You'd be hard-pressed not to come across AI-generated content sometimes. Thanks to the one-click text generators inside popular platforms like LinkedIn and Gmail, there's a chance an AI bot is already writing the emails you're responding to, the social media posts you're scrolling by, and the websites you're browsing.
Sniffing out AI text, however, isn't as nearly as effortless. You'll rarely see a disclaimer alongside artificially produced content, and because AI bots are trained on heaps of human writing, they've learned to write like us. That means spotting the details that distinguish AI-generated text isn't always obvious—and sometimes it's near impossible for an untrained eye.
At the same time, the ability to identify AI-written text matters more than ever. People are already using AI chatbots like ChatGPT to create misinformation—sometimes unknowingly, simply by trusting what it says. If you're not sure if what you're reading was written by a human, there's a variety of AI content detectors you can turn to. They're far from perfect, but they can pick up some of the clues that AI leaves behind to conclude how likely it is that a text is the work of a bot.
What is the best AI content detector?
There are scores of AI writing generators available online, and each uses a unique method to reach as close to human-level fluency as possible. As AI tech develops, it will continue to get better at hiding its patterns and errors. For that reason, no single tool is a silver bullet for spotting synthetic text with flawless accuracy, but there are some that come close.
AI content detectors are engineered to identify the most common tells—like a robotic rhythm or overuse of niche words—from a handful of popular AI text generators like ChatGPT, Google Bard, and Jasper. Some of them are also uniquely designed to cater to a specific use case like college essays.
To use these apps, all you have to do is copy in the text (some, like OpenAI's, will ask you to create an account first). There are often a few additional filters you can tweak to help the detector narrow down the search.
I decided to put three of the most popular AI content detectors through their paces: GPTZero, OpenAI AI Text Classifier, and Copyleaks AI Content Detector. To do so, I wrote three pieces of text on various subjects and asked ChatGPT to produce their AI-generated counterparts (using basic prompts like "explain how wireless charging works in 200 words").
Here's how each AI content detector performed.
The science behind AI detectors isn't perfect, or even consistent. The accuracy varies depending on lots of factors, including the types of content and writing styles.
GPTZero
GPTZero is aimed at professors looking to find out if their students are seeking help from ChatGPT on essays. It examines an excerpt on two parameters:
Perplexity measures how complex the text is. If GPTZero is "perplexed" by it, it means it's more likely to be human-written, but if it's familiar—indicating it's based on the same data GPTZero is also trained on—it's AI-generated.
Burstiness evaluates the variations of sentences. AI bots tend to stitch together sentences at a predictable uniform length, while humans write with greater variations.
GPTZero can review up to 5,000 characters in one go, and it can automatically extract text from an uploaded document. It can detect content from a range of AI services like ChatGPT and Bard, and it'll even tell you when it's a mix of AI and human composition.
Its performance in my tests was hit-and-miss. It correctly flagged the excerpt I wrote about wireless charging as human-written, but it failed against the AI-generated version and concluded it was also "likely to be written entirely by a human" despite its dull tone. What's more, it found the AI-written text to be more random and gave it a higher perplexity score. It produced similar results for the third piece of text as well.
GPTZero did far better on the second test, however. It not only confidently identified the AI-written text, but its scores for both versions also clearly reflected reality. ChatGPT's text was inundated with superlatives and a repetitive rhythm, and it expectedly scored poorly on both the perplexity and burstiness measures.
OpenAI AI Text Classifier
AI Text Classifier is another tool from OpenAI, the creators of ChatGPT. OpenAI trained it on mountains of both human-written and machine-generated text until it could tell the difference. It requires a minimum of 1,000 characters to run, and depending on how confident it feels about the results, it rates excerpts as "very unlikely, "unlikely," "unclear if it is," "possibly," or "likely" AI-generated.
OpenAI's detector did worse than GPTZero. While it flagged all my human writing as "very unlikely AI-generated," it failed to spot ChatGPT's work in all three tests. In one case, it said it was unclear if the ChatGPT-authored text was AI-generated, but in another, it said it was "very unlikely AI-generated." Not a great result.
Unlike GPTZero, AI Text Classifier has a linear approach to spotting AI-written text. Its judgment depends entirely on what it learned from reading heaps of human-written content during its training process and how close or far the submission is from that (as opposed to GPTZero, which scores text based on its proximity to a few obvious telltale signs of a machine).
Editor's note: Shortly after we published this article, OpenAI discontinued its AI Text Classifier because of accuracy issues.
Copyleaks AI Content Detector
The Copyleaks AI Content Detector is from a company that's been selling a plagiarism checker for nearly a decade. Copyleaks claims its AI detector takes advantage of its years of experience and feedback from refining its plagiarism checker, and therefore, has an unprecedented 99.1% accuracy. The Copyleaks detector is also the only tool on this list that's frequently updated for new AI engines and supports about a dozen languages.
Copyleaks delivered on its promises. In all of my tests, it accurately flagged the human and AI text. And each time, it produced a conclusive answer, instead of hedging its bets like GPTZero and AI Text Classifier. It even told me the exact probability of a particular sentence or paragraph being written by a human or a machine. It detected each excerpt's author with a probability of 99%.
Which AI content detector should you use?
I know I was working with a small sample size, but Copyleaks was, without a doubt, the most accurate AI text content detector I tested. It was barely a contest, honestly. So the next time you have a hunch an email you just got may have been written by a robot, you know where to confirm it.
Related reading:
