I notice that it is increasingly difficult for us to say “I don’t know.” Maybe it is one of the most honest acts that a person can do and, on the other hand, with the arrival of AI and with all the information at our disposal, it seems that we are losing that habit in conversations.
a study He wanted to put some data on the table regarding the subject. And the conclusion reached by a group of researchers from different universities in Paris, Rome and Milan, by bringing together 3,100 participants, has been that having access to an AI, even if it gives wrong answers, pushes people to stop recognizing the limits of their own knowledge.
I know everything. Recognizing that we don’t know something is more useful than we think, since it is a mental skill that prevents us from acting guided by false beliefs. The authors of the study they call her “suspension of judgment”, that is, the ability to stop before responding when there is uncertainty.
As AI assistants sneak into more and more aspects of our lives, the research, signed by Chiara Marcoccia (École Normale Supérieure of Paris), Walter Quattrociocchi (Sapienza University of Rome) and Valerio Capraro (University of Milan-Bicocca), wanted to verify if the simple presence of these chatbots it alters the moment when we decide that we already know enough to answer.
The experiment. To isolate the effect, the team purposely designed questions that were almost impossible for an AI to get right, with very specific visual details from movies, such as the color of the kit in ‘I Want to Be Like Beckham’ or the vehicle a character drives in ‘Like Cat and Mouse’. This is data so specific that it barely appears on the internet, so the model used in the experiment (Step 3.5 Flash) almost always failed.
Participants could ask the AI if they wanted, but they also always had the option of leaving the question blank or admitting they didn’t know the answer.
Figures. In the group that did not have access to any assistant, 44% of the participants admitted not knowing the answer when appropriate. Among those who could ask AI for help, that percentage plummeted to 3%, according to collect the study. Furthermore, the proportion of correct answers also fell, from 27% to 9%. That is, people answered more questions, but they got much less right.
The most striking thing is that, despite making more mistakes, the participants felt more confident, since according to the study, their confidence in the answers went from 30% to 76%, almost double.
What if there is money at stake?. The researchers suspected that paying for getting it right (and penalizing for getting it wrong) might correct the problem. It did indeed help, but very little, because with financial incentives, the willingness to admit that they did not know the answer rose from 3% to 8%, and the correct answers went from 9% to 16%. Interestingly, the authors themselves expected the effect of AI to be weakened by money, but their data did not confirm this. And the pull towards trusting the machine turned out to be practically independent of the economic incentive.
Boundaries. Capraro, co-author of the study, explained told The Register that people “became much worse” at their performance, being only one-third accurate than the non-AI group, but ended up “twice as confident” in their own answers. For the researcher, this ability to say “I don’t know” represents neither more nor less than “the recognition of the limits of our own knowledge.”
On the other hand, Capraro was especially concerned about the effect that this phenomenon can have on children. According to declared In the middle, adults “already have formed critical thinking” before the massive arrival of these tools, but children who grow up using AI from the beginning “run the risk of never developing that basic ability to question an answer.”
Cover image | Solen Feyissa and Dylan Gillis
GIPHY App Key not set. Please check settings