The Tell-Tale Heartbeat Is Fading #
Artificial intelligence writing's tell-tale heartbeat is fading, rendering traditional detection methods obsolete before they even gain traction. A massive study by Graphite recently mapped this evolving landscape, analyzing 90,000 AI-generated articles against 10,000 human-written ones. Researchers sourced the human articles from Common Crawl, all published before November 2022, then used GPT-4.1 to summarize them, instructing nine different AI models to expand these summaries into full articles.
This comprehensive dataset revealed 12,877 distinct linguistic "tells"—words, phrases, and patterns appearing at least twice as often in AI text. However, these linguistic signals are a rapidly moving target; less than half observed in one GPT version, such as GPT-5.6 Sol, carry over to the next iteration like GPT-6 Astra. This constant evolution means old detection methods quickly become irrelevant.
Consider the notorious em dash. Early models, including GPT and Gemini, overcorrected their initial punctuation habits so dramatically that they now use em dashes far less often than humans. GPT-6 Astra, for example, employs them 88% less often than humans did before ChatGPT even existed, rendering a lack of em dashes an unreliable signal for AI authorship.
Each Model's Quirky New Fingerprint #
A meticulous Graphite study analyzed 90,000 AI articles against 10,000 human texts, revealing major AI models rapidly develop and discard unique linguistic fingerprints. These idiosyncratic patterns are not static; they evolve with each new version, rendering older detection methods obsolete. Less than half the AI tells from GPT-5.6 Sol carried over to GPT-6 Astra.
GPT has shed its earlier "salesy" vocabulary, discontinuing words like "unlock" and "streamline." GPT-6 Astra now demonstrates a distinct preference for negation, often defining concepts by what they are not. Phrases such as "This is not simply a tool" surface 12 times more frequently than in human content, marking a measurable shift.
Claude, contrastingly, developed a pronounced
The Rise of a Global 'AI-ese' #
Graphite’s study revealed its most unexpected insight: all nine models tested now write more like each other than like actual people. This convergence suggests the emergence of a distinct 'AI-ese' language, a shared stylistic fingerprint that transcends individual model eccentricities. Rather than mimicking human expression, AI systems are increasingly mirroring each other.
Many assume newer models inherently sound more human. However, Graphite’s data challenges this directly. While Claude demonstrated progress towards human-like writing, GPT models moved in the opposite direction. Notably, GPT-6 Astra surprisingly exhibits 48% more AI tells than its predecessor, GPT-4.1, confounding expectations for advanced iterations.
This shared stylistic drift carries significant implications for AI detection. Identifying a specific model might indeed become more challenging as individual tells shift and blur across versions, complicating efforts to pinpoint authorship. Yet, as the 'AI-ese' solidifies, discerning AI-generated content as a category could paradoxically become easier.
Humans still differentiate themselves through genuinely unique writing. Graphite’s researchers highlight markers like exclamation marks and parenthetical asides, which humans use over 100 times more often than current models. For deeper insights into these evolving patterns, consult AI Tells — Five Percent - Graphite.io.
Enjoying this? Get one like it in your inbox each morning.
one email a day · unsubscribe in two clicks · no third-party tracking
The Unmistakable Messiness of Being Human #
While AI models converge on a standardized, often formal 'AI-ese', the Graphite study reveals a counter-narrative for human authorship. Our writing remains distinctly unique, characterized by traits AI struggles to replicate authentically. The data from their analysis of 10,000 human-written articles points to clear, powerful signals. Genuine personality emerges through specific stylistic choices. Human writers, unlike their AI counterparts, frequently employ:
- Exclamation marks
- Parenthetical asides
- Personal anecdotes
Humans use exclamation marks over 100 times more often than current AI models, a striking difference. These seemingly minor flourishes inject emotion, emphasis, and a conversational tone that AI, despite its advancements, still largely omits. They represent the subtle, often subconscious, fingerprints of a human mind at work.
This "messiness" — the unexpected interjection, the parenthetical thought, the personal story — is not a flaw; it's a feature. It grounds text in relatable experience and conveys a unique voice. As AI perfects its formal, often bland, style, the true differentiator for human writing remains its authentic, slightly imperfect, and emotionally resonant voice.
Frequently Asked Questions #
What are 'AI tells'?
'AI tells' are specific words, phrases, or stylistic patterns that appear significantly more often in AI-generated text than in human writing, acting as potential indicators of authorship.
Is AI writing getting harder to detect?
Yes and no. While individual tells for specific models change rapidly with each update, the study suggests AIs are converging on a shared 'AI-ese' style, making them collectively distinct from human writing.
Do newer AI models sound more human?
Not always. The Graphite study found that while Claude's word choices are getting closer to human writing, GPT-6 Astra actually has 48% more detectable 'AI tells' than its predecessor, GPT-4.1.
What is the biggest giveaway of human writing right now?
The data shows that humans use stylistic elements of personality and emotion far more than AIs. For example, humans use exclamation marks over 100 times more often, along with parenthetical asides and personal anecdotes.