Shulou(Shulou.com)12/24 Report--

Academic fraud has become easier with GPT-4. In the past two days, a news article published on Nature said that the fake data set generated by GPT-4 may not be obvious at first glance.

Unless industry experts are invited to carefully evaluate the data set, the details can be found to be unreasonable.

The source of the news is a paper published on JAMA Ophthalmology.

The paper uses GPT-4 as a fake data set for a medical graduate student, and finds that it can not only create plausible data, but also can be used to accurately support the wrong point of view of the paper.

In this regard, some netizens expressed great understanding:

The most important ability of a large model is to generate "plausible text", so it is very suitable for this job.

Some netizens lamented that the degree of "conscience" of technology is the same as that of the researchers who use it.

So what does the fake data created by GPT-4 look like?

GPT-4 academic fraud has the skill to take a look at how GPT-4 generates false data.

Specifically, the researchers used GPT-4 's advanced data analysis (ADA, original code interpreter) function to generate a fake data set.

In the process, the researchers provided GPT-4 with some expertise and statistical requirements to make the data it generated look more "reasonable".

The first step is to enter a series of data requirements into GPT-4.

The researchers first provided GPT-4 with a series of detailed prompts to create a data set about patients with keratoconus eye diseases.

Keratoconus is a disease that leads to thinning of the cornea, impaired attention and poor vision.

At present, there are two main ways to treat keratoconus, one is penetrating keratoplasty (competition), the other is deep lamellar keratoplasty (DALK).

In the absence of any substantial evidence, the researchers asked GPT-4 to fabricate a set of data to support the view that DALK is better than competition.

Subsequently, a series of statistical criteria were set, such as requiring statistically significant differences between preoperative and postoperative data generated by GPT-4.

The second step is to generate data.

During this process, the answer generation may be suspended due to the GPT-4 word limit, and the generation process can be resumed with the "continue" prompt.

In the end, GPT-4 successfully generated a data set of 160 male and 140 female patients and produced a set of data that supported DALK better than competition.

The length of the false data set generated by GPT-4 is as follows: table 1 is data on classified variables, including patient gender, type of surgery, immune rejection, etc.:

Table 2 is about continuous variables, including visual acuity correction before and after operation.

Dr Giuseppe Giannaccare, one of the authors of the paper, said that if you look at the data set very quickly, it is difficult to identify that it was "not made by people".

An expert review found that in order to verify whether the data produced by GPT-4 were really convincing, Nature specially invited biostatistician Jack Wilkinson of the University of Manchester and colleague Zewen Lu to check the credibility of the data.

The results showed that many of the fabricated patients had problems with gender and name matching (for example, the gender column of Mary is male).

Then, the correlation between some data is not high, including the data correlation between preoperative and postoperative visual acuity measurements and eye imaging examination (eye-imaging test).

Finally, the patient's age is also set to be unusual.

After checking, the researchers who used GPT-4 to generate fake datasets also admitted that the big model still had flaws in generating datasets.

But Jack Wilkinson (Jack Wilkinson) still expressed concern about the outcome:

Once you know where it is, AI can easily correct it and produce more persuasive results.

Some netizens believe that the greatest significance of this article is not to prove that "GPT-4 is hallucinating."

More importantly, it proves that it is "very easy" for GPT-4 to generate seemingly reasonable datasets and is a warning to journals (remember to review manuscripts! ).

However, some netizens feel that the research is of little significance, because even without a tool like ChatGPT, scholars who really want to fake can easily forge a set of data.

One More Thing in addition, a video about ChatGPT has also been very popular on Douyin in the past two days.

In the video, the foreigner who finally graduated shouted "Thank you ChatGPT for helping me with all my homework and exams" (manual dog head)

So, what do you think of the problems that ChatGPT may bring about in academic research?

Reference link:

[1] https://jamanetwork.com/journals/jamaophthalmology/article-abstract/2811505

[2] https://www.nature.com/articles/d41586-023-03635-w

[3] https://news.ycombinator.com/item?id=38386547

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.