AI chatbots may help people evaluate news at first, but new research from Patti Mays and colleagues at the MIT Media Lab suggests that relying on artificial intelligence could weaken users’ ability to identify misinformation over time.
In a four-week study, participants who used an AI chatbot to assess pairs of news headlines and images were initially 21% more accurate at distinguishing fake news from legitimate news. However, by the fourth week, participants were 15% less accurate at identifying misinformation without AI than they had been at the beginning of the study. About one-quarter of participants reported that their independent fact-checking skills had improved. The findings highlight the “AI dependency paradox,” a pattern also observed in fields such as healthcare.
“Users get excited about these ‘magic’ LLMs, but forget that they are just statistical models that predict the next ‘token’ in a sequence,” said Ankh Rani, a doctoral student in Media Arts and Sciences and one of the study’s lead authors, along with fellow MAS doctoral student Valdemar Danley SM ’23.
The researchers also found that the chatbot’s communication style influenced long-term learning. AI systems that “tell” users the answer directly may build trust and speed up decision-making, while systems that “ask” questions using the Socratic method can better help people develop their own ability to evaluate information and detect fake news.
“AIs that ‘tell’ by providing direct answers are more likely to foster trust, while AIs that ‘ask’ through Socratic questioning are better at actually learning how to discern the truth for themselves,” Danley said. “But it’s a trade-off between speed and effort.”
Source: www.technologyreview.com