Fact-checking information from large language models can decrease headline discernment.

Area of Science:

• Information Science
• Human-Computer Interaction
• Artificial Intelligence

Background:

• Fact-checking is crucial for combating online misinformation.
• Scaling fact-checking is challenging due to information volume.
• AI language models show potential for automated fact-checking.

Purpose of the Study:

• Investigate the impact of AI-generated fact-checks on belief and sharing intent for political news.
• Compare AI fact-checks with human-generated fact-checks.
• Identify potential harms and benefits of AI fact-checking.

Main Methods:

• Preregistered randomized controlled experiment.
• Assessed participants' belief and sharing intent for political headlines.
• Utilized fact-checking information from a popular large language model (LLM).

Main Results:

• LLM fact-checks did not improve headline accuracy discernment or sharing of accurate news.
• Human-generated fact-checks enhanced discernment.
• AI fact-checks decreased belief in true headlines mislabeled as false and increased belief in uncertain false headlines.
• AI fact-checking increased sharing intent for correctly labeled true headlines.

Conclusions:

• AI fact-checking information does not significantly improve users' ability to discern accuracy.
• AI fact-checking can introduce specific harms, such as misinformed beliefs.
• Human fact-checks are more effective in enhancing discernment.
• Policies are needed to mitigate unintended consequences of AI fact-checking applications.