AI biased toward established scientific ideas and hypotheses

 "In November, Meta took down the public interface for Galactica, its scientist-specific large language model, just days after it was unveiled—users had identified myriad factual errors in the generated text. And a 2022 preprint study of Sparrow, an information-seeking chatbot developed by a Google subsidiary, found that up to 20% of its responses contained errors. AI text may also be biased toward established scientific ideas and hypotheses contained in the content on which the algorithms were trained. Journal editors also worry about ethics, suggesting authors who use text generators are sometimes presenting the outputs as if they wrote them—a transgression others have dubbed “aigiarism.”

ban listing a large language model such as ChatGPT as a co-author, to underscore the human author’s responsibility for ensuring the text’s accuracy. That is the case for Nature and all Springer Nature journals, the JAMA Network, and groups that advise on best practices in publishing, such as the Committee on Publication Ethics and the World Association of Medical Editors. But at least one publisher has taken a tougher line: The Science family of journals announced a complete ban on generated text last month. The journals may loosen the policy in the future depending on what the scientific community decides is acceptable use of the text generators, Editor-in-Chief Holden Thorp says. “It’s a lot easier to loosen our criteria than it is to tighten them.”

"Searching for papers to include in a systematic review may be a legitimate use if the researcher follows proper methods in deciding which papers to include, for example, she says, whereas cutting and pasting it into a perspective or opinion piece “is not OK because it’s not your perspective.” 

==  34 algorithms from five different companies

==prefers the original text versus the altered versions is consistently different for human-written versus AI-generated text, allowing DetectGPT to predict the likelihood that a sample came from a particular machine

== The company TurnItIn, which markets a widely used plagiarism detector, said last week it plans to roll out a synthetic text detector as early as April. TurnItIn says the tool, trained on academic writing, can identify 97% of text generated by ChatGPT, with a false positive rate of one in 100.

== robot-generated text

Ref: doi: 10.1126/science.adh2937