What happens when AI can ask a blog what it thinks, and readers stop reading start to finish?
I’ve been thinking about suggesting we turn Signal (our newsletter) into an MCP (model context protocol) server. MCP is the standard way AI tools connect to other software, so any AI assistant could ask Signal directly what we think about something, instead of reading the pages. Technically that’s easy. What’s strange is what it does to the people writing and the people reading.
We’ve already built a small version of this at Charlie Oscar. Understudy holds the writing voice of people here in a form an AI can use: a guide to how they write, the rules they follow, and posts they’ve approved. Mine came mostly from my own essays. An AI can draft in that voice and then check the draft against it. People are quite literally querying their own writing.
Writing that answers questions
A blog post tells a story. You lead the reader through your thinking and hope they end up where you did. When an AI is doing the reading, the story matters less than whether it gets a clear answer to “what does Charlie Oscar think about X?”
So the job changes. You have to spell out what you think, why, and what would change your mind, clearly enough that someone can ask without you around to answer. Understudy showed me how much that takes. My voice guide even has a section on where our “anti-slop” rules get my writing wrong. Whatever it produces, a person still has to be the last to read and edit it.
Readers start comparing
Readers change too. Today you read one writer and take their argument on their terms. Once blogs can answer questions, a reader can put the same question to Signal, Ben Thompson, Benedict Evans and Gartner, and see where they disagree. That was always possible, but it meant reading everything yourself. Now it’s one question.
Easier to publish, harder to sound different
There’s an upside. A small blog with a sharp point of view can hold its own against a big media brand, because the AI cares about the quality of the answer, not the name on it.
The catch is that the tools making this easy are also making everyone sound the same. AI answers to creativity tests are far more alike than people’s answers are. Researchers who write with AI start to sound alike. Part of the cause is how the models are trained: the people rating AI generated answers prefer familiar sounding text. It’s why Understudy checks for more than banned phrases. The real giveaway is sameness.
Confident and wrong
The risk I’d watch for is people feeling informed without being informed. If you ask three sources a question and get one smooth answer, you can walk away confident and wrong. The AI might not tell you that you asked the wrong question, or that all three sources share the same blind spot. I see the same thing running AI agents at work: a fluent answer is cheap, and proving it’s true is slow, laborious and finicky.
A workshop, not a vending machine
In my experience, people tend to use artificial intelligence like a vending machine: question in, answer out. It’s more useful as a workshop. You bring a half-formed idea and get back a clearer version of it, a counter-argument, or a flaw you missed. Understudy works that way. It checks a draft, says what’s wrong with it, and sends it round again.
If we do turn Signal into something AI can ask, I’d want it to share our working as well as our conclusions: the half-formed views and the reasoning behind them. Otherwise we’re only adding to the sea of sameness. What it’s worth depends on what we bring that the machine can’t produce on its own.
