Detecting when text has been generated using tools like ChatGPT is a difficult task. Popular AI detection tools, such as GPTZero, can provide some guidance to users by telling them when something was written by a bot and not a human, but even specialized software is not foolproof and can return false positives.
As a journalist who began covering AI detection more than a year ago, I wanted to curate some of WIRED’s best articles on the topic to help readers like you better understand this complicated topic.
Do you have even more questions about how to detect results from ChatGPT and other chatbot tools? Sign up for my AI Unlocked newsletter and contact me directly with anything AI-related you’d like answered or want WIRED to explore further.
February 2023 by Reece Rogers
In this article, which was written about two months after the release of ChatGPT, I began to grapple with the complexities of AI text detection, as well as what the AI revolution could mean for writers publishing online. Edward Tian, the founder behind GPTZeroHe talked to me about how his AI detector focuses on factors like text variation and randomness.
As you read, focus on the section on text watermarks: “A watermark could designate certain word patterns so that they are out of reach of the AI text generator.” While it was a promising idea, the researchers I spoke to were already skeptical about its potential effectiveness.
September 2023 by Christopher Beam
This article, a fantastic piece from last year’s October issue of WIRED, gives you an inside look at Edward Tian’s mindset as he worked to expand the reach and detection capabilities of GPTZero. It is crucial to focus on how AI has impacted schoolwork.
AI text detection is a priority for many classroom educators when grading papers, potentially foregoing essay assignments altogether as students secretly use chatbots to complete assignments. While some students may use generative AI as a brainstorming tool, others use it to fabricate entire assignments.
September 2023 by Kate Knibbs
Do companies have a responsibility to flag products that could be generated by AI? Kate Knibbs investigated how potentially AI-generated copyright-infringing books were being made available for sale on Amazon, even though some startups believed the products could be detected with special software and removed. One of the central debates around AI detection revolves around whether the potential for false positives (human-written text that is accidentally flagged as AI work) outweighs the benefits of labeling algorithmically generated content.
August 2023 by Amanda Hoover
Beyond schoolwork, AI-generated text is increasingly appearing in academic journals, where it is often banned without proper disclosure. “Articles written with AI could also divert attention from good work by diluting the body of scientific literature,” writes Amanda Hoover. One possible strategy to address this problem is for developers to create specialized detection tools that search for AI content in peer-reviewed articles.
October 2023 by Kate Knibbs
When I first spoke to researchers last February about watermarking for AI text detection, they were hopeful but cautious about the possibility of printing AI text with specific language patterns that are undetectable to human readers but obvious to detection software. Looking back, his fear seems well founded.
Just half a year later, Kate Knibbs spoke to multiple sources who were breaking AI watermarks and demonstrating its underlying weakness as a detection strategy. While not guaranteed to fail, applying AI watermarks to text remains difficult to achieve.
April 2024 by Amanda Hoover
One tool that teachers are trying to use to detect AI-generated classroom work is Turnitin, a plagiarism detection software that added AI detection capabilities. (Turnitin is owned by Advance, the parent company of Condé Nast, which publishes WIRED.) Amanda Hoover writes: “Chechitelli says most customers have opted to purchase AI detection. But the risks of false positives and bias against English learners have led some universities to ditch the tools for now.”
AI detectors are more likely to falsely label the written content of someone whose first language is not English as AI than that of someone who is a native speaker. As developers continue to work to improve AI detection algorithms, the issue of erroneous results remains a critical hurdle to overcome.
