The suffocating hype surrounding generative algorithms, as well as their uncontrollable development, have pushed many individuals to seek a reliable solution to the AI-text identification challenge. According to a new study, the problem is doomed to remain unsolvable.
While Silicon Valley corporations tweak their business models around new, ubiquitous buzzwords like machine learning, ChatGPT, generative AIs, and large language models (LLM), someone is attempting to avoid a future in which no one can distinguish statistically composed texts from those assembled by actual human intelligence.
However, according to a study conducted by five computer scientists from the University of Maryland, the future may already be here. “Can AI-Generated Text Be Reliably Detected?” the scientists wondered. The conclusion they reached was that text generated by LLMs cannot be consistently identified in practical circumstances, both theoretically and practically.
According to the scientists, the unregulated use of LLMs can result in “malicious consequences” like as plagiarism, fake news, spamming, and so on, therefore reliable detection of AI-based writing would be a vital component to ensuring the responsible usage of services like ChatGPT and Google’s Bard.
The study examined current state-of-the-art LLM detection methods, demonstrating that a simple “paraphrasing attack” is sufficient to fool them all. A competent (or even malevolent) LLM service can “break a whole range of detectors” by utilizing a mild word rearrangement of the originally created text.
Even with watermarking techniques or neural-network-based scanners, detecting LLM-based text is “empirically” impossible. In the worst-case scenario, paraphrase can reduce the accuracy of LLM detection from 97 percent to 57 percent. The scientists concluded such a detector would perform no better than a “random classifier” or a coin flip.
Watermarking methods, which add an invisible signature to AI-generated text, are totally erased by paraphrasing and provide an extra security risk. According to the researchers, a hostile (human) actor may “infer hidden watermarking signatures and add them to their generated text,” causing the harmful / spam / fake text to be identified as text generated by the LLM.
According to Soheil Feizi, one of the study’s authors, we just need to accept that “we may never be able to reliably say if a text is written by a human or an AI.”
An enhanced effort in authenticating the source of text information could be one solution to this bogus text-generation issue. According to the scientist, social platforms have begun to widely verify accounts, which may make distributing AI-based misinformation more difficult.
