Skip to Main Content

Detecting Generative AI

Introduction

In response to this rise in AI content generation, there has been a corresponding rise in the development of AI text detection tools. However, while AI text generators are improving rapidly, the development of AI text detection tools has not kept pace.

AI detection tools are not reliable enough to use by themselves, with OpenAI and others acknowledging these tools should not be relied upon to definitively determine if the content is AI-generated. It should also be noted that texts can be edited to avoid detection. Before submitting work to third-party apps and websites, it is also crucial to consider the limitations of AI text detection tools and any privacy implications.

AI detectors will likely continue to improve as AI text generators become more advanced. However, it's unlikely that there will be a tool that can detect AI writing with 100% accuracy.  

Examples

OpenAI's AI text classifier has limitations in accurately detecting AI-generated text. This tool cannot be used on text less than 1,000 characters, has a 26% accuracy rate, performs poorly with languages other than English, and at times mistakes human-written text for AI-generated content. 

Another popular detector, GPTZero, measures the perplexity (randomness) and burstiness (complexity) of sentences and structure and apparently has a much higher success rate, with some sources saying as high as 98% accuracy. However, many tech websites and blogs claim to have disproven this with their own tests, and much like OpenAI text classifier GPTZero has its own limitations; struggling with short sentences and text that has been edited. It provides clearer feedback as well as providing reasoning for why a determination was made. 

Turnitin has developed an AI detection tool that they advise is designed to help educators identify texts that may have been generated by an AI tool. Whilst they tout a 98% confidence rate, this is caveated with a statement advising: "(o)ur AI writing detection model may not always be accurate (it may misidentify both human and AI-generated text) so it should not be used as the sole basis for adverse actions against a student. It takes further scrutiny and human judgment in conjunction with an organization's application of its specific academic policies to determine whether any academic misconduct has occurred." 

Tags: AI, artificial intelligence, ChatGPT, Generative AI