The Tell-Tale Words: Catching AI in the Act of Writing

Every human writer has their quirks—a favorite adjective, a specific sentence rhythm, a tendency to lean on certain transitional phrases. As it turns out, artificial intelligence is no different. AI models develop their own stylistic crutches, and these digital fingerprints are becoming our best tools for navigating an increasingly synthetic information landscape.

Recently, a fascinating analysis surfaced on the LessWrong forum that perfectly illustrates this modern cat-and-mouse game. Researcher Linch Zhang examined a document titled Magnifica Humanitas, attributed to a "Pope Leo XIV." The document's purported theme was a solemn warning about the impacts and dangers of AI on humanity. However, the medium might just have been the message.

Zhang ran the text through an AI detection tool called Pangram. The results were startlingly ironic: several paragraphs were flagged with a 40 to 100 percent probability of being AI-generated, while the opening chapter scored a 62 percent likelihood.

The most intriguing aspect of this analysis isn't merely the irony of using an algorithm to draft a warning about algorithms; it's how the text was caught. Zhang noted that the document exhibited specific linguistic tells, most notably an unusually high frequency of the word "genuinely." In the niche world of AI forensics, "genuinely" is a known red flag—a stylistic habit frequently exhibited by Anthropic’s Claude model.

When large language models are trained and aligned for safety and helpfulness, they inadvertently adopt certain conversational tics. They might overuse words like "delve," "robust," or "genuinely" in their attempt to sound balanced, objective, or empathetic. Because these models generate text by predicting the next most statistically likely word based on their specific training weights, these patterns become baked into their output.

This incident highlights a profound shift in how we must consume information. We often assume that AI-generated text is perfectly smooth and devoid of personality, making it indistinguishable from human writing. In reality, AI has a very distinct "personality"—it just requires a statistical lens to see it.

Whether the Magnifica Humanitas document was a deliberate satirical thought experiment or a genuine attempt at automated drafting, it serves as a crucial reminder. As the line between human and machine authorship continues to blur, maintaining trust will require more than just looking at the prestigious title of the author. It will require a keen eye for the digital tells hiding in plain sight.

Key Points

• An analysis of a document warning about AI dangers revealed it was likely partially written by AI.
• AI detection tools like Pangram flagged sections of the text with up to 100% probability of machine generation.
• The AI was identified through linguistic fingerprints, specifically the overuse of the word 'genuinely,' a known trait of Anthropic's Claude.
• AI models develop unique stylistic habits based on their training and alignment processes.

Why It Matters

As AI becomes capable of mimicking authoritative voices, understanding how models leave behind linguistic fingerprints is essential for verifying the authenticity of the media we consume.

Sources:

• Did the Pope use AI to write about the dangers of AI? — The Verge - AI