Author(s): Shivang Raikar

Originally published on Towards AI.

A punctuation mark turned into a giveaway, users complained, and the newest models learned to hide it. The data says something stranger is going on.

For about two years, we all could spot AI writing by looking for one punctuation mark.

The article argues that the em dash became an “AI fingerprint” because early models absorbed it from common training text, but a look at large-scale data shows the opposite of fading: em dashes (and other “AI-sounding” markers) have increased across the web even after people learned to prompt them away. The author explains that this isn’t just a training-data issue: the em dash survives because it leaks from Markdown structure into plain prose, and different post-training/fine-tuning recipes can make models either suppress or still emit it. As human feedback and preference signals increasingly enter the training loop, model behavior shifts in response—so when one obvious tell (like em dashes) is suppressed, the fingerprint “moves” to less noticeable patterns (sentence rhythm, phrasing habits, vocabulary quirks). The conclusion is that “AI writing style” won’t stabilize into something easily memorized; instead, it will keep evolving along with the tuning process and our reactions to it.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI