Multiagent Persuasion and Online Misinformation

In the study of complex multi-agent environments, understanding what incentive
structures result in cooperation is crucial to effectively encouraging cooperation and preventing
undue competition. Said understanding would be particularly useful in contexts in which large
groups of agents act largely independently from each other and in such a way where their actions
cannot be controlled, but rather indirectly influenced. One such real-world scenario is social
media communication online, as users act largely outside of the control of the platform holders
and various previous attempts to control their behavior have been ineffective. The study of this
indirect influence in the language of mathematics and economics is Bayesian persuasion, and
developing optimal strategies under a Bayesian persuasion setting is notoriously difficult. If one
were able to develop a learning algorithm for the Bayesian persuasion setting, one could apply
insights from it in the context of social media in order to incentivize certain user behaviors
My technical report covers the application of multi-agent reinforcement learning
algorithms to the Bayesian persuasion setting. To be more formally concrete, in the Bayesian
persuasion setting, one agent (the “Sender”) sends signals to another agent (the “Receiver”) and
each agent’s utility is determined by the actions of the Receiver and some “state of nature”
visible only to the Sender, usually some random value drawn from a set distribution. Thus, the
Receiver is “warily trusting” of the Sender to send true information based on the state of nature
that the Sender observes, and the Sender in turn can decide whether to send a true or false signal
based on that observation. The conclusion of our findings is that we can construct a state of
nature distribution such that agents using the EXP3-DH learning algorithm will fall into an
equilibrium in which the Sender will always truthfully report to the Receiver, regardless of the
utility values each agent receives for each action-state pair.
The STS research paper discusses the application of these insights to the context of social
media, with preventing misinformation as the primary target. The paper begins with an overview
of the existing social media landscape and the various attempts that have been previously made
to prevent the spread of misinformation on platforms like Twitter, and why such measures have
failed. I then go through various “intuitive” alternative proposals for ways to incentivize the
spreading of truthful information. Finally, I construct a proposal for social media platforms under
which users are incentivized to post truthful information (either through monetary means or onplatform content rewards). These incentives would be limited to a certain tier of users that are
initially selected manually by the platform holder, and afterwards would select themselves in
order to avoid potential abuse of this system to continue posting misinformation while still
getting the incentives
While the application of the results from the technical report are here limited to the
domain of social media misinformation, the potential applications are massive in scope, precisely
because they apply to any situation in which rational agents must indirectly influence each others
choices via aligned incentives. Despite this, social media misinformation is a notable real-world
problem that has only been growing in severity over the past few years. If the changes suggested
by the STS research paper were implemented into sites like Twitter or Facebook, the technical
research suggests that the spread of misinformation would be significantly reduced, clearly
demonstrating the research’s real-world value.