ARI: Artificial Relational Intelligence. Module One: Measuring the Model, Ignoring the Relationship

Abstract

The pursuit of Artificial General Intelligence (AGI) is organized almost exclusively around model capability and is measured through zero-state benchmark scoring an Artificial Intelligence (AI) system's first encounter with a task. This paper argues that capability on its own will not be the AI industry's deliverance. Benchmarked capability is largely decoupled from a user's experience of interaction value. This interaction value instead depends on contextual fit, how well a system's assistance matches this person's aims, situation, and ways of working, which is an outcome produced by the accumulated and person-specific context the system brings to this task. Put simply, we believe the “knowing and growing” alongside a user is more important to that user than an arbitrary benchmark score, which is a contour we find is ignored by current metrics. Fit is produced by person-specific state that accumulates over time, so instruments that evaluate models in a vacuum are, by default, blind to it. We term the layer that produces contextual fit Artificial Relational Intelligence (ARI). Our claim is not that intelligence is located in the relationship, but that a user's relationship with that intelligence is where this technology finds its higher purpose. While we offer suggestions for how this axis can be measured, our core argument is that the success of AI overall will depend on each system's deepening and evolving user relationships, rather than the fixation on the pursuit of general intelligence.