Reporter Holly Else in a news article for Nature:
An artificial-intelligence (AI) chatbot can write such convincing fake research-paper abstracts that scientists are often unable to spot them, according to a preprint posted on the bioRxiv server in late December.
So far so good. Per the preprint, researches collected 50 real abstracts, 10 each from JAMA, NEJM, BMJ, Lancet, and Nature Medicine, then asked ChatGPT to generate a new abstract out of each article’s title and journal name. They ended up with 100 abstracts, half of them AI-generated, that they were able to analyze using 3 methods: a plagiarism detector, an AI detector Or, to be more precise, the GPT-2 Output Detector. Note that ChatGPT is based on GPT-3., and blinded The preferred term nowadays seems to be masked over blinded, but either way you are bound to have funny-slash-distrubing associations pop into your head. human reviewers.
You can click through the link to read the outcomes, but per the pre-print’s own conclusion:
The generated abstracts do not alarm plagiarism-detection models, as the text is generated anew, Emphasis mine. but can often be detected using AI detection models, and identified by a blinded human reviewer.
So the “can often be detected” from the preprint itself becomes “often unable to be spotted” in the hands of a crafty human reporter. Gotcha.
Of course, no alarmist article is complete without some color comentary:
“I am very worried,” says Sandra Wachter, who studies technology and regulation at the University of Oxford, UK, and was not involved in the research. “If we’re now in a situation where the experts are not able to determine what’s true or not, we lose the middleman that we desperately need to guide us through complicated topics,” she adds.
We have always and forever will be in a situation where everyone — expert or not — had to engage their critical thinking to determine whether data presented are true and important, true but unimportant, true but misinterpreted, fragile, exagerated, overblown, or just plain fake. AI making it easier for the unscrupulous to do what they would have done anyway does not change the equation by an Earth-shattering amount.
Look, some people can’t handle a blank page but are good at editing, even if it means completely replacing the original text. In the olden days of 6 months ago trainees had no other recourse but to grind their teeth and just get on with it, hoping that at some point in their careers they will have trainees of their own writing those pesky first drafts. ChatGPT seems like a godsend for them. Whether what’s sent to journals for publication or posted on a pre-print server is real, fake, nonsense or profound still depends on the person doing the submitting.
Some side observations in no particular order:
- I have no issue with the pre-print itself, which I hope and trust will find a good home.
- Why does Nature deem the work important enough to cover in a news article, but not important enough to publish in one of its own journals?
- For an online news article, it is sadly lacking in that great breakthrough from six decades ago, the hyperlink. Even the URL for the pre-print itself is given as an un-clickable footnote. And no mention of the online and freely accessible plagiarism and AI detection tools.
- Nature’s news department is on a roll.