How to detect if a text was written by ChatGPT?

Texts generated by artificial intelligence share certain common identifiable features. However, while OpenAI claims to develop tools to identify text written by ChatGPT, there is currently no 100% secure technique.

Students whose copy was written by artificial intelligence. Fraudsters who use it to write phishing emails. After wowing the Internet with its ability to write human-like text, ChatGPT AI was quickly seized for fraud or malicious use.

Schools in New York have even banned the use of AI, created by California startup OpenAI. But how do you know if a text was generated by ChatGPT? Is that possible?

Visible clues while reading

Even without sophisticated software, certain formulations can give you a clue. For example, the repeated presence of common and impersonal words rather than rare words or expressions. For example, “the” (the), “it” (that) or “is” (is) for English text.

The reason? Contrary to appearances, ChatGPT is not really a chatbot, but an algorithm that calculates the most likely text continuation. If you ask him a question, he’ll understand that the most likely outcome is an answer – but the answer will consist of the words that are most likely to appear in his database. Neither boilerplate is like “this” or “that”, as the MIT Technology Overview explains.

Another hint: because ChatGPT works with probability, if multiple people ask the exact same question, it will generate the same answer for everyone, most likely – except for a few details. A good hint for teachers: if they proofread several copies that look very similar, with the same grammatical constructs, the same reasoning, the same examples… They can be generated by an algorithm. This is what caught the attention of the teacher of Lyon, half of his master’s students have used ChatGPT to write their copy.

But unlike many people, ChatGPT makes no mistake in French. If the text you are reading contains errors in agreement or grammar, then it is most likely written by a human.

detection software

As if to accompany the ChatGPT explosion, more and more sites are claiming to be able to detect the origin of text with impressive precision. But without an explanation of their methods, those promises often sound too good to be true.

One of the most transparent and popular tools is GPTZero. The site, developed by computer science student Edward Tian over his Christmas vacation, takes an approach that’s been used in AI before: if an algorithm creates it, a similar algorithm will know how to recognize it.

To let you know whether it’s from AI or humans, GPTZero will run your text through the previous ChatGPT model, called GPT-2. “It calculates ‘confusion’: does GPT-2 find text it’s familiar with? Or is it surprised by sentence lengths or expressions that don’t match the probabilities it’s learning?” explained Edward Tian to Tech&Co.

So all you have to do is enter your text on the site, hit enter, and if “refusal” is high, it’s most likely human-generated. Add to that “burstiness”, which assesses how much this confusion varies over the text: sentence lengths generated by an AI won’t vary much across text, whereas human sentences will be more random.

“I have nothing against AI, but its adoption must come with safeguards,” explained Edward Tian to Tech&Co. “People should be able to know the truth about what they read.”

There’s no quick fix

But GPTZero can’t assure you 100% that the text is from human or machine. For a simple reason: currently it’s not possible.

All current detection tools are imperfect, and make errors more or less regularly. For example, we asked ChatGPT to write like a 4 year old (in English): the software alternated between classic sentences and shorter exclamatory sentences. A text varied enough to fool GPTZero.

It’s also easy to manually add errors to the text written by the AI, or reformulate it slightly so that it isn’t detected. This process can even be automated: in twittera computer scientist explains that he has created a program that adds invisible spaces in the middle of certain words – thus turning them into words unknown to GPTZero, who are therefore “confused” and will assume the text to be human creation.

Finally, GPTZero is less efficient on languages ​​other than English, because GPT-2 is primarily trained on English text.

Edward Tian was aware of this shortcoming. He recalled that his current model was designed to detect academic cheating. Never mind he let the written text pass like a 4 year old.

“I look forward to starting full time soon in improving the software, to add new selection criteria and to make it available to everyone,” the student explained to Tech&Co.

“Flag” text generated by upstream AI

This detection method has another drawback: “To calculate the ‘confusion’ of an algorithm in front of text, you need to have a lot of information about the algorithm in question”, notes Edward Tian for Tech&Co. “The model itself, the parameters, the weight…”

GPTZero works thanks to GPT-2, a model released by OpenAI in 2019, but has been largely overtaken by ChatGPT. “If other companies build more sophisticated but non-transparent AI, detection can become much more complicated,” admits Edward Tian at Tech&Co.

This is why one strategy is to make these AI-generated texts clearly identifiable from their creations, by adding special markings to them that will leave no room for doubt.

This is a strategy followed by OpenAI: the company intends to modify its next algorithm to ensure that “every time GPT generates a long text, there is an invisible secret signal in its choice of words, which you can use to prove later that, yes, it was right. of the GPT,” according to researcher Scott Aaronson who recently joined the startup. In universities as on the internet, the chase is just beginning.