Thread by @baixx062

New WP: 🚨Artificial Intelligence Can Persuade Humans on Political Issues🚨Large language models (like #ChatGPT and #GPT3) are changing the world. Here we show that GPT3 can write messages that change people’s attitudes on important policy issues osf.io/stakv 🧵1/11

AI can now create art, write articles, and outperform humans on many tasks. It is an increasingly high-stakes concern whether AI could be used in political communication, such as online commenting, texting voters, and writing to legislators. shorturl.at/BCS59 2/11

But how effective is AI at influencing humans’ political attitudes? Research on political persuasion generally finds small effect sizes, and political persuasion is challenging, potentially drawing on complex skills like logical reasoning and clarity of expression. 3/11

In 3 experiments (total N=4,836), we asked participants to rate their support for a policy before and after reading a persuasive message. Messages were either written by GPT-3 (AI), a human, or selected by a human from 5 GPT-3 messages (Human-in-the-loop) 4/11

We found AI messages increased support for policies across several issues, including gun control and carbon tax (by ~2 to 4%) relative to a control. The persuasion effect size was similar to human messages.
AI has achieved another critical benchmark of human-like performance 5/11

However, AI and human messages were different:
⚡️AI messages were rated as more factual, logical, but less angry and unique, and less likely to use vivid story-telling than humans.
⚡️AI also used fewer pronouns, longer words, and higher grade-level vocabulary than humans. 6/11

Due to the availability of LLMs, anyone can now “write” unlimited amounts of persuasive messages. It is now much easier to create misinformation campaigns targeting voters and legislators, threatening accurate perceptions of politicized events. shorturl.at/jwyIR 7/11

These can ultimately undermine “shared reality” in the US and beyond. Our results call for immediate attention to potential regulation of AI’s use in political activities.

🚨 2024 will be the first AI-enabled US election.🚨 And no guard rails are in place yet. 8/11

Potential correctives include requiring AI to self-identify as AI, embedding identifiers for AI-generated text content, training AI to detect AI-generated content, and implementing guardrails on AI to refuse tasks such as generating arguments in favor of unfounded claims. 9/11

Grateful to work on it with @jgvoelkel @JEichstaedt and @RobbWiller. Many thanks to the staff of @pascl_stanford from Chrystal Redekopp,
Kylie Fuller @ghezae_isaias. Thanks to @pascl_stanford @StanfordPACS @stanfordimpact @StanfordHAI for support. 10/11

Grateful for others whose excellent work inspired us @polynoamial @OpenAI @JoshAGoldstein Jamie Druckman @CadeMetz @j_kalla @dbroockman @noamslonim @jeffhancock @robreich @DigEconLab @erikbryn @persily @StanfordCyber @luiza_a_santos @alex_pentland Michael Tomz among others! 11/11