The Best AI Detection Tools to Catch Cheating and Plagiarism

• OpenAI, a research company aiming to advance artificial intelligence (AI), released ChatGPT, a new tool that made national waves at the end of 2022.
• Now, educators and school professionals are concerned about academic integrity and the future of learning.
• Several companies, including companies that work to advance AI, have worked to create AI detection tools.
• Although none of the tools are foolproof, they can be used to feel out whether a piece of text was written by a human or a robot.

Since the release of ChatGPT — a tool created by OpenAI in November 2022 that can spit out entire essays, computer code, and poetry — educators have been worried about how it might change the course of teaching and learning.

But just as quickly as ChatGPT and other AI-generating tools have made national headlines, new detector tools have been launched to sniff out if a piece of text is written by a human or by a bot.

These tools are not foolproof, and the companies that created them do not claim they are. Jan Leike, head of OpenAI's alignment team, even warned that their recently released detection tool "shouldn't be solely relied upon when making decisions."

According to TechCrunch, "As text-generating AI improves, so will the detectors — a never-ending back-and-forth similar to that between cybercriminals and security researchers… That’s all to say that there’s no silver bullet to solve the problems AI-generated text poses. Quite likely, there won’t ever be."

Still, these tools can at least detect lower-quality AI-generated content and give users a feel of whether or not a text was human-generated. Generally speaking, the longer the body of text in a passage, the better the tool works.

According to Search Engine Journal, OpenAI plans on introducing a cryptographic watermarking feature. By embedding statistical patterns into word choices and even punctuation, the watermark would make it easy for a detection tool to catch it.

Scott Aaronson, a computer scientist working on the OpenAI watermarking project, said, "We want it to be much harder to take a GPT output and pass it off as if it came from a human. This could be helpful for preventing academic plagiarism, obviously, but also, for example, mass generation of propaganda… Or impersonating someone's writing style in order to incriminate them."

Until watermarking features are implemented, however, AI detection tools may be needed to help prevent cheating and plagiarism. Here's a list of the current and future tools that students and professors can use to help detect if something was written using AI.

OpenAI's AI Text Classifier

OpenAI on Jan. 31 released their iteration of an AI detection tool, which they confirm is "not fully reliable." The "work-in-progress" tool "correctly identifies 26% of AI-written text (true positives) as 'likely AI-written,' while incorrectly labeling human-written text as AI-written 9% of the time (false positives)."

The classifier is free to use, and is more accurate in detecting AI in longer texts. According to Glenn Gabe, a digital marketing veteran of 27 years, the tool provides one of five responses to a body of text: very unlikely, unlikely, unclear, possibly, and likely of being AI-generated.

OpenAI said that when used in tandem with other tools, its text classifier could help prevent AI generation tools from being abused, according to TechCrunch.

GPTZero

Edward Tian — a computer science major and journalism minor at Princeton University — recently created GPTZero during his winter break. According to NPR, more than 30,000 people have already used the app since its launch on Jan. 3.

Tian tweeted, "in short, there's so much chatgpt hype going around. is this and that written by AI? we as humans deserve to know!" His original tweet about the app's launch received over 7 million views on Twitter.

According to NPR, Tian trained the bot to detect two different factors: "perplexity," which measures the text's complexity, and "burstiness," which compares sentence length and complexity variation.

Tian recently released a model update called GPTZeroX, intended primarily for educators. The updated tool highlights the parts of the text that are AI-generated, supports larger inputs — including files for an entire class — and operates faster than the pilot version, according to Tian's Twitter. According to Tian, these were the most requested features by educators.

Originality.AI Chrome Extension

Originality.AI's detection tool is designed to be used by writers and web publishers. It claims a 94% detection rate on GPT-3 generated content, and it is easily downloadable as an extension on Chrome for reviewing any piece of content a person is writing or reading on the web.

The detection module scores text, and even entire websites, on a scale of 0-100 on the chance that it was AI-generated. The trial for the detection tool is free, but after the trial period ends, it will cost $0.01, or one credit, per 100 words for users.

The tool can also scan for plagiarism. According to Gold Penguin, an agency that writes free tutorials about the latest AI software, "With the abundance of information available online, it's easy to accidentally plagiarize without even realizing it. Originality can be a valuable tool for students looking to verify the originality and authenticity of their work before submitting it for a class or assignment."

Content at Scale AI Detector

Content at Scale is best known for its promise of generating full length and SEO-optimized blog posts with just a few clicks, all the while surpassing even the toughest AI detectors. This is why its version of the AI Detector, which measures a text's human content score, is one of the most popular free detectors on the market.

Users can input up to 25,000 characters into the tool.

Writer's AI Content Detector

Just like Originality.AI's model, Writer's AI Content Detector is intended to help writers make decisions about what content they are posting online. But anyone can use the tool for free. After entering text into the box, the tool will spit out a detection score — the higher the percentage, the higher the likelihood the content was written by a bot.

The tool can now detect GPT-2, GPT-3, GPT-3.5, and ChatGPT, according to Gabe.

GLTR (Giant Language Model Test Room)

The MIT-IBM Watson AI Lab and Harvard Natural Language Processing Group collaborated to create GLTR, a tool aimed "to take the same models that are used to generate fake text as a tool for detection."

A language model, generally speaking, is a model trained to predict the "very likely" next word in a sentence or phrase. GLTR uses the language model to detect if a word in a sentence is too likely, which strays away from often-unpredictable human writing.

If a word is highlighted in green, it is in the top 10 of most-likely predicted words to follow the word in behind it, according to Gabe. If a word is highlighted in red or purple, the word is in the top 1,000 or more predicted words.

This tool takes more intuition than others, but generally speaking, if a piece of text is mostly highlighted in green and yellow, the odds are high that AI generated the content.

Hugging Face OpenAI Detector

The community at Hugging Face claims dedication to "good machine learning," including an open access AI detection tool free for all.

Content is ranked from a scale of "real" to "fake" in the Hugging Face OpenAI Detector, which can detect GPT-2 produced content. The tool uses the GPT-2 output detection model, a product created by OpenAI in 2019, and currently does not detect the latest AI models, such as GPT-3 and GPT-3.5.

OpenAI says that the detection model "needs to be paired with metadata-based approaches, human judgment, and public education to be more effective."

Gold Penguin classified this detector as an extra detection method as opposed to a standalone tool. The blog said, "Generally you can get a decent clue based on the response, but I've pasted some fully robotic AI writing and it has told me it was 99% real. I've also pasted an advanced academic essay written by a human and claimed 99% AI."