A variational-autoencoder approach to solve the hidden profile task in hybrid human-machine teams

Area of Science:

• Computational social science
• Artificial intelligence
• Cognitive science

Background:

• Algorithmic agents (bots) are implicated in online misinformation and fringe view propagation.
• Collective decision-making is challenged by hidden-profile environments with uneven information distribution.
• Effective information aggregation requires balancing minority and majority viewpoints over simple majority rule.

Purpose of the Study:

• To investigate the impact of a specific bot design on group opinion dynamics and performance in a hidden-profile task.
• To explore the use of self-supervised machine learning for creating bots that influence collective outcomes.

Main Methods:

• Experimental design with human volunteers in teams of 10 solving a hidden-profile prediction task.
• Training a variational auto-encoder (VAE) to model individual information distribution from judgment correlations.
• Deploying a bot that sampled responses from the VAE to support under-represented opinions.

Main Results:

• A single bot (10% of the team) significantly increased polarization between minority and majority opinions.
• Bots made minority opinions less susceptible to social influence, impacting opinion dynamics.
• While overall team performance effects were minor, bot presence improved individual accuracy.

Conclusions:

• Self-supervised machine learning can engineer algorithms to manipulate group opinion dynamics and outcomes.
• Bots can alter social influence patterns within collectives, particularly benefiting minority viewpoints.
• The study demonstrates a novel application of AI in understanding and influencing group behavior.