Last week, OpenAI released valuable insights for educators showing that some teachers are leveraging ChatGPT as an educational tool. The company also provided suggested prompts to assist educators in getting started.
However, in the related FAQ section, the company also revealed what everybody knew: AI detectors don’t do their only job adequately. As a result of their improper functioning and bias, many students ended up being punished unjustly.
AI detectors don’t work
When asked if AI detectors actually work, OpenAI shortly stated that they don’t. Many companies, OpenAI included, have introduced tools claiming to detect AI-generated content. However, none have consistently proven capable of reliably distinguishing between AI- and human-generated content so far.
Namely, AI detectors frequently generate false positives due to their reliance on unverified detection metrics. Ultimately, there is nothing inherently distinctive about AI-generated text that reliably sets it apart from human-written content. Plus, determined writers can easily circumvent these detectors through simple rephrasing.
For this reason, OpenAI discontinued its AI Classifier—an experimental tool designed for detecting AI-written text, which exhibited an alarmingly low 26% accuracy rate.
Can ChatGPT recognize AI-generated text?
OpenAI’s new FAQ also dispels another common misconception that ChatGPT can discern whether text is AI-generated or not. OpenAI clarifies that ChatGPT lacks any ‘knowledge’ of what content might be AI-generated. It does occasionally generate responses to queries like ‘did you write this [essay]?’ or ‘could this have been written by AI?’ Yet they are all arbitrary and devoid of factual basis.
In line with this, OpenAI also addresses the inclination of its AI models to produce fabricated information. The company stated that ChatGPT may sound convincing occasionally but could also provide incorrect or misleading information, that is, hallucinate. The company added that the popular chatbot can even make up quotations or citations, which should not be trusted.
Is AI-written text unrecognizable?
Although automated AI detectors exhibit a lack of efficacy, this does not imply that humans cannot identify AI-generated writing. Educators familiar with a student’s typical writing style can discern deviations in style or capability.
Additionally, sloppy attempts to present AI-generated work as human-written often leave unmistakable traces, such as the phrase “as an AI language model.” These further indicate that someone carelessly copied and pasted ChatGPT output. Notably, a recent article in the scientific journal Nature highlighted instances where readers detected the phrase “Regenerate response” in a scientific paper, which corresponds to the label of a button in ChatGPT.
In the current technological landscape, it is safest to steer clear of automated AI detection tools altogether. AI analysts have observed that AI writing remains virtually undetectable and is likely to stay so. AI detectors exhibit high false positive rates and should be avoided.