Why AI detectors can be wrong?

When using AI-based content detection, it is important to be aware of false positives. False positives occur when the AI assigns a high score to content that was produced by a human. False positives happen in approximately 3% of cases. This article aims to explain the most common causes of false positives and provides recommendations on how to mitigate them.

Short Content

Content that is scanned should ideally be at least 100 words long (that is about 400 characters), as longer content tends to yield more accurate results. To illustrate this, I conducted an experiment where I wrote a 50-word introduction for a blog post. Surprisingly, the AI scored it at 79%!

What happened here?

The short length of the content played a significant role. With insufficient text to analyze, our AI detection software had limited information to make an accurate judgment, resulting in an inflated score.

Formulaic Content

Formulaic content, such as introductions, conclusions, standardized essays, statistical articles, or scientific papers, often triggers higher AI scores. While not all formulaic content will be flagged as AI-generated, it is more likely to receive a higher score.

Scanning Public Domain Content

Another common source of false positives is the scanning of public domain content, including books, journals, or internet content. Even if the content was written centuries ago, it can still trigger false positives.

As an example, I scanned the first page of Jane Austen’s “Pride and Prejudice,” written in 1813. Surprisingly, it received a 67% probability of being AI-generated.

How can a work written over 200 years ago be mistaken for AI?

The answer lies in the training data used for AI models. Many large language models, like ChatGPT and Bard, are trained on public domain books, papers, and journals. If the content is in the public domain, there is a high chance that AI models have been trained on similar data.

This means that AI models have learned to replicate public domain works, and the AI detection algorithms are trained to identify such works. Consequently, scanning public domain content or content with a similar style may result in false positives.

It is important to note that not all public domain works trigger false positives.

For instance, scanning excerpts from “The Adventures of Sherlock Holmes” and “The Tragedy of Romeo and Juliet,” and both were correctly identified as 100% original human-authored works.

Reducing False Positive Rate

To minimize false positives, we recommend scanning longer content whenever possible. Longer content tends to be more unique and mitigates the issues associated with short or formulaic content.

Keep in mind that scanning formulaic content, such as articles with numerous statistics or scientific papers, may yield higher AI scores. Similarly, scanning public domain works or other content that AI models have been trained on may result in higher-than-usual scores.

The best defense against false positives is awareness of the causes, scanning multiple works instead of a single piece, and focusing on longer content. For optimal results, we suggest scanning 3-5 works of at least 100 words each from the same author.

Scroll to Top