Montreal AI detection tool can counter ‘ChatGPT anomalies’

The process is simple: open the AI ​​detection tool’s web page in a browser (A new window)we copy and paste a text in English (French version is in preparation) of at least 200 characters for the test, we press analysisThe program offers a percentage of the chance that the essay will be written by a human or artificial intelligence (AI).

If you are testing content whose nature you already know, you can also indicate at the bottom of the screen whether the score obtained by the company is correct.

” We ran our first analysis on Wednesday, based on 2,000 English-language texts. The results showed a reliability of 93%. »

For the second version we are working on, with an improved algorithm, it is estimated that this percentage will be higherThe CEO of the company told Radio Canada.

Zoom in (A new window)

Draft & Goal’s AI detection tool analyzes submitted texts to determine whether they were written by a human or a machine.

Image: draft and target

To be able to detect if a text has been generated by a chatbot, Draft & Goal uses machine learning.

Generative AI models are often probabilistic, that is, models that ensure that what is written is not completely randomNabil El-Tayeb explains.

This brings the draft and goal to Detection of fingerprints left by artificial intelligence, in particular by looking at several elements of writing style.

If the tool is only available in English for now, a French version of the service will be online by the end of January, according to Nabil Tayeb.

There is a lot of English text on the web, so it is easier and faster for beginners to practice with English than with Frenchsays the CEO of Draft & Goal, who plans to expand the service to other languages ​​in the future.

One combination of circumstances

Draft & Goal has been working on its own for two years to build an automated content creation tool. She wants to help other companies be more productive writing blog articles, posting static web pages, or writing product descriptions, while keeping some human intervention.

By generating the texts, we needed to know if they would be perceived as text written by an AISays Nabil Tayeb, CEO of Draft & Goal, specifying that the goal is not to mislead readers.

” So we created a tool to help us out, and quickly realized we were pretty good at spotting AI-generated content. »

The team saw this as an opportunity to share it with others, because there are many risks associated with the democratization of AI software that simulates human typing in an ultra-realistic way, as is the case with ChatGPT d’ Open AI.

We see AI as a tool, a typewriter that should be improved, but it should not become the center of creativity as suchthought the CEO of Young Shooting.

Several areas of work

Teachers fear For example, these chatbots infiltrate their students’ essays at all levels.

Professors have been using software to spot plagiarism for years, but it’s designed to analyze texts already published on the web.

Since the AI ​​generates original text, it does not copy text from the webHe says.

A tool like Draft & Goal can also be useful in press rooms to ensure the credibility of the press release.

” [Une personne malintentionnée] It can send out completely fake press releases without being suspicious, and it can be done on a massive scale. »

in the testing phase

At the moment, the draft tool and target are still in development. Each order improves its own artificial intelligence.

The company wants to keep the software free and available to everyone as long as possible. Soon it still intends to offer paid plans with additional features, which could take the form of, among other things, an extension on the Chrome browser.

A new version of the detector will be put online In the coming days It will allow finer sections of suspected AI-generated text to be targeted.