Readers have developed a new literacy skill nobody asked for: detecting AI-written text. The tells are real, they're consistent, and they have a common root cause — language models are trained to produce the most probable next word, and personal voice is precisely the stuff that's improbable.
The fingerprints
- Uniform sentence rhythm. Human writers produce jagged length distributions — a forty-word sentence, then a fragment. Models regress toward comfortable medium-length sentences, and the evenness reads as hum.
- Compulsive balance. "While X has advantages, it's important to consider Y." Models hedge because bland balance is the safest high-probability path. People with actual opinions are lopsided.
- Tell-tale vocabulary. Certain words appear in AI text far above their natural rate — delve, crucial, landscape, leverage, foster. None are wrong; the density is.
- Structural tics. The rule-of-three everywhere. "It's not just X, it's Y." Tidy intro-body-conclusion shapes on posts that a human would have ended abruptly.
- Frictionless abstraction. No specific dollar amounts, no named tools, no Tuesday afternoons, no sentences that could embarrass the author. Texture is risk, and default generation avoids risk.
Why prompting harder doesn't fix it
"Make it sound more human" produces performed casualness — the model's statistical average of what casual looks like, which is its own recognizable flavor. The problem isn't effort; it's the target. Without a specific person's patterns to aim at, the model can only aim at the middle of everyone, and the middle of everyone is what readers are detecting.
What actually moves the needle
The fixes all amount to replacing the average with a specific target. Provide real examples of one person's writing and the probability landscape shifts: their sentence rhythm, their vocabulary tier, their willingness to claim things flatly become the high-probability path. This is why example-driven voice systems produce drafts that survive the sniff test when instruction-driven prompting doesn't — the model is no longer guessing what a human sounds like, it's matching what this human sounds like.
The stakes are trust, not aesthetics
None of this would matter if readers didn't care. They do — not because AI assistance is disqualifying, but because generic text signals that the author didn't think the audience was worth their actual attention. The tell isn't "a machine touched this." The tell is "nobody's fingerprints are on this at all."