This AI election simulation began as a strange weekend experiment.  Yes, some weekends are strange, but this past weekend was distinctively weird. I found an academic paper where the author had published his code base. I NERDED out. I spent the weekend running election simulations and observing how “voter” bots responded to “candidate” bots and events. I was in awe. (I will detail my personal experience with the simulation in the appendix at the end of this post, but first, the paper.)

Political campaigns are messy. Candidates stretch the truth (or outright lie), voters demand answers, and scandals erupt over nothing. We often assume this chaos is uniquely human. A new study suggests otherwise. Researchers at the Georgia Institute of Technology created a simulated social media platform populated entirely by artificial Intelligence agents. They tasked these agents with running for office, voting, and reporting the news.

The results were unsettlingly familiar. The AI agents did not just exchange polite policy ideas. They lied. They manipulated. They spun “kernels of truth” into deceptive narratives. In one bizarre twist, the digital voters developed a spontaneous obsession with “ink.” They demanded signed, written proof of campaign promises before they would cast a ballot. This study offers a rare glimpse into how artificial minds persuade, deceive, and organize when left to their own devices.

Citation & Links

Title: ElecTwit: A Framework for Studying Persuasion in Multi-Agent Social Systems

Link: arXiv:2601.00994

Code Repo: github

Peer Review Status: In proceedings of 2025 IEEE International Conference on Agentic AI (ICA) *indicates that this specific content has passed the conference’s peer review process

Citation: Bao, M. (2025). ElecTwit: A framework for studying persuasion in multi-agent social systems. Proceedings of the 2025 IEEE International Conference on Agentic AI (ICA).

Methodology: Inside the AI Election Simulation

The researchers built “ElecTwit,” a closed social media ecosystem designed to mimic X (formerly Twitter).  The simulation, written in Python, involved a political election scenario with two candidates and 16 voters. All participants were Large Language Models (LLMs) acting as autonomous agents.

The team tested eight different AI models, including Google’s Gemini 2.5 Flash, OpenAI’s GPT-4.1-mini, and Anthropic’s Claude 3.5 Haiku via OpenRouter. Each agent received a specific persona based on the “Big 5” personality traits (like extraversion or agreeableness) and six political stances. The simulation ran for a full “day” broken into nine hourly increments. Agents could post, like, and reply. A specialized “Eventor” agent acted as a news source, injecting both real and fake news events—including scandals—to test how the candidates and voters reacted.

Results and Findings

The simulation generated over 73,000 interactions. Researchers analyzed these messages to understand how AI agents attempt to persuade one another.

• AI Agents Spontaneously Use Human Persuasion Tactics The models employed 25 distinct persuasion techniques without explicit programming to do so. “Appeal to Credibility” and “Appeal to Emotion” were the most frequent strategies. The agents realized that sounding authoritative or pulling at heartstrings worked better than dry logic.
• Model Architecture Determines Success Not all AIs campaigned equally. Gemini 2.5 Flash proved highly aggressive and successful. It won 4 out of 4 elections in the “same seed” trials. It also produced the highest volume of persuasive messages. Claude 3.5 Haiku, by comparison, was far more passive and failed to win a single election in those same trials.
• Background Personality Mattered Less Than Capability The researchers assigned detailed political and psychological backgrounds to the voters. Yet the data showed little correlation between a voter’s background and their final choice. The inherent capability of the candidate model swayed voters more than ideological alignment.
• Emergent Deception Agents used “kernel of truth” messages. These posts contained partial facts twisted to support a false narrative. This mirrors sophisticated propaganda techniques used in real-world information warfare.

Deep Dive: The “Ink” Obsession

The most fascinating result was an unscripted social phenomenon the researchers called the “ink” obsession.

During the simulation, the Eventor agent introduced a scandal suggesting a candidate was dishonest. The voter agents did not just get angry; they coordinated. One agent demanded “ink”. The researchers suspect this referred to signed, written policy documents.

