Why It Matters
The erosion of identifiable linguistic 'watermarks' in AI-generated text has profound consequences for media, academia, a...
Can you spot AI writing? | The Economist
The Economist
The Signal
Recent analysis by The Economist reveals that AI writing style is shifting away from formerly reliable 'bot' markers like em-dash overuse toward more subtle patterns. As models train on human feedback, they are shedding obvious stylistic clichés in favor of longer sentences and scientific vocabulary, making detection increasingly difficult for casual readers.
The Case
Detection Heuristics
- Older AI giveaways like the phrases "deep dive" and "rich tapestry" are largely obsolete, as models have pivoted to utilizing more complex, scientific-sounding language.
- Contrary to popular belief, em-dashes are no longer a hallmark of AI; the study found ChatGPT uses significantly fewer em-dashes than human writers.
- A more effective heuristic for identifying AI-generated text is a lack of punctuation, specifically a comparative dearth of commas and semicolons.
- LLMs currently exhibit a distinct rhythm characterized by excessively long sentences and the repetitive overuse of the word "and," which often results in text that feels rambling or dull.
Trends and Limitations
- The study concludes that AI prose is steadily converging toward human patterns, driven by training on human feedback that prioritizes what readers find impressive.
- These markers are inherently unstable; because detection heuristics rely on current model behaviors, the signals remain moving targets that may lose predictive power as soon as models update.
The 1 Minute Signal Take
Do not rely on static 'AI-speak' clichés to filter content, as models are actively evolving to mimic human stylistic norms. Focus instead on structural patterns like sentence length and punctuation density, though recognize that even these are temporary tells in an environment of rapid model convergence.
Pro Analysis
Full analysis always available on Pro.
Time saved:
Tags
Written by: 1 Minute Signal Editorial Team