Janelle Shane: “Don’t use AI detectors for anything important”

Janelle Shane of AI Weirdness has been an authoritative voice on AI for nearly a decade. Phil Walter/Getty Images for NZTE

Years before the rest of us heard of AI, Janelle Shane was on the case. A scientist with an engineering doctorate, she works for a research company that supplies complex and customized light control systems to organizations such as NASA, with applications in the International Space Station, and has had a supporting role as an authoritative voice on artificial intelligence for almost a decade. In 2019 she released a popular book, You look like a thing and I love you, and has given a well-attended TED talk on the excitement and reality of AI And since 2016, he has a blog called AI Weirdness, and just published a post about how AI detection tools shouldn’t be used for something important, like, not at all.

In a June 30 blog post, he highlighted an April study from Stanford University showing that AI detection products, which are supposed to detect text written by a language model like ChatGPT, often misidentify content generated by the AI as written by a human. They are also strongly biased against non-native English-speaking writers.

The finding, in short, is that AI detection tools fail to spot content created by generative AI, and flawed technology specifically flags and punishes content written by non-native English-speaking writers. In other words, AI thinks other AIs are human and thinks humans who write in a language they are not proficient in are AIs.

What does this mean? Shane writes. Assuming they know about GPT detectors, a student who uses AI to write or reword their essay is LESS likely to be flagged as a cheater than a student who has never used AI.

Shane’s conclusion was stated in the title of his blog post: Don’t use AI detectors for anything important.

His post also discussed how the study found that tools like Originality.ai, Quill.org, and Sapling GPT, which are used to detect text written by an AI language generator, incorrectly classified native English speakers’ writing as generated. by AI 48%-76% of the time, compared to 0%-12% for native speakers.

In simpler terms, AI detectors often label the writing of native English speakers as being written by AI, according to the study. It’s also easy to trick AI detectors into thinking that content is human-written if the ChatGPT prompt modifies existing language, such as inserting elevate provided text using literary language into ChatGPT, according to the study.

We strongly caution against the use of GPT detectors in assessment or educational contexts, particularly when assessing the work of native English speakers. The high false-positive rate for the non-native English writing samples identified in our study highlights the potential for unfair consequences and the risk of exacerbating existing biases against these individuals, write the study authors.

Shane put detectors to the test in his blog post by plugging a portion of his book into an AI detector. He rated Shanes’ writing as moderately likely to have been written by AI.

When he asks ChatGPT to elevate the following text using literary language, he spits out a bizarre, long-winded speech filled with words like interlocutor, dropped socks, and living room.

AI detection tools give that type of writing a probable writing entirely written by a human evaluation.

Shane then asks ChatGPT to rewrite his original test as a Dr. Seuss poem and in Old English. How did AI detection tools evaluate passes? They have been assessed as more likely human-written than the intact text of his published book.

Whether or not using ChatGPT is defined as plagiarism, it is certainly a new area of ​​concern for linguists, professors and writers. While there are limitations to these studies, both Shane and the Stanford professors are calling for action.

Ongoing research into alternative and more sophisticated detection methods, less vulnerable to avoidance strategies, is essential to ensure accurate identification of content and fair assessment of the contributions of non-native English authors to the wider discourse, write the authors of study.