AI has disrupted the main cue: the text no longer proves what the author was thinking

For three thousand years, text has served as a “proof-of-thought”: if a person writes coherently and persuasively, it means they have a firm grasp of the subject. Writing well used to be difficult, and this barrier weeded out those who didn’t understand. LLMs have broken this signal—the quality of the text no longer proves anything, and the burden of verification has been shifted entirely to the reader. Juliette Ryan’s article on Substack examines this from a perspective that practitioners rarely discuss: the problem isn’t plagiarism, but cognitive debt. Writing well used to be difficult. To produce coherent text, you had to condense a three-dimensional thought into a linear stream of words that another person could follow. This process isn’t a side effect—it’s a mechanism that reveals gaps in your own understanding. When you ask Claude to “polish” a text, you’re skipping precisely the stage at which you realize that you don’t actually understand half of the topic. The author of the article shares an observation: an acquaintance of hers spent a year writing Facebook posts using AI “just for polishing” and sincerely believed the thoughts were her own. A Pangram check revealed 100% AI generation. The person wasn’t lying—they simply couldn’t distinguish their own thoughts from the text the model had assembled from their notes. That is the crux of the problem. The fluency of the text creates a false sense of understanding—not only for the reader, but also for the author. You read a polished paragraph, a light bulb goes off in your head, and you think, “Yes, that’s exactly what I wanted to say.” But try to paraphrase it without the text in front of you—and you’ll find there’s no coherent framework in your head. The model filled in the gaps you didn’t notice because they were hidden behind beautiful phrasing. Awareness doesn’t work as a safeguard. The author emphasizes: knowing about fluency bias doesn’t turn it off. It’s a subcognitive reflex—you still trust polished text more than clunky text. The more AI-generated text you read, the more your internal gauge for “what the truth sounds like” shifts toward the model’s style. The most dangerous type is neither a novice nor an expert. A novice knows how much they don’t know. An expert knows that, too. The “competent middle-of-the-road” person is the dangerous one: someone with domain knowledge who has never taken it to the stage of synthesis. Previously, they lacked the skill or confidence to speak up. Now Claude provides both—it picks up on ideas, selects arguments, and offers encouragement. A person publishes a text and believes they’ve understood what they’ve written. Their readers believe it too—and spread it further. The practical takeaway for those who work with AI daily: don’t delegate the final stage of writing to the model. Prompting is unchecked thinking. Until you’ve laid out the entire chain of arguments in a long text without the crutch of AI, you don’t know whether you had a complete understanding or just a sense of it. The author of this article uses AI for everything herself—research, brainstorming, spell-checking, data analysis, and automation. But not for writing. And she doesn’t consider herself ready to write an essay until she starts catching AI in inaccuracies within her field. This is a strict but honest measure of maturity: a model is trustworthy only to the extent that you are an expert on the subject. If you can’t catch it making a mistake—you’re not ready to either entrust it with your text or publish the result under your own name. Source: Juliette Ryan — AI broke writing as proof-of-thought

Source: Hacker News - Newest: “AI” “LLM”