Generative AI and Cognitive Bias: Online Experiment on Fake News

How to cite

Ishihara, Takunori. 2025. "Generative AI and Cognitive Bias: Online Experiment on Fake News." AEA RCT Registry. AEARCTR-0015361. https://doi.org/10.1257/rct.15361-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Takunori Ishihara*

* Primary investigator

Registry ID
AEARCTR-0015361
DOI
https://doi.org/10.1257/rct.15361-1.0
First registered
Last updated
Status
Completed
Trial dates
2025-02-07 – 2025-02-12
Countries
Japan ()
Registry record
http://www.socialscienceregistry.org/trials/15361

Abstract

This study examines the effect of generative AI use on cognitive biases. Specifically, it empirically investigates the extent to which judgments on fake news are influenced by the use of generative AI, such as ChatGPT.
Generative AI is a technology that utilizes large-scale language models to automatically generate new content, including text, images, and audio, based on existing data. ChatGPT, in particular, can generate text responses to questions and continuously improve its dialogue accuracy by learning from new data. This technology is being considered for adoption by businesses and governmental organizations to enhance work efficiency and generate new ideas, and its implementation is expected to increase in the future.
However, it is known that generative AI has an issue called "hallucination," where it produces non-existent data or information. As a result, outputs generated by AI may include inaccuracies, posing a potential problem in its application. Additionally, the use of generative AI could lead users to unintentionally create and disseminate misinformation through social media.
If humans could process information correctly, the impact of fake news would be minimal. However, cognitive biases make it difficult for individuals to detect misinformation accurately. Several reasons can explain this difficulty, such as confirmation bias, where individuals focus only on information consistent with their prior beliefs, or inattentional blindness, where critical information is overlooked. Traditional research has often employed nudging techniques to correct such cognitive biases (Pennycook et al., 2021). However, the effectiveness of nudges is generally low, averaging around 2% (DellaVigna and Linos, 2022).
Although generative AI has hallucination issues, it may also help reduce cognitive biases when used effectively. Previous studies indicate that AI can sometimes outperform humans in systematically error-prone tasks (Chen et al., 2023). Additionally, since generative AI enables interactive use, its personalized nature may result in higher acceptance by users, making it potentially more effective than traditional nudges.
Therefore, this study empirically examines whether using generative AI improves individuals' ability to discern fake news accurately.

Intervention

ChatGPT-generated responses to each news item
ChatGPT usage

Experimental design

This study will conduct an online survey experiment in the winter of 2025, with an expected sample size of approximately 1,200 participants. The subjects will be 1,200 individuals registered with a web survey company, selected based on the following criteria: Japanese residents aged 20–65, with equal representation of genders and evenly distributed age groups. Individuals who have never used generative AI will be excluded from the survey.
Participants will be randomly assigned to one of three groups and will complete a questionnaire consisting of the following items:
1. Questions about generative AI usage
2. A set of ten true-or-false questions, including fake news items
3. Questions about demographic attributes
Respondents will receive a participation reward of 100 yen. Additionally, they will earn a 10-yen bonus for each correct answer in the true-or-false section, with a maximum bonus of 100 yen.

In the true-or-false task, participants will assess whether each news item is real or fake. These items are derived from fact-check articles provided by the Japan Fact-Check Center (https://www.factcheckcenter.jp/). Participants will answer each question using a three-choice format (True / Unsure / False). The order of the questions will be randomized.
Participants will be assigned to one of three groups, each subjected to a different intervention condition:
1. Control Group (C): No AI intervention. Participants will simply judge whether the news is fake or not.
2. AI Information Group (T1): Participants will be shown ChatGPT-generated responses to each news item before making their judgment. The provided information will be factually accurate.
3. AI Usage Group (T2): Participants will personally interact with ChatGPT and use its responses to determine whether the news is fake or not. They will also submit their input and the AI's response.

Design details

Randomization method
randomization done in office by a computer
Randomization unit
individual
Clusters
1
Observations
1,200
Arms
1,200 individuals
Primary outcomes
Correct answer rate

Cited by

No citation data available for this registration yet.

Keywords