jeu, Juin 29th 2023
(Keystone SDA) Chat GPT has a knack for creating persuasive misinformation – even more persuasive than what humans create and proliferate, according to a new study from the University of Zurich (UZH) published this week in the journal Science Advances.
For the study, the researchers at the University of Zurich asked 700 participants between the ages of 26 and 76 guess whether a tweet was written by a real Twitter user or by Chat GPT-3. Participants were also asked about their understanding of the tweets and whether they thought they were truthful or not.
The tweet topics included Covid-19, vaccine safety, climate change, and homeopathic treatments for cancer, among other trending subjects. The participants could not tell the difference between the tweets; moreover, they understood the tweets generated by Chat GPT more clearly than the human ones.
“GPT-3 is a double-edged sword,” the researchers wrote in the study. Human-composed tweets took longer, on average, for participants to determine whether a tweet was correct or not. According to the study, this shows that GPT-3 informs “more efficiently” than humans.
The participants were able to recognize misinformation created by humans with a 92% accuracy, but only 89% with Chat-GPT generated tweets.
These results indicate that information campaigns created by GPT-3 (but evaluated by trained people) would be more effective in crisis situations, for example, as the UZH researchers wrote.
However, the findings also show risks. The researchers therefore advocate strict, evidence-based and ethically sound regulations to counter the potential threat posed by these technologies.
French 
English 
German 
Spanish 
Arabic