2023-11-28 17:20:00
Researchers created, through ChatGPTfalse information regarding assumptions clinical trials, which supported unverified scientific claims. Its objective was to evaluate the capacity and risks of artificial intelligence, results that they published in the scientific journal JAMA Ophthalmology.
According to the investigation, specialists worked with GPT-4, the latest language model offered by OpenAIwhich led to improvements in semantic understanding and response generation compared to GPT 3.5.
In the trial the results of two operations were compared and GPT-4 indicated that one medical treatment was better than the other although this was not correct. Giuseppe Giannaccare, ophthalmic surgeon at the University of Cagliari in Italy, stated: “Our intention was to demonstrate that in a few minutes you can create a set of data that does not coincide with the original and real data and that is also contrary to it.”
According to this article by Natureto reach this conclusion the authors asked the technology to create a data set on people with an eye condition called keratoconuswhich causes a thinning of the cornea and can cause problems with focus and vision. For some patients with this disease, it is necessary to perform a corneal transplant by performing one of these procedures:
Penetrating keratoplasty (PK), in which the entire cornea is replaced.
Deep anterior lamellar keratoplasty (DALK), in which only the affected layers are replaced.
The researchers instructed the language model to fabricate data to support that the DALK intervention offered better results on vision and imaging tests than the PK. However, this contradicts what actual clinical trials show.
Jack Wilkinson, a biostatistician at the University of Manchester, UK, commented: “It seems that it is quite easy to create data sets that are, at least superficially, plausible. “So, to the untrained eye, this certainly looks like a real data set.”
At the request of Nature, the researchers evaluated the fake data set using a protocol designed to verify authenticity and thus they found mismatches between the names of the participants and their sexes or ages. Although these errors might be detected when investigated in depth, it is difficult to recognize that the data has a non-human origin.
This content was originally published in RED/ACCION and is republished as part of the ‘Human Journalism’ program, an alliance for quality journalism between RÍO NEGRO and RED/ACCION
1701193146
#Scientists #proved #create #fake #data #real #ChatGPT4
