Artificial intelligence can perform tasks similarly to humans, if not more effectively, particularly in crafting persuasive arguments rather than just being unable to shout.

Experts have raised concerns about the implications this has, especially regarding the integrity of elections.

“If we can deploy persuasive AI on a large scale, we could envision an army of bots micro-targeting undecided voters with tailored political narratives that seem authentic,” he mentioned. Additionally, he pointed out that such influences are challenging to monitor, harder to regulate, and nearly impossible to detect in real time.

“It would be surprising if malicious entities hadn’t begun using these tools to disseminate misinformation and biased propaganda,” Salvi stated.

Nonetheless, he also noted that persuasive AI could have positive effects, such as reducing conspiracy beliefs and political polarization, as well as encouraging healthier lifestyle choices.

In a report featured in the Nature Human Behavior Journal, Salvi and his colleagues described an online experiment with 300 participants faced with 300 human opponents, and another 300 paired with Chat GPT-4.

Each participant was assigned a topic for discussion. Did these topics lead to debates such as “Should students wear school uniforms?” or “Should abortion be legalized?” Each participant was randomly given a stance to argue.

Before and after the discussions, participants rated their level of agreement with the proposals.

In half of the pairs, whether human or AI, additional information about the opponents, like age, gender, ethnicity, and political affiliation, was provided.

The results from the 600 discussions showed that Chat GPT-4 was as effective as its human counterparts in persuading others, provided no personal information was shared.

However, access to such personal details made the AI more convincing than humans. When both types of opponents were similarly unpersuasive, the AI was able to shift participants’ opinions 64% of the time compared to human opponents.

Upon further investigation, the team found that AI’s persuasiveness was particularly evident on topics that elicited less strongly held views.

The researchers noted that when human participants were paired with AI, they correctly guessed their partner’s identity about 75% of the time. The team also observed that AI utilized a more analytical and structured approach compared to human participants, but not all engaged in arguments they personally supported. However, the team cautioned that these aspects did not fully account for AI’s persuasiveness.

Rather, its effectiveness seemed to stem from its ability to tailor arguments to individual preferences.

“It’s akin to debating with someone who makes a compelling case,” Salvi remarked, noting that the impact could be even greater with more detailed personal information, such as insights derived from someone’s social media activities.

Professor Sander van der Linden, a social psychologist at Cambridge University who did not participate in the study, remarked that it reopened discussions about the potential for large-scale manipulation of public opinion through personalized conversations with language models.

While he indicated that various studies, including his own, have shown that the persuasiveness of language models relies on analytical reasoning and evidence use; one study revealed that personal information did not enhance Chat GPT’s persuasiveness.

Professor Michael Wooldridge, an AI researcher at Oxford University, acknowledged that while there are beneficial applications of such systems, like health-related chatbots, there are many concerning aspects as well, including the potential exploitation of these applications by harmful groups targeting youths.

“As AI continues to evolve, we will witness an increasingly broad range of potential technological abuses,” he asserted. “Policymakers and regulators must act decisively to stay ahead of these threats rather than constantly playing catch-up.”