As GPAI models grow in scale and complexity, their capabilities, i.e. the range of tasks or functions that they can perform competently, manifest as behaviors. Emerging capabilities are as yet poorly understood and characterized, leading to misconceptions about their definition, nature, predictability, and implications. Acknowledging that capability emergence is sociotechnical, not solely mathematical and computational, KnoWare aspires to make the “emerging capabilities prefiguration” the prime methodology for forecasting, explaining, assessing, and validating (pre-)emerging capabilities in GPAI systems.
Our goal is to shift GPAI capabilities understanding from reactive observation to proactive detection. The key concept is “prefiguration” — the detectable early signs that a capability is forming: partial behaviors, latent skills, calibration shifts, and enabling conditions. KnoWare builds tools and frameworks to spot these signs early and assess their associated risks.
KnoWare will utilize psychometrics-inspired practices and tests and explainable AI techniques to make gradual improvement visible rather than sudden, and avoid emergence illusions. For evaluating GPAI models, KnoWare will develop a novel benchmarking approach, beyond narrow task-based performance and aggregated metrics, that considers prefigured capabilities and includes open tasks that measure not just “what the model can do” but also “whether the emerging capability is aligned with human intent” (risk-aligned governance).
KnoWare aims to define milestones for capability development against human values (e.g. autonomy, democratic principles, ethics) and legal compliance, monitor capability-risk co-evolution, and provide guidance for systemic risk management under the AI Act for supporting safe, trustworthy GPAI deployment in high impact sectors.
