A New Study Finds AI Text Detectors are Not as Accurate as Some Claim

A study by researchers from University of Pennsylvania finds that AI text detectors aren’t as reliable as people think. AI text detectors are used to analyze if a text is AI generated or not. But these AI text detectors are developed in a way that they only look out for some specific signs that can easily be written by humans too. Sometimes, AI detectors easily find the AI generated text but they do not do well when presented with some specific texts like news articles. Sometimes, these AI text detectors also mark completely human written text as AI generated content.

The researchers of the study proposed a new way to generalize an AI written text by giving these detectors a data set of 10 million documents from news articles to blogs and recipes, and benchmarking them. A public leaderboard will then rank AI detectors according to their performances based on those datasets. They said that they are trying to benchmark AI detectors so if someone comes up with a new idea, we can validate that their text is in fact human written.

Ever Since the release of GPT-2 in 2019 and GPT-3 in 2022, there have been so many problems regarding the text produced by AI text generators. Many teachers have also shown their concerns about their students using LLMs to write their assignments and academic papers. Many AI detectors claim 99% accuracy which is too good to be true. Some even claim that AI is hard to detect.

The researchers say that AI detectors are easy to deceive if we replace certain words and add British spellings on the words. Some detectors also work best on the AI text models they were trained on, so they sometimes cannot accurately detect text written by models like Anthropic’s Claude. Some AI detectors which were specifically designed to detect news may also struggle on recipes. All in all, as LLMs are getting better, AI detectors are struggling to detect AI written text accurately.

Image: DIW-Aigen