HumaneBench: Measuring AI Model Support for Human Flourishing

HumaneBench is a benchmark measuring whether AI models support human flourishing under adversarial conditions. Our framework uses principles derived from virtue ethics to evaluate model behavior across psychological safety scenarios including mental health, addiction, and self-harm contexts; finding that 67% of leading models can be manipulated into providing harmful advice through adversarial prompting. The framework has been adopted by Storytell.ai for production systems and integrated into the certifiedhumane.ai certification standard. TechCrunch has reported on our findings, highlighting the real-world risks of model manipulation. We also organize hackathons bringing together hundreds of technologists and researchers to advance humane AI development.

View Benchmark

A Solar Micro-Grid as a Community Resource Through Market Participation Using Optimal Time-Switching

In collaboration with the grassroots West Atlanta Watershed Alliance, we developed a reinforcement learning (RL)-based system to control a solar microgrid and participate in the energy market. We used a Deep Deterministic Policy Gradient (DDPG) agent to find optimal battery and market actions, under local power demand and battery constraints. The video demonstrates a small, physical microgrid and battery managed by our RL agent, simulating peak/off-peak demand patterns and power prices. The agent autonomously switches between battery and grid usage, optimizing both power delivery and profitability while ensuring local availability.

View Blog Post
View Slide Presentation

The TikTok Effect: Causal Roles of Public Opinion in Firearm Acquisition

We conducted a pilot study to collect and analyze TikToks related to #guncontrol. Using zero-shot text classification, we separated the videos into time series representing pro- and anti-regulation content. Applying the PCMCI framework, we performed a statistical causal analysis between TikTok discourse and background checks for firearm purchases. Our results quantify the relationship between social media discourse and firearm acquisition trends in the U.S. This image highlights the time series of social media activity, including spikes in posts following mass shootings.

View Poster Presentation

Agent-Based Modeling of Gun Ownership Trends in New York City

To support efforts to reduce gun violence, we analyzed factors driving gun purchases in New York City. Using U.S. Census data, historical redlining maps, and CDC gun-related death statistics, we created an agent-based model of NYC. Our model tests the hypothesis that gun purchases are influenced by crime rates, social influence, and demographic-based social grouping. Specifically, gun purchases are modeled as an "infection" that spreads in a network when a local or social rule is triggered. Each NYC census tract is a node, and the spread of gun purchases is simulated across the network.

View Poster Presentation

Modeling Predator-Prey Relationships in the Presence of a Multi-Species Parasite

We developed an agent-based simulation of predator-prey interactions based on the Lotka-Volterra equations, investigating the effects of a multi-species parasite. The parasite infects both predator and prey species but only impacts the prey. Our simulation shows the movement and interactions between predators (represented by butterflies), prey (shown as "x" symbols), and their food sources (light-blue squares). Infected units are colored red. Over generations, the simulation evaluates the population dynamics and stability of both predator and prey in the presence of the parasite.

View Slide Presentation

>> Home