As large language models have taken off, they've had huge implications for any industry that relies on text, most notably education. ChatGPT is one such model that has been accused time and time again of facilitating students cheating on essays and other assignments, leading to a number of instances of students unfairly being accused of cheating. Can ChatGPT be detected though? It's complicated, but the answer is largely no. There's more to it than that though, so keep reading.

ChatGPT can't really be detected

Nor can Microsoft Copilot or Google Gemini

As a whole, AI detectors can't and don't work, and the reason why is fairly nuanced. Essentially, they try to recognize patterns and the statistical likelihood of words appearing one after another in the order that they do. For anything academic, where a writing style is typically enforced, then that falls apart immediately, as it tends to be written similarly. Content written by an LLM like GPT-4, GPT-3.5, or Gemini can also be restructured and changed, too.

As for why tools like ZeroGPT, Turnitin, and other services don't work, I think the best example of it is OpenAI's own GPT detector that the company released in the summer of last year. Launched only to be quickly shelved, it was a tool that the makers of ChatGPT claimed could be used to detect AI, but unsurprisingly, it failed. It couldn't detect AI-generated content consistently, and it often flagged content that wasn't AI generated as being AI generated. Not exactly a powerful tool then, is it?

If OpenAI can't get it right, with their unparalleled access to the magic behind ChatGPT, then it's impossible to claim that the others can to any degree of consistency. ZeroGPT, for example, detects the constitution of the United States as being generated by AI. I may be wrong, but I don't think ChatGPT was available when the United States government was established in 1787. There are ways to try and detect AI content though, and none of it really involves software.

How to detect AI content

It doesn't involve software

There are a few ways to analyze text to figure out if it's AI generated, but to be honest, you'll never be sure. They'll serve as indicators that can give you more information to raise if you need to speak to someone about it, but they won't give you a smoking gun.

• Word usage: One of the most tell-tale signs of AI writing is the usage of extravagant or otherwise exuberant words. These words stand out, especially in a body of text that doesn't require it or is different to the writer's usual style.
• Repetition: AI models will often repeat the same descriptors or phrases throughout, especially if the content is a combination of multiple different prompts. Keep an eye out for these, as they may be phrases that are in its training data that don't really fit the content.
• Consistency and coherence: AI-generated text may sometimes lack consistency or have coherence issues over long passages, so keep an eye out for any logical errors. Language models don't employ logic in their response generation (typically unless it's mathematical), so keep an eye out for any logical contradictions.
• References to being ChatGPT: You may be surprised to learn that students have managed to out themselves in the past by accidentally including references the AI makes to itself in the text. Students have copy and pasted whole responses, including the closing where ChatGPT says that it's just ChatGPT.

These are the best ways to consistently spot AI, but you'll need to do it yourself. There aren't any tools that will do it for you consistently, especially with the wide variety of models that are available now to use.

If you're looking to try and identify AI-written content, then you'll need to look closely; and even then, you still mighn't find it.