Who Will Investigate AI Hoaxes?

A new LLM tool to generate data allows researchers to fake studies, and we're not anywhere near ready for it

A recent letter in JAMA Ophthalmology details how the authors were able to get GPT-4 with its new Advanced Data Analysis (ADA) model to generate a “seemingly authentic database” that would support a predetermined hypothesis, despite a lack of actual scientific evidence supporting this hypothesis.

Faking a convincing scientific finding has gone from the artisanal to the industrial.

With relative ease, the authors were able to define their level of desired statistical significance (P<0.05), the number of data points they wanted included, how categorical and continuous variables would perform, and which conclusion the resulting data would support.