Fluency is not evidence

This is the module that matters most. It is part of the Lrnon Certificate route: finishing it earns your Module Record, and the Certificate needs an assessment on top.

E1·L7 explained why an assistant sounds confident when it is wrong — it is producing likely text, and likely text has the cadence of knowledge whether or not the content is right. That explains the machine.

This module is about the harder half: why that works on you, and what to do about it.

The signal that stopped meaning anything

For all of human history until very recently, fluent, well-organised, confidently-specific writing on a subject was evidence that someone had put work in.

Not proof. But a decent heuristic, and your reading instincts are built on it. Effortful writing indicated an effortful writer.

Generated text severs that link entirely while producing the identical signal. The fluency is now free. Your instinct has not updated, and will not, because it is not the kind of thing that updates by being told.

That is the whole problem, and noticing it is most of the defence.

Four things you mistake for reliability

Confidence. No hedging reads as certainty. In generated text it reflects nothing — the same system produces equally unhedged prose for a fact and a fabrication.

Specificity. Names, dates, section numbers, percentages. Specificity feels like a claim somebody checked. Detail is exactly as cheap to generate as vagueness, and fabricated citations are specific by construction.

Structure. Numbered points, clean headings, a balanced conclusion. Reads as disciplined thinking. It is a format.

Fluent handling of your objection. You push back, and it responds smoothly and reasonably. That feels like it has thought about it. Smooth responsiveness is what the system does regardless of whether the new answer is any better than the old.

That last one catches the most careful people, because they mistake their own diligence in objecting for having tested something.

Do not become uniformly sceptical

The wrong lesson is to doubt everything. It does not survive contact with a working week — it is exhausting, so it gets abandoned entirely, usually within days.

Scepticism is a budget. Spend it where being wrong costs something, which is a minority of what you read, and let the rest go unchecked knowing that you chose to.

L2 is that triage. The rest of the module is what to do once a claim earns attention.

The uncomfortable part

You have already accepted things this way. Not hypothetically — everyone reading this has repeated something from an assistant that turned out to be wrong, and usually did not find out.

That is not a character failing. It is the predictable result of a reliable signal being broken. The response is not to feel bad about it; it is to build a habit that does not depend on your instincts being correct, because they are not.

Try it now (5 minutes)

Think of something you have repeated from an AI answer in the last month.

Check it now. Whichever way it goes, notice what made you believe it at the time — that specific feature is the one to distrust from here on.

Check your understanding

1. Why does fluent text suppress scepticism? It is usually more accurate Fluency has always signalled effort and expertise, and generated text breaks that link while keeping the signal People do not read carefully

2. Specificity — names, dates, percentages — indicates: That someone checked the claim Nothing about accuracy; detail is as cheap to generate as vagueness A more capable model

3. The right response to all this is: Doubt everything an assistant says Spend scepticism where being wrong costs something Only use assistants for creative tasks

Recap

Fluency, confidence, specificity, structure and smooth handling of objections are all free to produce and none of them evidence anything. Your instincts still read them as reliability and will not update on their own. So do not try to doubt everything — build a habit that does not rely on instinct, starting with deciding which claims deserve checking at all.