AI-Generated Writing: Why Readers Stopped Noticing

I read something like forty student emails a week. I’d like to tell you I scan each one for signs of a chatbot. I don’t. The question doesn’t cross my mind while I’m reading, and it only surfaces later, usually because somebody else raises it.

A new experiment by Jiaqi Zhu and Andras Molnar suggests I’m entirely typical, and that being typical here is the problem worth talking about.

Zhu and Molnar ran a pre-registered study with 647 US participants across four scenarios: a gratitude email from a close friend, a nanny application, a cover letter, and feedback on a work proposal. Each participant read a message from someone named Alex, then rated Alex on ten traits. What varied was what they were told about how Alex produced the message.

When participants knew Alex used AI, the ratings collapsed. Zhu and Molnar report a composite impression of -0.94, against 0.39 when the same message was described as human-written. That’s a Cohen’s d of 1.69, which is about as large as social psychology results get. Told outright that a message came from a chatbot, people judge the sender as lazy and insincere.

AI-Generated Writing

What Happens When Nobody Mentions AI-Generated Writing

The condition that matters is the fourth one, and it’s the condition earlier studies never ran. Zhu and Molnar gave one group no information at all. No mention of AI, no description of chatbots, nothing. Just a message and a rating scale.

That group rated Alex at 0.43. Statistically indistinguishable from the group told a human wrote it.

Hold that beside another number from the same study. Zhu and Molnar found that 45.7% of their participants had personally used AI to write or edit a message to someone else within the previous two weeks. Half the sample does this. Almost none of them assume anyone else does.

The authors then added a fourth wrinkle that I find genuinely clever. A separate group was told explicitly that the message’s origin was uncertain, with no further information supplied. Those ratings dropped to 0.13, a significant drop from the uninformed group at d = 0.51. Nothing changed except where attention pointed. As Zhu and Molnar put it, the differences “can only be explained by different levels of attention to AI involvement” (p. 22).

Even prompted, the correction stayed small. Working backward from the impression scores, the authors estimate participants implicitly assigned about a 19.4% chance the message came from AI.

Writing Was Never Only Writing

Zhu and Molnar frame all of this through signaling theory, and this is the part I keep turning over. A signal carries social information when it’s difficult to fake, verifiable, and costly to produce. Handwritten apology notes work because all three hold. Cover letters worked for the same reason.

Generative AI breaks all three simultaneously, which should have collapsed the signal value of written text almost immediately. Zhu and Molnar’s data say it hasn’t. Readers aren’t calculating the odds of AI involvement and concluding they’re low. They aren’t calculating at all.

That gap between what a signal is worth and what readers still treat it as worth cannot hold forever. It’s holding right now because attention hasn’t caught up.

What This Means for Assessment

Every teacher grading a stack of essays is standing in the no-information condition.

That’s the uncomfortable translation. We talk endlessly about AI detection tools and their false positive rates, as though the bottleneck were accuracy. Zhu and Molnar suggest the earlier bottleneck is attention. Most of the time, the question of authorship simply isn’t active in the reader’s mind, and no detector helps with a question nobody asked.

The finding also cuts against the reflex to make suspicion permanent. The uncertain condition in this study shows what happens when we push authorship to the front of a reader’s mind.

Judgments get harsher across the board, including for work that was entirely human. Anyone who has watched a colleague start reading every polished sentence as evidence knows exactly what that looks like in a classroom.

This is why I keep arguing that the interpretation we draw from student writing has to be built, never assumed. Kane’s framework asks what chain of inference connects a performance to a claim about competence, and Zhu and Molnar have identified a link in that chain nobody was watching. The inference now depends on the grader’s attentional state.

Hou and colleagues offer one alternative worth borrowing. They measured AI adoption as the cosine similarity between a group’s final report and its own chat transcript, sidestepping impressions completely. That’s evidence about process, not vibes.

Design the task so the process leaves a trail. Ask for drafts, prompts, revision decisions, and reasoning. Then the question of who wrote the polished sentence stops being the question you need answered.

Attention is a terrible foundation for assessment. It always was. AI just made that visible.


References

Hou, C., Zhu, G., Liu, Y., Sudarshan, V., Chong, J. L. L., Zhang, F. Y., Tan, M. Y. H., & Ong, Y. S. (2026). The effects of critical thinking intervention on reliance behaviors, problem-solving quality, and creativity during human-Generative AI collaborative learning. Computers & Education, 247, 105576.

Zhu, J., & Molnar, A. (n.d.). Blissful (a)ignorance: People form overly positive impressions of others based on their written messages, despite wide-scale adoption of Generative AI [Manuscript]. University of Michigan.

The AI in Education Research Digest

Keep exploring the research with me.

Get my reviews of new studies on AI, teaching, and learning, with thoughtful takeaways for your work as an educator.

Free to join. Unsubscribe anytime. Privacy policy.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top