Janelle Shane of AIWeirdness speaks at the Better By Design CEO Summit 2019 on March 19, 2019 at Villa Maria Estate, Auckland, New Zealand. Phil Walter/Getty Images, NZTE
Years before we knew about AI, Janelle Shane was involved in this case. A scientist with a Ph.D. in Engineering, she works for a research company that provides organizations such as NASA with complex custom lighting control systems and applications to the International Space Station. She has also had her sideline job as an authoritative voice on AI for almost a year now. 10 years. In 2019, she published her popular book, You Look Like a Thing and I Love You, and gave her widely-reviewed TED talk about the hype and reality of AI. And since 2016 she has been writing a blog called “AI Weirdness”. And she came up with her post that AI detection tools shouldn’t be used for anything important—not at all.
In a blog post dated June 30, she highlighted an April study from Stanford University in which AI detection products that are supposed to detect text written by language models, like ChatGPT, can detect AI-generated content. , showed that they often misidentified as written by AI. human. They also have a strong prejudice against non-native English-speaking writers.
The short answer is that AI detection tools fail to detect content created by generative AI, and flawed technology flags and punishes content specifically written by non-English-speaking writers. That’s it. In other words, AI thinks other AIs are humans, and it thinks humans who write in languages it isn’t proficient with are AIs.
“What does this mean?” Shane wrote. “Assuming knowledge of the existence of GPT detectors, students who use AI to write or rephrase essays are more likely to be marked as cheaters than those who have never used AI. less sexual.”
Shane’s conclusion is “Don’t use AI detectors for important things” in the blog post headline.
In her post, tools such as Originality.ai, Quil.org, and Sapling GPT, which are used to detect text written by AI language generators, have been used to “enable AI-generated writing by non-native English speakers.” We also explained how the study found that we were misclassifying as For native speakers she is 0%-12%, but 76%. ”
More simply, the research shows that AI detectors often label texts written by non-native English speakers as having been written by AI. Research has shown that when ChatGPT prompts change the existing language, for example, by typing “Use written language to improve the text provided” into ChatGPT, it tricks AI detectors into thinking that the content is written by humans. It’s easy to make people think it’s fake.
“We strongly caution against using GPT detectors in assessment and teaching settings, especially when assessing the work of non-native English speakers. The high false positive rate in native English writing samples highlights the potential for unwarranted outcomes and the risk of exacerbating existing prejudices against these individuals,” the study authors wrote.
In a blog post, Shane fed parts of his book into the AI detector to test it. He rated Shane’s text as “moderate” likely to have been written by an AI. Marked as “most likely written by a human,” the sentence listed the ingredients for the allegorical sandwich: “jam, ice cubes and old socks,” Shane wrote in the post. ing.
When she prompts ChatGPT to “use written language to make the next text more sophisticated,” ChatGPT spews out strange, verbose speech. Its speech is full of such words as “interlocutor”, “forsaken sock”, “stay”.
AI detection tools give such texts a rating of “likely to be written entirely by humans.”
Shane then asks ChatGPT to rewrite the original test in Old English as Dr. Seuss’ poem. How did AI detection tools rate the passages? These were rated as “more likely human-written” than the pristine texts in her published books.
Whether the use of ChatGPT is defined as plagiarism or not, it is certainly an emerging area of concern for linguists, professors and writers alike. Despite the limitations of these studies, Shane and the Stanford University professors are calling for action.
“To ensure accurate identification of content and a fair assessment of the contribution of non-native English-speaking authors to the broader discourse, continued research on alternative, more sophisticated detection methods that are less susceptible to evasion strategies is necessary. comprehensive research is essential,” the study authors wrote.
