As language models like GPT continue to improve, it is becoming increasingly difficult to differentiate between AI-generated and human-written text. But, in some cases, like academics, it’s necessary to ensure that the text isn’t written by AI.
This is where AI text detectors come into play. Though none of the tools currently available detect with complete certainty (and neither do they claim to do so), a few of these tools do provide pretty accurate results. So, here, we list down the eight most accurate AI text detectors you can try.
GPTZero was built to help teachers figure out if the work submitted by their students was generated using AI. But since GPTZero detects ChatGPT-generated content accurately, you don’t necessarily have to be a teacher to benefit from it.
GPTZero analyzes the text and tells if the text was written by humans or AI, highlighting the specific sentences it believes are AI-generated. It also provides a perplexity and burstiness score, which are measures for randomness and variation in randomness. Generally, these scores are high for human-written text and low for AI.
To test the tool, first, we pasted an excerpt from Ernest Hemingway’s Soldier’s Home. GPTZero highlighted a few sentences as potentially AI-generated text. However, it was able to correctly identify ChatGPT-generated text as “written entirely by AI”.
Lastly, we entered an excerpt from AI-written text mixed with human content. Astonishingly, GPTZero identified all the sentences that were written by AI correctly. But it marked a few human-written ones as well.
Overall, GPTZero is good at detecting AI text, but also produces false-positive results. The tool is free and has an API too.
OpenAI has created an AI Text Classifier to counter its own GPT model. Though far from being completely accurate, this Classifier can still identify AI text. Unlike other tools, OpenAI’s Classifier doesn’t provide a score or highlight AI-generated sentences.
Instead, it tells the likelihood of the content being AI-generated using five categories from very unlikely to likely. OpenAI’s Classifier was able to recognize Hemingway’s story and marked it as “unlikely to be written by AI”.
Similarly, it correctly identified an excerpt from ChatGPT chat and labeled it as “likely to be AI-generated”. Lastly, it marked the mix of human and AI text as “unclear”, which is the most suitable option among the five labels it uses.
Content at Scale is itself an AI text generator that focuses on producing long-form content with a few clicks. It has an impressive AI detection tool too.
Content at Scale AI Detector provides scores for probability, pattern, and predictability, along with an overall score. These are expressed as percentages. The higher the percentage, the more likely that the text was written by a human. The tool also highlights sentences that it believes are AI-written.
In our tests, it gave a 96% score to human-written content, 22% to ChatGPT’s text, and 61% to the combination of both. Talking about the specific sentences, it nearly identified all the AI-written sentences. But, like GPTZero, it wrongly labeled a few of the human-written sentences as AI-generated.
This tool is free too and produced quite similar results as GPTZero.
Originality AI is a popular AI text detector that claims to accurately detect text produced by GPT 3, GPT 3.5, and ChatGPT. It gives a percentage of the likelihood that the text was generated by humans or AI.
Originality AI marked human-written text as 71% likely to be written by a human, which isn’t ideal. But Originality AI accurately detected ChatGPT’s answer, labeling it as 100% AI. When asked to analyze a mix of human and AI text, Originality marked it as 94% human.
So, while it was able to detect GPT content, mixing it with human text puzzled it. Unlike most other tools, Originality AI is a paid detector. However, it provides a few credits for trial if you download its Chrome extension.
Writefull aims to make academic writing easier. To that end, it has a number of AI tools, including a GPT detector. The tool is free to use and encrypts the data to ensure privacy.
Writefull GPT Detector can identify content written by GPT-3 as well as ChatGPT. After analyzing, it provides a score indicating the probability that the content was written by AI.
In our tests, Writefull correctly detected human-written text, giving it a score of 1% (i.e., a 1% chance that the text was AI-produced). Likewise, it gave a 100% score to ChatGPT’s text. However, when we combined AI-generated content with human text, Writefull marked it as “1% likely this comes from GPT-3 or ChatGPT”.
So while it did recognize AI and human texts individually, Writefull erred when we mixed the two. The tool offers integration with Word and Overleaf and has an API for GPT Detector as well.
Hive Moderation does not only have a tool to detect AI text, but also to spot AI-generated images.
The text detection tool accurately analyzed human-written text as “likely to contain 0% AI text”. Similarly, AI-generated text was identified as such. But like most others, it wasn’t able to detect the combination of both texts as accurately.
On its site, you can use the demo tool for free without even signing up. But you’ll have to contact them for full access. Similarly, the image detector can help you spot images generated by AI tools like DALL-E 2 or Midjourney.
Paraphrasingtool.ai has an AI content detector with two options: Detect by Jarvis and Detect by Veronica. Jarvis highlights the potentially AI-generated sentences, whereas Veronica assigns a probability percentage.
In our first test, Detect by Jarvis wrongly marked a few sentences as written by AI, while Detect by Veronica correctly showed that the content was 100% human-written. The ChatGPT-written text was, however, detected by both correctly.
Detect by Jarvis performed better when asked to analyze the combined text, highlighting most of the AI sentences. But Veronica showed it as 100% human content. A free tool, AI Content Detector by Paraphrasingtool.ai, provided decent results overall.
Copyleaks, one of the most popular plagiarism detection tools, has another tool to detect ChatGPT, GPT-3, and other similar language models. What makes it a standout is that it can detect content written in other languages, like Spanish and French, as well.
The tool was able to correctly identify Ernest’s short story as human-written with around 85% probability. Likewise, it detected AI text with 99.99% certainty. Where it did err was the mix of AI and human content, as it marked it as 96% likely to be human text.
AI Content Detector by Copyleaks has a Chrome extension. Best of all, it offers API and LMS integration.
Spot AI-Generated Text With a Single Click
Thanks to the above-mentioned tools, you can spot AI-written text quite easily. However, these are still far from being completely accurate. Just by changing a few sentences, it’s possible to trick these detectors. And a more serious concern with these tools is the false positives.
Even if you don’t want to rely completely on these detectors, AI text contains several telltale signs that you can spot if you pay a little attention.
Read the full article here