AI model GPT-3 (dis)informs us better than humans

Area of Science:

• Digital Health
• Artificial Intelligence
• Information Science

Background:

• The current infodemic significantly impacts global health.
• Artificial intelligence (AI) is transforming information creation and evaluation.

Purpose of the Study:

• To assess human ability to distinguish AI-generated disinformation from accurate information.
• To determine if participants can identify tweets from AI (GPT-3) versus human users.

Main Methods:

• A preregistered study involving 697 participants.
• Analysis of tweets generated by AI (GPT-3) and humans.
• Evaluation of participant accuracy in identifying disinformation and AI-generated content.

Main Results:

• GPT-3 produces accurate information that is easier to understand than human-generated content.
• GPT-3 also generates more compelling disinformation compared to humans.
• Participants could not reliably distinguish between AI-generated and human-written tweets.

Conclusions:

• AI presents a dual risk in information environments, capable of both enhancing clarity and creating sophisticated disinformation.
• The inability of humans to detect AI-generated content necessitates strategies to combat AI-driven disinformation.
• Findings inform improvements for public health information campaigns in the age of AI.