It is difficult to detect when text has been generated by tools like ChatGPT. Popular artificial intelligence detection tools like GPTZero can provide users with some guidance by telling them when something was written by a bot rather than a human, but even dedicated software isn’t foolproof and can throw up false positives.
As a journalist who started covering AI detection over a year ago, I wanted to compile some of WIRED’s best articles on the topic to give readers like you a better understanding of this complicated topic.
Do you have any other questions about detecting results from ChatGPT and other chatbot tools? Sign up for my AI Unlocked newsletter and contact me directly if you have any questions about AI that you would like answered or WIRED to learn more about.
February 2023 by Reece Rogers
In this article, written about two months after ChatGPT launched, I began to explore the complexities of AI text recognition and what the AI revolution could mean for authors publishing online. Edward Tian, the founder of GPTZero, spoke to me about how his AI detector focuses on factors such as text variance and randomness.
As you read, focus on the section on text watermarking: “A watermark can potentially mark certain word patterns so that they are off-limits to the AI text generator.” Although it was a promising idea, the researchers I spoke with were , already skeptical about their potential effectiveness.
September 2023 by Christopher Beam
A fantastic piece from last year’s October issue of WIRED, this article gives you a glimpse into Edward Tian’s thinking as he worked to expand GPTZero’s reach and detection capabilities. Focusing on how AI has impacted school work is crucial.
AI text recognition is at the forefront of many classroom teachers’ minds as they grade papers and may forego essay assignments altogether because students are secretly using chatbots to complete homework. While some students may use generative AI as a brainstorming tool, others use it to create entire assignments.
September 2023 by Kate Knibbs
Are companies responsible for labeling products that may be generated by AI? Kate Knibbs investigated how potentially copyright-infringing AI-generated books were put up for sale on Amazon, despite some startups believing the products could be detected and removed using special software. One of the core debates about AI detection revolves around whether the potential for false positives – text written by humans that is inadvertently flagged as the work of AI – outweighs the benefits of labeling algorithmically generated content.
August 2023 by Amanda Hoover
Beyond just homework, AI-generated texts are increasingly appearing in academic journals, where they are often banned without proper disclosure. “AI-authored papers could also divert attention from good work by diluting the pool of scientific literature,” writes Amanda Hoover. One possible strategy to solve this problem is for developers to create dedicated discovery tools that look for AI content in peer-reviewed articles.
October 2023 by Kate Knibbs
When I first spoke to researchers about watermarking for AI text recognition last February, they were hopeful but cautious about the potential to imprint AI text with specific language patterns that are undetectable to human readers but obvious to recognition software . Looking back, their concerns seem justified.
Just six months later, Kate Knibbs spoke to multiple sources who broke AI watermarks and demonstrated their underlying weakness as a detection strategy. Although it is not guaranteed to fail, it remains difficult to watermark AI text.
April 2024 by Amanda Hoover
One tool that teachers are trying to use to detect AI-generated classwork is Turnitin, a plagiarism detection software that has AI detection capabilities. (Turnitin is owned by Advance, the parent company of Condé Nast, which publishes WIRED.) Amanda Hoover writes: “Chechitelli says the majority of the service’s customers have opted to purchase AI recognition. But the risk of false positives and bias against English learners has led some universities to abandon the tools for now.”
AI detectors are more likely to incorrectly label written content from someone whose native language is not English as AI than content from someone who is a native speaker. As developers continue to improve AI detection algorithms, the problem of erroneous results remains a key obstacle to overcome.