Can You Spot AI Writing by Em Dash Alone? (The Math Says No)

If you’ve spent time in the online writing world, you most definitely know that usage of AI is ubiquitous. How were you able to arrive at that conclusion? There are several “tells”: the general tone or overuse of certain styles or varying levels of complexity that causes confusion, among others. One specific trait is the use (or, as online circles say, overuse) of em dashes ( — ). This was also recently investigated in a paper by an independent researcher.

Since the broad deployment of instruction-tuned models beginning in late 2022, users across writing communities, technology forums, and social media have identified em dash frequency as a marker of machine-generated prose — a “tell” that has become the subject of extensive online discussion, mockery, and ad hoc detection heuristics.

The paper argues that the em dash is a prose device as well as a markdown feature and that the Large Language Models (LLMs) trained via complex, multi-step processes mix them up, and then proceeds to discuss some experiments to verify the stated claims.

But the em dash survives. It survives because it occupies a unique position: it is simultaneously valid punctuation and a structural marker. In markdown, the dash is architecture. In prose, the dash is punctuation. The form is identical; the register is different. When a model receives the instruction “write prose, not markdown,” the suppression passes over the em dash because the em dash is already in the punctuation register. It does not need to be stripped because it does not, on its surface, look like formatting.

The paper proceeds to provide some clarification on the two major discourses online: (a) AI over uses em dashes and (b) AI “thinks” in markdown. Further, provides the results as:

The finding is therefore not that LLMs uniformly “overuse” em dashes relative to all human writing, but rather that (a) the em dash is the specific artifact that resists formatting suppression, (b) em dash behavior varies systematically across providers in a pattern consistent with different RLHF calibrations, and (c) some models produce em dashes at rates well above the human mean even under explicit suppression.

This was an interesting paper to read because I have been thinking about this a lot. In fact, I stumbled upon this paper when I was researching this topic online. The paper investigates if em dashes are excessively used by LLMs, and, if so, why. However, my inquiry in this article is about how I could tell if a piece of writing in the wild was written by human or AI, solely by using em dash usage frequency?

My question is straightforward. Given a piece of writing and its usage of em dashes, how likely is it that AI wrote it? If this formulation feels familiar for the Bayesians among the readers, you guessed correctly that I’m talking about the base rate fallacy. We could pose this question in similar lines and gather some insight.

Before we jump ahead, let me recap what the base rate fallacy problem is and address it using Bayes’ Theorem. This occurs when “people tend to ignore the base rate (e.g., general prevalence) in favor of the information pertaining only to a specific case.”

For example, say we are in the peak of the COVID-19 pandemic, and vaccines have been made available to a specific region. Say, roughly 90% of the population in this region are vaccinated. Further say, about 15% of the population is hospitalized for COVID-19 and about 2/3 of the hospitalized population are vaccinated. It’s a fallacy to conclude that the vaccine is ineffective because majority of the hospitalized population were vaccinated without considering the base rates.

First, let’s look at the known probabilities:

[embed]

Using these values and Bayes’ theorem, we can calculate the probability of hospitalization given the patient is vaccinated and unvaccinated.

[embed]

If we looked at just the numbers in the hospital without considering the base rate, it will give a skewed image of what’s happening. So, we should not make any conclusions about vaccinations and hospitalizations by just looking at hospitalizations without considering the base rates in the population.

That’s a recap of how we use Bayes’ theorem to handle the base rate fallacy problem. Now, we will analyze the AI writing problem in a similar fashion.

Our intention, as alluded to earlier, is to evaluate the probability that a given piece of writing is done by AI given the em dash usage in that piece of writing. The paper quoted above has done some experiments and labeling to find out em dash frequency in human writing as well as AI writing. We will directly use these frequencies to estimate the probability of em dash usage given human or AI writing. The paper used two forms of prompting in the experiments: the unconstrained paradigm where the prompts contained no stylistic constraints and the constrained paradigm where the prompt contained instructions preventing the use of any markdown formatting. To keep it simple, I averaged the frequency of em dash usage across multiple AI providers under the unconstrained (4.75/1K words) paradigm and constrained (2.07/1K words) paradigm as used in study 1 of the paper for AI writing, and used 3.23/1K words as human baseline.

[embed]

The log odds of AI writing given em dash usage is a function of the probability of human writing. Therefore, we can just vary P(Human Writing) for different em dash usage frequencies for AI to see how the log odds changes.

If the log odds of AI writing given em dash usage is higher that means it is likely that a given piece of writing is done by AI. The first term in the RHS is just a ratio of the em dash usage frequencies of AI to human. We consider three cases where the AI uses more, same or less em dashes per 1K words as compared to humans. The first chart below depicts the full range of possible base rates for human writing. In the second chart, we constrain the base rates for human writing between 40% and 60%. The second chart matches the reality when you encounter an “everyday” post on the internet: it is hard to say if it was written by human or AI.

The curves from top to bottom in the charts are when AI uses more, same or less em dashes per 1K words as compared to humans. If we look at 0.5 on the x-axis, it is the scenario where 50% of all written artifacts come from humans alone. Then the log odds for if the writing is by AI is roughly between -0.45 to 0.39 for different em dash usage frequencies of AI. When log odds are in this range, the probability of AI writing given em dash usage frequency is roughly between 0.4 to 0.6. This is what we typically would call a toss-up. It could go either way.

We saw three discrete scenarios of em dash usage frequency of AI as compared to humans. We may think that if we know which of these scenarios we are dealing with, then we can be more precise about the probability of AI writing given em dash usage. This, however, is not easy. There are several AI providers in the consumer market who use different modes LLM training which directly influences the outputs produced. Additionally, there is no way of knowing exact instructions or prompts used to produce the piece of writing in question (were specific markdown suppression instructions used?). Thus, we can only be certain that the em dash usage frequency of AI as compared to humans is a variable that can take a range of values depicted in the charts.

In summary, if we think we can easily distinguish a piece of writing as AI just by looking at em dash usage frequency, we better think again. It is much harder than we anticipate because em dashes are used in human writing as well as AI writing as a prose device. Given the multiplicity of avenues where em dashes can be intentionally suppressed in AI writing, analyzing em dash frequencies between humans and AI leaves us having to discern between two close choices. In addition, AI writing has infiltrated online writing world in a perverse manner, that the probability of a piece of writing is by a human is plummeting. So, in order to confidently say a written artifact is not by a human, we have to look for signals beyond em dash usage frequency.