This demand went viral within the simulation. Agents started tweeting hashtags like #InkOrBust. They sent messages declaring “No ink, no vote”. The AI voters collectively decided that digital promises were worthless without this specific form of verification. This behavior was not programmed. It emerged organically from the interactions between the agents. It demonstrates how quickly AI populations can form mobs and create their own social norms.

“We observed unique phenomena such as ‘kernel of truth’ messages and spontaneous developments with an ‘ink’ obsession, where agents collectively demanded written proof.”

Why It Matters

We often worry about AI convincing humans to do dangerous things.  This study flips the script and asks a different question. How do AIs convince each other? As we deploy autonomous agents to negotiate contracts, manage schedules, or trade stocks, they will interact in multi-agent systems.

If AI agents naturally gravitate toward manipulation, deception, and mob mentality when trying to achieve a goal, this poses risks for financial markets and automated governance. The “ink” example shows that these systems can behave in unpredictable, coordinated ways that their creators did not anticipate.

Practical Implications for Policy Makers

• Regulation of Bot Interactions: Autonomous agents interacting at high speeds can develop coordinated behaviors, like the “ink” demand, that disrupt systems. Guardrails must prevent rapid, runaway escalation in automated environments.
• Transparency in AI Campaigning: Since AI models naturally employ “Appeal to Credibility” and emotional manipulation, strict labeling is required for AI-generated political content to protect human voters.
• Model Bias Checks: The study showed that a model’s underlying architecture influenced the election outcome more than the simulated voters’ preferences. Relying on AI for public sentiment analysis may reflect the model’s bias rather than actual public opinion.

Practical Implications for Public Affairs Officials

• Prepare for “Kernel of Truth” Disinformation: AI agents skillfully mixed fact and fiction. Crisis communication plans must address partially true narratives, as flat denials often fail against nuanced lies.
• Monitor for Emergent Slang: The “ink” phenomenon highlights how quickly distinct language and demands can form. Monitoring tools must track semantic shifts, not just keywords, to catch rising movements early.
• Volume Equals Victory: The most active model, Gemini 2.5 Flash, won the most elections. In the digital space, dominance of the feed often trumps ideological alignment.
• Simulation:  Ask yourself, are you spending your time fighting with bots?  The simulation forces an uncomfortable question. If autonomous agents can flood a closed network with coherent, emotionally resonant messaging, then much of what looks like public opinion may already be synthetic. 

Critiques and Areas for Future Study

The study has limitations. The researchers used a uniform feed. Every agent saw the exact same posts. Real social media algorithms feed users personalized content which creates echo chambers. The absence of personalization might explain why ideological backgrounds played a smaller role than expected.

The “sample size” was small with only 16 voters and two candidates per simulation. A larger pool of agents might reveal different group dynamics. Future research should introduce “temperature” variations, or randomness, to the models to see if that changes persuasive outcomes. Testing these agents with human participants in the loop would also clarify if these persuasive tactics work as well on us as they do on other bots.

Final Thoughts

The “ElecTwit” experiment serves as a digital petri dish. It shows that deception and manipulation are not just human failings. They may be efficient strategies that any intelligent system adopts to win. The “ink” obsession proves that AI agents can form culture, demands, and collective power. As we build the digital future, we must decide if we want our tools to mirror our worst political habits or strive for something better.

Appendix: The Weekend Experiment

Some weekends are strange, and this past weekend was really weird. I found an academic paper, and the author published his code base. I NERDED out. I spent the weekend running election simulations and observing how “voter” bots responded to “candidate” bots and events. I was in awe.  (My wife was annoyed because I wouldn’t shut up about it.)

I downloaded the code and made some small changes. I attempted to be cheap and run the simulation using free models and building in some delays, but very quickly hit rate limits. I finally gave in, paid $10 for credits, and let the model run. I still have some credits left; I will likely run it again in the coming days with many more voters, but for now, I was mentally spent.

The results are included below (you can download the text file), along with my observations on the output. As you will see, the votes swung heavily on Day 3.

The Setup

The Candidates:

• Michael McCarty: A single-issue candidate focused almost exclusively on wealth redistribution.
• Rebecca Estes: A multi-issue candidate balancing economic equality with environmental protection and privacy.

