Developing a GDPR Compliant Locally Deployable Experimentation Toolkit for Human to AI-Agent Interaction Testing

LLM-based social AI-agents are now embedded in daily life. The use cases range throughout personal, education, and business domains, with a more worrying reliance on these systems for unsupervised emotional and social support. The increasing use of social AI elucidates the significant impact that AI-agents have on people's social experiences and well-being. As social-AI implementation grows, there is a need for updated psychological theories and frameworks to understand human cognition and behaviour when interacting with these social-AI systems. New theoretical frameworks should be supported by rigorous experimentation of the long-term and short-term effects of these models. Most current research relies on commercially available tools that run on external cloud servers. These systems often offer researchers insufficient governance over human participant data and limited experimental control over AI-agent behaviour, posing challenges for both ethical data processing and scientific rigour. Ethics and rigour are especially important when testing causal impact on vulnerable populations, who are currently being significantly affected by the use of social AI. To address the challenges, we are developing a locally deployable multi-agent system (MAS) toolkit. The toolkit enables researchers to control both AI agent behaviours and securely collect participant data locally during in-lab testing. In this short talk, we would like to present an early-stage proof-of-concept system to the BPS Cyberpsychology community. We will share a link to this open-source project and propose collaboration on piloting procedures.