Queryable Thinking
What happens when your writing becomes a callable knowledge service and the reader becomes a synthesiser.
I've been thinking about turning this blog into an MCP server, and the more I sit with the idea, the stranger it gets. Not technically, because the technical part is trivial. You expose the posts as structured resources, you register a few tools like "query the writer's view on X" or "retrieve the reasoning behind claim Y", and any AI system that speaks MCP can consume your thinking directly rather than scraping a rendered page. The strangeness is in what it does to the things on either end of the pipe. The writer becomes something different, and so does the reader. And what sits between them is no longer a website in any meaningful sense.
This is already happening in pieces. Schema markup, knowledge graphs, the Model Context Protocol being donated to the Linux Foundation in December 2025, and publishers quietly building APIs for AI consumers. What nobody seems to have said out loud yet is that the endpoint of this trajectory is the replacement of the website as the primary vehicle for published thought with something closer to a callable knowledge service, and once you're in that arrangement, the reader is querying. The writing craft that survives that transition is the one that understands it's producing queryable thinking, and looks meaningfully different from the one we have now.
From storyteller to ontologist
The writer in the current arrangement is a storyteller. You arrange your thinking into a linear narrative, you lead the reader through it, and you hope they reach the conclusion with you. The article is a performance of reasoning, and the reader's job is to follow along.
In a world where an AI system is the primary consumer of your writing, the performance stops mattering. What matters is whether your positions are addressable, whether your reasoning is legible as structure, and whether a system that queries "what does this person think about X" gets a useful answer. The craft shifts from sequencing insights in prose to clearly defining the ontology of your own thinking so it can be queried without you in the room. Your output stops being articles and starts being queryable thinking.
The technocratic misreading is "every writer should learn RDF," but the real shift is closer to what good thinkers have always done internally. You have positions. Those positions have dependencies on other positions. They have evidence attached to them. They have the conditions under which you'd change your mind. A storyteller hides most of this scaffolding behind the narrative, but an ontologist surfaces it, because surfacing it is the point.
The practical implication for someone writing on the internet is that the discipline of "what exactly is my position, what does it depend on, what would falsify it" becomes load-bearing in a way it wasn't before. You can still write long-form prose. But the prose is now a view onto a structure, and the structure is what gets consumed.
The reader becomes a synthesiser
On the other side of the pipe, something equally large happens to the reader. Right now, reading someone's writing means consuming their argument on their terms. You follow their narrative, and you adopt, provisionally, their framing. You can disagree later, but the encounter is serial.
Once writing is exposed as a queryable structure, the reader can compare multiple writers' positions directly against a question they actually care about. Not "what does Useful Machines say about AI adoption" in isolation, but "what do Useful Machines, Ben Thompson, Benedict Evans, and the Gartner research say about enterprise AI adoption, and where do they disagree, and on what evidence". The reader, or the system acting on their behalf, becomes a synthesiser by default.
This is the genuinely new thing. Synthesis across multiple sources has always been possible, but the friction was enormous. You had to read all the sources, hold them in your head, notice the disagreements, and trace the reasoning. Now that friction collapses, and the cost of a comparative, structured encounter with five writers' thinking drops to roughly the cost of one query.
The democratisation and homogeneity paradox
The upside of the shift is a kind of democratisation. Writers who couldn't get distribution in the traditional publishing economy can make their thinking discoverable to the systems that are now doing a lot of the reading. A structured, addressable body of work by someone with a sharp view on a narrow topic can compete with a media brand's output, because the AI consumer doesn't care about the brand; it cares about the quality and specificity of the position. In principle, the long tail of structured thinking gets heard for the first time.
The problem is that the same tool that enables this democratisation is systematically compressing the variance in how we think and write. The research here is detailed.
Doshi and Hauser, in a January 2025 study comparing human and LLM responses on standardised creativity tests, found that LLM outputs are measurably more similar to each other than human outputs are to each other. A 2025 arxiv paper on academic writing found that researchers who adopt GPT for revisions exhibit a marked shift toward GPT-style writing, and the convergence varies by demographic, with junior and non-native English researchers converging hardest. And on the mechanism, an October 2025 paper on mode collapse showed that post-training alignment reduces LLM diversity through RLHF dynamics that amplify common, majority-style responses, meaning the compression is not a temporary artefact; it's baked into how these systems are trained.
The paradox is straightforward once you hold both sides at once. The tools making it easier than ever to structure and publish your thinking are the same tools compressing the thinking toward the mean. Democratisation of publication and homogenisation of voice. The more people use LLMs to help organise their work for the AI-native content layer, the narrower the distribution of ideas in that content layer becomes.
I think this is the defining tension of the post-website era, and anyone writing online over the next five years will have to decide how they relate to it.
Meaning was always reinterpreted
The semiotic tradition, starting with Peirce and Saussure and running through Eco, has been clear for a century that meaning is never transmitted cleanly from writer to reader. It's reconstructed at the point of reception, using the reader's context and priors. Every reading is a reinterpretation.
What MCP and the content layer do is make that reinterpretation computational and visible. Where previously a reader's reconstruction of your argument happened privately in their head, now a system does it explicitly, with steps you can inspect. This is less of a break than it sounds. The interpretation has always been happening; we've just moved it from inside human heads into queryable infrastructure, and the discomfort many writers feel about this is less about meaning being distorted and more about losing control over the distortion.
I don't find this entirely reassuring, but it does suggest the right response is a different craft entirely.
Weaponised ignorance
A post like this needs to be honest about what goes wrong, and I'd not focus on the obvious misinformation story.
When reading becomes querying, the reader's ability to construct a partial, confident, completely wrong understanding of a topic increases dramatically. You can ask three queryable bodies of work a question, get a synthesised answer, feel that you've done the research, and walk away with a view that is confident, fluent, and built on exactly zero of the context that would make the view defensible. The AI intermediary will happily compose the synthesis. It has no mechanism for telling you that the question you asked was the wrong one, or that all three sources are operating from a shared blind spot, or that the answer you now hold depends on a premise none of them stated.
This is weaponised ignorance in a specific sense. Confident partial understanding becomes cheap to generate and nearly indistinguishable, from the inside, from actual expertise. The loss is in their calibration of what they know. A writer operating in this environment has to assume their readers will increasingly arrive at their work already holding confident positions that were assembled at query speed, and the craft question is what you do with that.
Workshops, not vending machines
The resolution, such as there is one, comes from getting clear about what these tools are for.
The dominant cultural framing treats LLMs as vending machines for answers. You put in a question, you get a finished response, and the transaction is complete. The framing I think survives the next few years is different: these tools are workshops for thinking. What you put in is a half-formed position; what you get out is a clearer articulation, a counter-argument you hadn't considered, a structural flaw you missed.
Writers who treat queryable thinking as a distribution channel for finished vending-machine answers will contribute to the homogeneity problem and probably not have much to say in five years. Writers who treat it as a way to expose the structure of their workshop, the partial views, the evolving positions and the explicit reasoning, will be doing something the current arrangement didn't make possible.
Queryable thinking is coming either way. What it's worth depends on who shows up and what they bring that the machine can't generate on its own.
References
Series:
Research cited:
- Doshi & Hauser, "We're Different, We're the Same: Creative Homogeneity Across LLMs" (arxiv 2501.19361, January 2025)
- "Divergent LLM Adoption and Heterogeneous Convergence Paths in Research Writing" (arxiv 2504.13629, April 2025)
- Zhang et al., "Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity" (arxiv 2510.01171, October 2025)