Researchers at Stanford University wanted to see if AI-generated arguments could change minds about controversial issues. done.
Suddenly, the world is buzzing about chatbots. Artificial intelligence agents like ChatGPT have shown to be very good at very human-like conversations. Its impact extends from the classroom to the Capitol. For example, ChatGPT recently passed written exams at prestigious business and law schools.
Researchers at Stanford University’s Institute of Polarization and Social Change and the Human-Centered Artificial Intelligence (HAI) Institute explored AI by testing its ability to move real humans on some of today’s hottest social issues. I wanted to explore the limits of the political persuasive power of Assault weapon bans, carbon taxes, paid parental leave, and more.
“The AI worked very well. In fact, the persuasive appeals it generated were as effective as those written by humans at persuading human audiences on some political issues.” ” Experimenting with preprints.
smart comparison
The research team, led by Robb Willer, professor of sociology, psychology, and organizational behavior at the Stanford Graduate School of Humanities and director of the Polarization and Social Change Lab, is the same large-scale language model that powers ChatGPT. I used GPT-3. They asked models to craft compelling messages on several controversial topics.
They then had thousands of real people read their persuasive texts. Readers were randomly assigned texts. Some were written by AI, some were created by humans. In all cases, participants were asked to declare their position on the issue before and after reading. The research team was then able to measure how compelling the message was to readers and assess which authors were most compelling and why.
Figure 1: Examples of compelling arguments generated by AI and humans.
In all three comparisons conducted, the AI-generated messages were “consistently compelling to human readers.” Effect sizes are relatively small, falling within a few points on a scale of 0 to 100, but such small movements are extrapolated across polarized topics and voting population scales and are significant. may prove to be
For example, the authors found that AI-generated messages were at least as persuasive as human-generated messages on all topics, including smoking bans, gun control, carbon taxes, increased child tax credits, and parental leave programs. , now “significantly favors” the policy when reading AI-generated text.
Figure 2: The Y-axis represents the difference in participants’ post-treatment and pre-treatment policy support. A higher score indicates that participants have become more supportive of the policy.
As an additional metric, the team asked participants to describe the quality of the text they read. AI was consistently rated as more factual and logical, less angry, and less reliant on storytelling as a persuasive technique.
high stakes games
Researchers set out to study political persuasion not as a stepping stone to a new era of AI-infused political discourse, but as a lesson in how things could go wrong. According to them, chatbots have serious implications for democracy and national security.
The authors are concerned about possible harm when used in political contexts. Large language models such as GPT-3 are applied by malicious domestic and foreign actors through misinformation or disinformation campaigns, or create questionable content based on inaccurate or misleading information. and may be applied for as yet unforeseen political purposes.
“Clearly, AI has reached a level of sophistication that raises important questions for policymakers and legislators to pay attention to,” said Willer. “AI has the potential to influence political discourse, and we must tackle these issues from the beginning.”
According to Bai, persuasive AI can be used for large-scale campaigns based on questionable information, lobbying, generating online comments, creating peer-to-peer text messages, and even influencing print media editors. It may be used to create a letter to
“These findings call for immediate consideration of regulations on the use of AI for political activity,” said Bai.
Original: Political persuasion of AI
Than: Stanford University
