Fluency is not evidence of quality
AI-generated text is grammatically correct, well structured, and coherent, but none of that tells you whether the content is accurate or well reasoned. Fluency has become noise, and unusually convincing noise, because it exploits a habit readers developed when clear writing usually meant clear thinking. That habit is now a liability.
AI-generated text is fluent: grammatically correct, well structured, and coherent. That fluency is why AI writing sounds convincing, and it’s a problem, although not for the reason that’s usually given.
By fluency I mean how easily the text reads, which psychologists call processing fluency. The same phrase is also used for a person’s skill in working with AI, which is a separate idea with its own note. Here it means the text.
The usual worry is that fluent text gives away AI authorship. The bigger problem is that readers take fluency as a sign of quality, a sign that the idea underneath is sound. People judge a statement that’s easier to read as more likely to be true, even when the statement itself is identical (Reber & Schwarz, 1999).
Fluency says nothing about whether a claim is right, whoever or whatever produced it. A language model produces well-formed prose whether or not its content is accurate, well reasoned, or even meaningful, so the words cohere and the argument moves forward with nothing on the surface to show what lies beneath.
The practical implication is that fluency needs to be filtered out entirely and treated as irrelevant to the question of quality. Confident delivery has no bearing on whether a speaker’s facts are correct, and fluency has no bearing on whether a claim is sound. The question to ask of a fluent text is whether it’s true, and how well it reads tells you nothing about the answer.
The response that’s emerged in AI literacy frameworks is to teach the ability to evaluate AI outputs: spotting hallucinations, checking citations, identifying logical gaps. The AI Fluency Framework calls this discernment (Dakan et al., 2025). This is an argument I’ve made myself, and it was reasonable at the time.
But it assumes the quality gap between AI and expert output will stay where it is, and I don’t see any sign that it will. The models available now are materially better than those available two years ago, and the trajectory is consistent. What happens when AI output is PhD-level across every artefact we might care to measure, with citations accurate, reasoning sound, and evidence appropriately qualified?
Evaluating model output doesn’t work in that world, because it’s a temporary fix that relies on a gap that is closing. Today an AI might hallucinate a reference if the user doesn’t know what they’re doing; tomorrow it won’t. Learning to spot today’s failure modes leaves open the question of what you’ll do when those failure modes are gone.
Output quality isn’t a stable criterion, so treating fluency as signal, at any level of quality, is a dead end.
Reading fluent prose without being swayed by its fluency runs against deep cognitive habits. The brain finds coherent text convincing. And that instinct isn’t irrational; in student writing especially, fluency was a reasonable trace of genuine engagement with ideas. A student who wrote clearly had usually thought clearly. That relationship no longer holds, and the instinct it trained is now a liability.
Fluency and output quality are no longer signals. They’re noise, and unusually convincing noise at that.
Sources
- Dakan, R., Feller, J., & Anthropic. (2025). The AI fluency framework. Anthropic. https://www-cdn.anthropic.com/b383cf6baddbfc72fdf8b0ed533a518e2872d531.pdf
- Reber, R., & Schwarz, N. (1999). Effects of perceptual fluency on judgments of truth. Consciousness and Cognition, 8(3), 338–342. https://doi.org/10.1006/ccog.1999.0386