IN MARCH 2024 , author Joanna Maciejewska tweeted: “I want AI to do my laundry and dishes so that I can do art and writing, not for AI to do my art and writing so that I can do my laundry and dishes.” Her comments reflect the growing trend of using AI tools like LLMs (Language Learning Models) to generate text, images, and other types of media that mimic human creativity.

Many LLMs are trained on existing content without the consent of the creator, so they aren’t always compensated properly for derivative works. In academic circles, bad actors sometimes pass off AI-generated content as their own to earn recognition and qualifications. What’s even worse is that AI doesn’t get things right all the time. Recent analysis of OpenAI’s latest ChatGPT4o model gives it an accuracy rating of 88 percent. If you’re well-versed in a particular subject, this can be an excellent way to detect if text content is AI-generated.

In this guide, we’ll explore how to identify inconsistencies that plague AI-generated content. You’ll learn more about the tone and style of AI text, as well as telltale signs of generated images like flawed hands and backgrounds. We’ll also touch on detecting AI-generated audio and videos.

–NATE DRAKE

A

1 RECOGNIZE FACTUAL ERRORS IN AI TEXT

One of the easiest ways to detect AI-generated text is to read through it carefully. LLMs work by predicting the next word in a sentence. This means the style tends to be more uniform and less prone to rare words than text written by a human.

» The same is true for careful writers—especially those who don’t speak English as a first language. Bad actors may doctor AI-generated text to introduce more obscure terms or spelling/ grammatical errors to make it harder to recognize. This may explain why a 2021 study using the GPT-3model found that humans were unable to reliably detect AI text just by reading it.

» Still, LLMs can sometimes still give themselves away through inconsistencies in the text. These may be factual, or just abrupt shifts in the tone, style, or topic.

» For instance, in the example above, ChatGPT’s GPT-4o model was asked to write instructions to activate the ‘blood code’ cheat in Mortal Kombat (1993). Experienced gamers will know that both Sega and Nintendo coded the home conversions of the game to be more family friendly by rendering blood in gray.