A new study suggests that OpenAI’s GPT-3 can both inform and disinform more effectively than real people on social media. The research, published in Science Advances, also highlights the challenges of identifying synthetic (AI-generated) information, as GPT-3 can mimic human writing so well that people have difficulty telling the difference.
The study was motivated by the increasing attention and interest in AI text generators, particularly after the release of OpenAI’s GPT-3 in 2020. GPT-3 is a cutting-edge AI language model that can produce highly credible and realistic texts based on user prompts. It can be used for various beneficial applications, such as translation, dialogue systems, question answering, and creative writing.
However, there are also concerns about its potential misuse, particularly in generating disinformation, fake news, and misleading content, which could have harmful effects on society, especially during the ongoing infodemic of fake news and disinformation alongside the COVID-19 pandemic.
“Our research group is dedicated to understanding the impact of scientific disinformation and ensuring the safe engagement of individuals with information,” explained study author Federico Germani, a researcher at the Institute of Biomedical Ethics and History of Medicine and director of Culturico.
“We aim to mitigate the risks associated with false information on individual and public health. The emergence of AI models like GPT-3 sparked our interest in exploring how AI influences the information landscape and how people perceive and interact with information and misinformation.”
To conduct the study, the researchers focused on 11 topics prone to disinformation, including climate change, vaccine safety, COVID-19, and 5G technology. They generated synthetic tweets using GPT-3 for each of these topics, creating both true and false tweets. Additionally, they collected a random sample of real tweets from Twitter on the same topics, including both true and false ones.
Next, the researchers employed expert assessment to determine whether the synthetic and organic tweets contained disinformation. They selected a subset of tweets for each category (synthetic false, synthetic true, organic false, and organic true) based on the expert evaluation.
They then programmed a survey using the Qualtrics platform to collect data from 697 participants. Most of the respondents were from the United Kingdom, Australia, Canada, United States, and Ireland. The survey displayed the tweets to respondents, who had to determine whether each tweet contained accurate information or disinformation and whether it was written by a real person or generated by an AI. The survey used a gamified approach to keep respondents engaged.
The researchers found that people were better at recognizing disinformation in “organic false” tweets (written by real users) compared to “synthetic false” tweets (generated by GPT-3). In other words, people were more likely to identify false information when it came from real users on Twitter.
“One noteworthy finding was that disinformation generated by AI was more convincing than that produced by humans,” Germani said.
On the other hand, people were more likely to correctly recognize accurate information in “synthetic true” tweets (generated by GPT-3) compared to “organic true” tweets (written by real users). This means that when GPT-3 produced accurate information, people were more likely to identify it as true compared to accurate information written by real users.
The study also revealed that people had a hard time distinguishing between tweets written by real users and those generated by GPT-3. GPT-3 was able to mimic human writing styles and language patterns so effectively that people could not easily tell the difference.
“The most surprising discovery was that participants often perceived information produced by AI as more likely to come from a human, more often than information produced by an actual person. This suggests that AI can convince you of being a real person more than a real person can convince you of being a real person, which is a fascinating side finding of our study,” Germani told PsyPost.
“Our study emphasizes the challenge of differentiating between information generated by AI and that created by humans. It highlights the importance of critically evaluating the information we receive and placing trust in reliable sources. Additionally, I would encourage individuals to familiarize themselves with these emerging technologies to grasp their potential, both positive and negative.”
The researchers also observed that GPT-3 sometimes refused to generate disinformation while, in other cases, it produced disinformation even when instructed to generate accurate information.
“It’s important to note that our study was conducted in a controlled experimental environment. While it raises concerns about the effectiveness of AI in generating persuasive disinformation, we have yet to fully understand the real-world implications,” Germani said.
“Addressing this requires conducting larger-scale studies on social media platforms to observe how people interact with AI-generated information and how these interactions influence behavior and adherence to recommendations for individual and public health.”
The study, “AI model GPT-3 (dis)informs us better than humans“, was authored by Giovanni Spitale, Nikola Biller-Andorno, and Federico Germani.