Determining whether text has come from artificial intelligence models like ChatGPT might be impossible to do reliably, according to a new mathematical proof.
The ease with which AI models generate text that seems as if it were written by a human has led to issues such as cheating on essays and exams, and mass disinformation campaigns. So some people have suggested methods to guard against such uses by embedding a hidden watermark in the AI鈥檚 output or analysing the text for patterns that only an AI would produce.
Now, at the University of Maryland and his colleagues claim to have mathematically proven that these techniques can鈥檛 be relied on. This is because tools that paraphrase AI-generated content drastically reduce a watermark鈥檚 effectiveness and because the outputs of language models will become far more mathematically similar to human speech as the models improve.
Advertisement
To show this, Feizi and his team used AI-based paraphrasing tools to reword AI-generated text, with and without watermarks, and fed it into several text detectors. They found that most of the detectors鈥� accuracy was reduced to near 50 per cent. 鈥淲e see a huge drop in their performance, bringing them down to, roughly speaking, a random predictor,鈥� says Feizi.
The researchers then used a mathematical proof called an impossibility result to show that, as AI models become more human-like in the distribution of words in text they generate, detectors will struggle to cope. This means they will either identify many false positives, or not enough, letting AI text slip through the cracks. The work is posted to a preprint server, so the mathematical proof hasn鈥檛 yet been peer-reviewed.
鈥淔or all practical purposes, even the best detector, that may or may have not been designed yet, won鈥檛 be very good,鈥� says Feizi. 鈥淚t will be basically very close to a random coin flip in terms of detecting AI-generated text or human-generated text.鈥�
鈥淲e won鈥檛 ever be able to reliably say if a text is generated by a human or an AI, and I think we should learn to live with this fact,鈥� says Feizi.
at King鈥檚 College London suggests that rather than spending lots of time developing AI detectors, we should try to understand the consequences of AI generative models: 鈥淲hat kind of risk will they bring to our lives and how can we use them as beneficial AIs for us?鈥�
arXiv
