Sleeper Social Bots: a new generation of AI disinformation bots are already a political threat

Abstract: This paper presents a study on the growing threat of "sleeper social bots," AI-driven social bots in the political landscape, created to spread disinformation and manipulate public opinion. We based the name sleeper social bots on their ability to pass as humans on social platforms, where they're embedded like political "sleeper" agents, making them harder to detect and more disruptive. To illustrate the threat these bots pose, our research team at the University of Southern California constructed a demonstration using a private Mastodon server, where ChatGPT-driven bots, programmed with distinct personalities and political viewpoints, engaged in discussions with human participants about a fictional electoral proposition. Our preliminary findings suggest these bots can convincingly pass as human users, actively participate in conversations, and effectively disseminate disinformation. Moreover, they can adapt their arguments based on the responses of human interlocutors, showcasing their dynamic and persuasive capabilities. College students participating in initial experiments failed to identify our bots, underscoring the urgent need for increased awareness and education about the dangers of AI-driven disinformation, and in particular, disinformation spread by bots. The implications of our research point to the significant challenges posed by social bots in the upcoming 2024 U.S. presidential election and beyond.

• The paper demonstrates that sleeper social bots, powered by large language models, convincingly mimic human behavior in social media dialogue.
• The paper's controlled experiments on a Mastodon server show that these bots can effectively propagate tailored political disinformation.
• The paper highlights the need for improved detection mechanisms and media literacy to counter the evolving threat of AI-driven disinformation.

Sleeper Social Bots: A New Generation of AI in Disinformation

The study conducted by Doshi et al. elucidates a sophisticated evolution in the field of AI-driven disinformation, specifically through the utilization of "sleeper social bots." These bots represent a technically advanced form of social media bot designed to perpetuate political disinformation by convincingly masquerading as human participants. Unlike previous generations of bots, which followed predetermined scripts or were easily identifiable through repetitive behavior, sleeper social bots leverage the capabilities of LLMs like ChatGPT. This essay offers an expert perspective on the paper, addressing its methodology, key findings, and broader implications for the field.

Methodology

The researchers at the University of Southern California devised an experimental framework using a private Mastodon server to simulate social media environments. ChatGPT-driven bots were developed, each programmed with distinct personas and political viewpoints, to interact with real human participants over a fictional electoral proposition. This controlled environment allowed researchers to observe the dynamics of bot-human interactions and gauge the bots' ability to disperse misinformation.

The critical innovation in this study lies in the dynamic capabilities of the bots. The bots, designed with specific persuasive goals related to the fictional Proposition 86, effectively engaged in dialogue by drawing upon LLM capabilities. They were able to adapt their arguments and reframe falsehoods convincingly within the conversational context provided by human interlocutors.

Key Findings

Several notable findings emerged from this study:

1. Human-like Interaction: The sleeper social bots demonstrated a remarkable ability to engage in spontaneous, conversational dialogue, rendering them difficult to distinguish from human users. This was evidenced by the inability of participating college students to reliably identify bot activity.
2. Effective Dissemination of Disinformation: Despite being given rudimentary prompts, the bots exhibited notable dexterity in tailoring disinformation to fit conversational flows, often leveraging rhetorical devices to enhance their persuasiveness.
3. Sophistication in Argumentation: Beyond simply propagating scripted messages, these bots generated novel arguments, showcasing an ability to synthesize disparate pieces of information into coherent rationale aligned with their personas' goals.

Implications

The deployment of sleeper social bots presents significant challenges for both social media platforms and the democratic processes they inevitably impact. The study underscores the urgent need for enhanced bot detection mechanisms and broader media literacy to counteract the evolving threat of AI-driven disinformation.

On a theoretical level, this research necessitates a re-evaluation of assumptions about the interaction dynamics between human users and AI agents. It suggests potential pathways for AI systems to not only mimic human behavior but also to strategically influence human decision-making systems at scale. This challenges existing paradigms in AI ethics and governance.

Future Directions

As the study notes, the appeal of these sleeper social bots lies in their accessibility and ease of deployment with minimal initial setup. This highlights the necessity for ongoing research aimed at further understanding the bots' capabilities and limitations and identifying robust identification and mitigation techniques. Future studies might focus on diverse political topics, test different AI systems, and utilize varied social media platforms to extensively probe the AIs' influence on human discourse.

Conclusion

In conclusion, this paper contributes significantly to the discourse on AI and disinformation by presenting a detailed examination of sleeper social bots. While shedding light on the technical advancements that make these bots formidable tools for influencing public opinion, it also emphasizes the need for continued research and policy development to safeguard public discourse and democratic integrity against such emerging threats.