Troll Farms

1. Problem Statement

The paper investigates the impact of coordinated disinformation campaigns (specifically "troll farms") on electoral outcomes. Unlike traditional propaganda, troll farms utilize social media to:

1. Mimic genuine signals: They send uninformative messages that are indistinguishable from voters' exogenous, informative private signals.
2. Microtarget: They exploit platform data to tailor the intensity and content of messages to specific voter types.
3. Scale cheaply: They can deploy vast numbers of fake accounts at near-zero marginal cost.

The central research question is: How does a sender (a troll farm) optimally manipulate information to maximize the probability of a preferred political outcome (e.g., re-electing a government) when voters possess private, imperfect information and cannot distinguish between genuine signals and troll-generated noise?

2. Methodology

The authors develop a constrained information design model (a variant of Bayesian persuasion) with the following structure:

• Agents:

  • Voters: A continuum of mass 1, characterized by a type $x \in [0,1]$ representing their political preference. A voter votes for the government if their posterior belief that the state is "good" ($\theta=1$) exceeds their type $x$.
  • Sender: A troll farm aiming to maximize the share of votes for the government.

• Information Structure:

  • State: Binary $\theta \in \{0, 1\}$ (Government competent or not).
  • Private Signals: Each voter receives an independent, continuous signal $s$ drawn from distributions $F_0$ and $F_1$ satisfying the Monotone Likelihood Ratio Property (MLRP).
  • Sender's Action: The sender cannot create new information but can garble existing information. For each voter type $x$, the sender chooses:
    • Intensity ($\alpha_x$): The probability that the voter receives a signal from the troll farm instead of the genuine signal.
    • Slant ($\tilde{F}_x$): The distribution of the troll-generated signals.
  • Constraint: Voters observe a signal but cannot distinguish its source. The sender's induced information structure must be a "garbling" of the exogenous signals.

• Equilibrium Concept: The authors solve for the sender's optimal strategy $(\alpha_x, \tilde{F}_x)$ that maximizes the expected vote share, subject to the constraint that voters update beliefs rationally (Bayesian updating) based on the mixture of genuine and troll signals.

3. Key Contributions

1. Source Uncertainty in Information Design: Unlike standard Bayesian persuasion (where the sender commits to an experiment) or signal jamming (where the sender shifts the mean), this model features source uncertainty. The sender mimics the receiver's private signal structure. This creates a unique constraint: the sender is restricted to information structures less precise than the exogenous signals, rather than more precise.
2. Counter-Intuitive Targeting Strategy: The paper derives a strategy where the sender sends pro-government messages to anti-government voters and anti-government messages to pro-government voters.
3. Paradoxical Effects of Information Quality: The paper challenges the standard intuition that better information always leads to better democratic outcomes. It demonstrates that higher signal precision can actually increase the sender's ability to manipulate outcomes.

4. Key Results

A. Optimal Strategy (Proposition 1)

The sender employs a dual strategy based on the voter's prior inclination:

• Pro-Government Voters ($x \leq 0.5$): These voters are initially inclined to vote for the government. The sender targets them with anti-government messages (low signal realizations).
  • Mechanism: By flooding these voters with low signals, the sender dilutes the impact of any high genuine signals they might receive. The sender adjusts the intensity and slant such that the posterior belief after any signal remains exactly at the voter's indifference point (or slightly above), ensuring they vote for the government regardless of the signal.
• Anti-Government Voters ($x > 0.5$): These voters are initially opposed. The sender targets them with pro-government messages (high signal realizations).
  • Mechanism: The sender sends high signals to increase the probability these voters receive a "favorable" signal. However, to prevent the voter from becoming too convinced (which would make the message suspicious or ineffective if the signal is too strong), the sender dilutes the credibility of these high signals. The strategy ensures that if the voter receives a signal $\geq s^*(x)$, their posterior belief is exactly $x$, making them indifferent but likely to vote for the government.

Result: The sender ensures that all pro-government voters vote for the government, while anti-government voters vote for the government if and only if they receive a signal above a certain threshold. Crucially, the sender sends both pro- and anti-government messages to different segments to maximize the total vote share.

B. Effect of Signal Informativeness (Theorem 1)

• Standard View: In standard information design, more precise independent signals make it harder for a sender to manipulate beliefs because the receiver's private information is harder to override.
• Paper's Finding: In this model, higher precision of independent signals increases the sender's payoff in both states.
• Reasoning: The sender is constrained to induce beliefs closer to the prior than the exogenous signals. As the exogenous signals become more informative (further from the prior), the "gap" between the prior and the signal widens. This relaxes the sender's constraint, allowing them to induce a wider set of posterior beliefs, thereby increasing their ability to manipulate the election outcome.
• Implication: Societies with high-quality information online are paradoxically more vulnerable to troll farms.

C. Effect of Polarization (Theorem 2)

• Finding: Increased polarization (a shift in the distribution of voter types $H$ toward the extremes, away from the median) reduces the sender's ability to manipulate outcomes.
• Reasoning: Polarized voters have stronger priors and require stronger signals to change their vote. This acts similarly to making signals less informative. The sender finds it harder to "garble" the information enough to sway extreme voters without making the manipulation obvious or ineffective.

D. Information Aggregation (Proposition 2)

• The presence of trolls can sometimes improve information aggregation (i.e., the government wins in state 1 and loses in state 0) when signals are moderately informative and polarization is low.
• However, if signals are very precise, trolls can cause the government to win in both states (a failure of aggregation), unless polarization is high enough to counteract the manipulation.

5. Extensions and Robustness

• Costly Trolls (Section 5.1): If sending messages is costly, the sender adjusts differently:
  • For pro-government voters: Reduces the extremity of anti-government messages (moderates the slant) but keeps targeting all of them.
  • For anti-government voters: Stops targeting moderate opponents but continues targeting extreme opponents with pro-government messages.
• Non-Manipulable Information (Section 5.2): If voters receive some information from offline sources that trolls cannot mimic, the core result (Theorem 1) holds: more informative manipulable signals still help the sender.
• Naive Voters (Section 5.4): If some voters are naive (fail to account for the troll farm), the sender's strategy changes (e.g., they stop sending anti-government messages to pro-government voters if it risks alienating naive ones). However, the comparative statics regarding signal precision and polarization remain robust.

6. Significance and Policy Implications

• Theoretical Contribution: The paper bridges the gap between information design and the study of disinformation, showing that the "source uncertainty" constraint fundamentally alters the sender's incentives compared to standard persuasion models.
• Policy Insight:
  • Regulation of Signal Quality: Efforts to improve the quality of information on social media (e.g., fact-checking, better algorithms) might inadvertently empower troll farms by widening the gap between prior beliefs and signals, making manipulation easier.
  • Polarization as a Defense: While generally seen as negative, high polarization can act as a natural defense against coordinated disinformation campaigns because extreme voters are harder to sway.
  • Efficiency Paradox: In some cases, the presence of a troll farm can move an election closer to the "efficient" outcome (where the government wins only when competent) if the baseline information is too noisy, though this is a fragile equilibrium.

In summary, the paper provides a rigorous mathematical framework demonstrating that the very tools designed to improve democratic decision-making (high-quality data, microtargeting) can be weaponized by bad actors in ways that are counter-intuitive to standard economic theory.