Can AI Detect AI ? - OpenAI Thinks Not...

Educators want a tool to detect when students use AI programs to write instead of doing it themselves. There are even programs for teachers to use that allegedly detect when a student has "cheated" by using AI.

Some of the common signs of AI-generated text include:

• Oddly generic or repetitive writing. AI models are trained on large datasets of text, so they may simply repeat what they have seen before. This can lead to oddly generic or repetitive writing.
• Factual errors. AI models are not always accurate, so they may make factual errors in their writing.
• Lack of creativity or originality. AI models are not as creative or original as humans, so their writing may lack these qualities.
• Unnatural word order or sentence structure. AI models may not always understand the nuances of human language, so their writing may have unnatural word order or sentence structure.

In an effort to help educators identify when students use AI, OpenAI trained a tool (AI Text Classifier) to distinguish between text written by a human and text written by AIs from a variety of providers. 

However, OpenAI determined that its classifier is not reliable. In its own tests, the classifier correctly identified 26% of AI-written text (true positives) as “likely AI-written,” while incorrectly labeling human-written text as AI-written 9% of the time (false positives).

Other companies and websites also claim to have AI-detection tools. Some of these tools include GPTZero and Copyleaks AI Content Detector. From my review, Copyleaks appears to be the best performer, however it is not perfect and also makes mistakes.

It probably should not be surprising that AI has limitations... including not being able to accurately identify AI content.