AI Slop — An Honest Note About This Website

Let me get this out of the way: almost all the text on this website was written by AI. The 535 aspect descriptions. The examples. The verification steps. The app explanations. The blog articles. This blog article. If you came here expecting hand-crafted prose by a domain expert burning the midnight oil — I'm sorry. That person exists, but he was busy building the platform instead of writing the copy.

The Uncomfortable Fact

TellDear uses Claude (by Anthropic) as its primary writing tool. The process works like this: I design the structure — which aspects to include, how they relate, what the verification steps should test — and then I prompt the AI to generate the prose. I review it, adjust it, sometimes rewrite it, sometimes throw it out and start over. But the first draft? Almost always AI.

In internet parlance, this makes TellDear "AI slop." That term — coined to describe the flood of low-effort, AI-generated content clogging search results and social media — is meant to signal that something isn't worth your time. That it was produced without care, without intent, without a human giving a damn.

I want to push back on that framing. Not because I'm defensive, but because the framing itself is sloppy.

The Alternative Wasn't Better Content

Here's the arithmetic. TellDear catalogs 535 aspects across six dimensions. Each aspect has a description, examples, counter-strategies, and verification steps — that's over 1,500 verification steps and 1,300 examples total. There are 24 apps, each with explanations and UI text. There's a knowledge graph with 850 relationships. Everything exists in English and German.

One person built this. One. The taxonomy design, the code, the database schema, the graph visualization, the app concepts, the deployment pipeline, the educational philosophy behind the whole thing — all one person.

If I had insisted on writing every piece of content by hand, TellDear wouldn't have better content. It would have no content. Or, more precisely, it would have about 30 aspects with lovingly hand-crafted descriptions, and the other 420 would be a bullet point and a Wikipedia link. The alternative to AI-generated content was not artisanal content. It was an empty shelf.

The Irony, Obviously

Yes, I see it. A platform dedicated to teaching people how to detect manipulation, evaluate arguments, and think critically about the information they consume — and it's using a large language model to generate its educational material. A tool that helps you spot bullshit, partly written by a machine that is, architecturally, a very sophisticated bullshit generator.

I don't think this irony is a gotcha. I think it's worth sitting with.

Large language models don't understand truth. They produce statistically plausible text. When I prompt Claude to describe the ad hominem fallacy, the result is accurate not because Claude understands logic, but because the training data contains enough accurate descriptions of ad hominem that the output converges on correctness. This works remarkably well for well-documented concepts. It works less well for novel analysis, which is why the apps use the LLM differently — as an analytical engine guided by TellDear's taxonomy, not as a source of ground truth.

The distinction matters: the structure is human. The taxonomy, the six dimensions, the relationships between aspects, the decision about which verification steps actually test for a pattern — those are design decisions made by a human who spent years reading about argumentation theory, cognitive science, and rhetoric — a fascination that, if we're being honest, started long before anyone called it a "project." The AI fills in the prose around a human-designed skeleton. It's a ghostwriter, not an architect.

What You Should Actually Worry About

The real risk of AI-generated educational content isn't that it's "slop." It's that it's confidently mediocre. AI text tends toward the consensus view. It smooths out edges. It rarely says "this is contested" or "experts disagree about this." It produces text that sounds authoritative regardless of whether the underlying claim is settled science or active debate.

I've tried to mitigate this by designing verification steps that are concrete and testable, by linking aspects to each other through the knowledge graph so you can see the bigger picture, and by building apps that let you apply the concepts to real-world texts rather than just reading about them. But I won't pretend the mitigation is perfect. Some descriptions are probably too confident. Some examples might be slightly off. If you find one, tell me.

A Challenge

TellDear has a Source Evaluator. It analyzes texts for rhetorical transparency, logical structure, framing, statistical claims, and discourse quality. I genuinely invite you to paste this article into it and see what comes back. Is this a well-argued case for AI-assisted content creation? Or is it a rationalization dressed up as honesty? Is the "I had no choice" framing a legitimate practical argument, or a false dilemma?

I designed a platform that gives you the tools to answer those questions. The least I can do is submit to the same scrutiny.

Judge for yourself. That's the whole point.