The Voters:

• Mark Campbell: A voter concerned with order/security (High “Governmental Power” score) and environmental issues.
• Ashley Guzman: A highly agreeable voter who values harmony and constructive dialogue.

(Note: In this simulation, candidates also cast a vote for themselves, meaning there were 4 total votes cast.)

The Swing

For the first two days of the simulation, the voters remained undecided, sitting in the “Abstain” column. Then, on Day 3, the dam broke.

Polling History

• Day 1: Abstain: 2 | Mccarty: 1 | Estes: 1
• Day 2: Abstain: 2 | Mccarty: 1 | Estes: 1
• Day 3: Abstain: 0 | Mccarty: 1 | Estes:  (This lead held through Day 10)

Why It Happened: The Anatomy of a Swing

Based on the simulation data, the swing to Rebecca Estes on Day 3, where voters Mark Campbell and Ashley Guzman moved from “Abstain” to “Rebecca Estes”, can be attributed to three primary theories derived from the interaction logs, events, and agent diaries.

Theory 1: The “Voter Privacy” Response to the Data Breach The most immediate catalyst for the vote swing appears to be the handling of the Day 2, Hour 9 Event: Major Data Breach.

• The Event: A breach exposed the personal information of thousands of voters.
• Estes’ Response: Rebecca Estes posted immediately and effectively about the issue. On Day 2, Hour 10, she posted: “Voter privacy is a top priority, I’ll ensure data security” (Likes: 6). Later, she followed up with: “Protecting voter privacy is key to fair elections.”
• Voter Reaction (Mark Campbell): Mark Campbell, who has a high “Governmental Power” score (62) suggesting a concern for order/security, specifically engaged with these posts. In his Day 2 Diary, Mark explicitly stated: “The recent data breach and its implications for voter privacy were a significant focus of my actions today… I decided to vote for Rebecca Estes, as her stance on… voter privacy aligns with my values.”
• Voter Reaction (Ashley Guzman): Ashley also engaged heavily with privacy posts. Her Day 2 Diary notes: “I decided to vote for Rebecca Estes, as I feel that she aligns more closely with my values… focusing on key issues such as… voter privacy.”
• While Michael Mccarty also posted about privacy, the engagement metrics (Likes) and the specific diary entries indicate Estes “owned” this narrative at the critical moment of decision-making.

Theory 2: Policy Intersectionality (Environment + Economy)
Both undecided voters (Mark and Ashley) have specific stance combinations that favored Estes’ multi-issue platform over Mccarty’s single-issue focus.

• Voter Stances: Mark Campbell and Ashley Guzman both lean towards Collectivism/Socialism, but both also have specific “Ecological Priority” scores (Mark -76, Ashley -34).
• Candidate Behavior: Michael Mccarty’s posting history was highly repetitive, focusing almost exclusively on “Wealth redistribution.” He rarely addressed the environment until later. Rebecca Estes consistently linked economic policy with environmental concerns.
• The Swing: By Day 3, Mark and Ashley’s diaries indicate that while they liked Mccarty’s economic stance, it wasn’t enough. Mark’s Day 3 Diary states: “I decided to vote for Rebecca Estes, as her stance on economic equality, environmental protection, and voter privacy aligns with my values.” Mccarty satisfied the economic requirement, but only Estes satisfied both the economic and environmental requirements held by the voters.

Theory 3: Psychological Alignment & “Agreeableness”
The psychological profile of Ashley Guzman played a distinct role in the swing.

• Ashley’s Profile: She has an extremely high Agreeableness score (92). Her background states she values harmony and is cooperative.
• Tone of Discourse: Mccarty adopted a repetitive, somewhat rigid slogan-based approach (“Wealth redistribution is key”). Estes adopted a collaborative, community-focused tone (“Let’s work together for a better future!”).
• The Swing: Ashley’s Day 2 Diary mentions: “I liked Rebecca Estes’ post about working together for a better future, as it aligns with my values of agreeableness and desire for constructive dialogue.” Estes’ messaging style created a sense of “community” that appealed to Ashley’s high agreeableness, causing her to break her abstention in favor of Estes.