How Common Are AI Tells in Online Content?

There are several tells that a text may have been written by AI. These are certain types of phrases, words and punctuation that AI tools tend to use much more frequently than human writers. And while none of these tells are perfect predictors of AI authorship, and therefor AI detection tools will always deliver some false positives, they tend to give a good enough indication of AI authorship when applied to a large enough sample.

These tells include for example the em dash, Oxford commas and certain words that show up at higher rates in AI-written texts. As always with AI, the explanation as to why this is the case lies in the training data. As AI models are trained on large datasets of human writing, journalistic and academic writing, which represents a large share of publicly available text, tends to be overrepresented in training data. That explains why things like the em dash or certain words that may be less common in non-professional writing, tend to be overrepresented in texts written by AI.

According to an analysis of 490,000 online texts conducted by the Pew Research Center, the use of the em dash in texts published after the release of ChatGPT has roughly doubled over the past three years. Use of the Oxford comma has increased by 63 percent, and the occurrence of AI-typical words (e.g. pivotal, underscore, vibrant) has more than doubled. The use of negative parallelism, i.e. “it’s not just this, but also that”, has nearly tripled since early 2023, though it’s still rarer than the other AI tells used in Pew’s analysis.

Looking at the overall trend, however, explains why these are widely considered AI tells. Since the release of ChatGPT, all of these stylistics quirks have started showing up more and more often in online content, closely mirroring the increase in text likely written with the help of AI.