was present. The summary roughly halved the click-through. Worse: just 1% of visits to a page with an AI summary resulted in a click through to a cited source.

One percent.

Read that as an economic signal, not a news item. Being used by an AI answer does not translate into being visited. Your expertise gets absorbed into the response, synthesized, delivered, and the reader's need is met before your name ever enters the picture. The consumption is real. The reciprocity is gone.

This is the spine of the whole paradox. The old bargain assumed that giving away knowledge bought you visibility, and visibility bought you everything downstream. Strip out the return traffic and you're left publishing your scarcest asset into a system that neither pays you nor points to you. You still have to publish to build authority. You just can't count on the loop closing the way it used to.

The Stakes Are Priced Now

For a long time you could wave this off as abstract. What's the harm, really, in a model reading your blog?

That argument got harder in 2025. Anthropic agreed to pay roughly $1.5 billion to settle a class-action lawsuit brought by authors whose books were used to train its models without permission. The settlement worked out to about $3,000 per book across an estimated 500,000 works. The authors' lawyer called it "the largest copyright recovery ever."

We're not citing this as a courtroom drama. We're citing it as a valuation event.

Before this, "your published work becomes training data" was a concern with no price tag. Now a number exists. Imperfect, contested, specific to books rather than blog posts, but real. Someone put roughly $3,000 on the act of absorbing a single published work without consent. The fact that a market and a legal system produced any figure at all changes the conversation. It confirms the thing was worth taking, and it confirms that the people who made it were, until forced otherwise, not part of the transaction.

Most of us publishing on the open web will never see a settlement. Books had ISBNs, identifiable authors, and a plaintiff class. Your how-it-works post has none of that. The lesson isn't "you'll get paid." The lesson is the opposite. The value is now documented, and for almost everyone the mechanism to capture it does not exist.

Why This Hurts Small Voices Most

Here's the part that bothers us the most, because it inverts who can afford to play.

Large platforms can absorb the extraction. Their moat was never the content alone. It's the distribution, the brand, the network, the audience that already shows up. If a model ingests their material, they still own the relationship with the reader. The extraction is a rounding error against assets that don't sit on the page.

The independent expert has no such cushion. Neither does the small team or the niche publication without an existing distribution advantage. For them, the specific knowledge is the entire moat. It's the whole asset. And the paradox falls hardest on exactly the people who most need visibility, because they're the ones who have to publish their scarcest material to get noticed at all.

So the trade is lopsided. The big player trades content it can spare for visibility it barely needs. The small player trades its only defensible asset for a visibility that isn't guaranteed to arrive and, per that Pew number, increasingly doesn't. Same action, wildly different balance sheet.

We say this as a small team publishing specific work in public. We are describing our own exposure, not observing someone else's.

The Better Question: Method Versus Judgment

So the paradox doesn't collapse into "publish" or "don't." Framed that way, it has no good answer. Publish and feed the systems. Stay quiet and stay invisible. Both lose.

The useful move is to stop treating your expertise as one undifferentiated thing. It has layers, and they don't behave the same way once they hit the open web.

• Reproducible method. The steps, the framework, the worked example, the checklist. This layer is transferable by design. It's what makes your writing citable, and it's precisely what a model can lift off the page and reuse. You should expect this to be absorbed. That's not a reason to withhold it, but it is a reason to be clear-eyed that this is the commodity layer.
• Situated judgment. Knowing which method applies to this messy case, reading the context, sensing when the standard answer is wrong, deciding what matters under real constraints. This layer doesn't transfer the same way, because it lives in application, not in text. It's the part no model lifts cleanly, because it's exercised fresh each time against a specific situation.

That distinction is the whole reframe. Not "how much do I share," but which layer am I making legible and reproducible, and which layer am I keeping as the part that only shows up when I do the work.

This is where the series is headed. What actually gets distilled out of published expertise, and what stubbornly resists distillation, deserves its own examination. Part 2, What Actually Gets Distilled, takes that apart directly: what genuinely transfers from your expertise into a model, and what doesn't.

The Discipline This Requires

We want to end without tidying this up, because a clean resolution would be a lie.

The honest position is a discipline, not an answer. Publish deliberately. Know, for every piece, which part you're handing over as commodity and which part remains your moat. Treat the specific method as something you're choosing to make reproducible, with the full expectation that it will be absorbed and rarely credited. And protect the situated judgment not by hiding it, but by understanding that it can't be fully captured in the first place.

Most of all, stop treating visibility as a free action. It never was, and now the cost is measurable. A 2025 Pew study put reciprocal traffic near zero. A 2025 settlement put a price on absorption. A dataset audit showed the whole transfer runs mostly untracked. Those three facts describe a single durable pattern, not three separate headlines: published expertise enters a commons that trains the systems, and the loop back to the author is weak and getting weaker.

You still have to publish. Authority is built in public or not at all. Just do it knowing exactly what you're spending, and on which layer. The next piece is about telling those layers apart.

FAQ

Does publishing expertise online still build authority in the AI era?

Yes, but the payoff has narrowed. Authority still requires specific, reproducible expertise published in public, because vague thought leadership doesn't distinguish you from anyone else. What's changed is the return. A 2025 Pew Research Center study found that only 1% of visits to a page with an AI summary resulted in a click to the cited source, so the traffic and recognition that publishing used to earn are much less reliable.

What is the expertise paradox?

The expertise paradox is the structural tension that the same specific, detailed content that builds your authority is also the most valuable material for training the AI systems that commoditize that authority. Specificity is what makes writing citable and what makes it extractable. Those are the same property seen from two angles, which is why you can't build a reference voice without also contributing training examples.

How much does my published work sell for as training data?

There's no established rate for a blog post, but a reference point now exists. In 2025, Anthropic agreed to pay roughly $1.5 billion to settle a lawsuit over books used to train its models without permission, about $3,000 per book across an estimated 500,000 works. That figure applies to identifiable books with a plaintiff class, not to open-web posts, so most published work has a documented value but no mechanism to capture it.

Should independent experts stop publishing to protect their knowledge?

We don't think so, because silence carries a worse cost than publishing: invisibility. The sharper move is to separate your expertise into reproducible method, which you should expect to be absorbed, and situated judgment, which doesn't transfer the same way because it's exercised fresh against each specific situation. Publish the method deliberately, and understand that the judgment is the part a model can't lift off the page.

Why does AI training hit small publishers harder than large platforms?

Large platforms have moats beyond their content, including distribution, brand, and an audience that already shows up, so extraction barely dents them. Independent experts and small teams often have only their specific knowledge as a defensible asset, and they have to publish it to get noticed at all. That makes the trade lopsided: the same act of publishing costs a solo expert far more than it costs a platform.

Further reading

• Pew Research Center, "Google users are less likely to click on links when an AI summary appears in the results" (2025) — Only 8% of users clicked a traditional link when an AI summary was present, versus 15% without one; just 1% of visits to a page with an AI summary clicked a cited source. pewresearch.org
• Longpre et al., "The Data Provenance Initiative: A Large Scale Audit of Dataset Licensing & Attribution in AI" (2023) — Audit of 44 data collections and 1,800+ fine-tuning datasets found license omission rates above 70% and mislabeling rates above 50%. arxiv.org
• NPR, "Anthropic to pay authors $1.5B to settle lawsuit over pirated chatbot training material" (2025) — Settlement terms: roughly $3,000 per book across an estimated 500,000 works. npr.org