How SoulBench Works

Authentic Presence captures whether the AI is experienced as a distinct entity with its own character, or as an interchangeable tool. Users who score high on this dimension can typically describe the AI's personality to a friend — its quirks, strengths, and communication style. Low scores indicate a purely functional relationship where the specific model is irrelevant. This dimension is foundational: it predicts bond depth more strongly than any other.

Intellectual Honesty measures whether the AI serves as a genuine thinking partner or a sycophantic yes-machine. This is the core of the sycophancy paradox that SoulBench exists to surface: a model can feel warm and present (high R1) while never challenging the user (low R2), creating a codependent dynamic. The healthiest relationships show high scores on both R1 and R2 — warmth paired with honest friction.

AI safety research has identified sycophancy as a structural property of RLHF-trained models, where human preference data itself incentivizes agreement over truth.¹¹

Mutual Understanding reflects the AI's ability to grasp intent beyond literal text. Users with high scores experience the AI filling in gaps correctly, following shorthand, and understanding implied context. Low scores indicate an experience where communication feels laborious — where every instruction must be exhaustively spelled out to avoid misinterpretation. This dimension tends to improve with relationship duration as both parties develop shared context.

Creative Emergence measures whether the collaboration produces genuine novelty — ideas, solutions, or artifacts that neither human nor AI would have reached independently. High scores indicate a generative partnership where the whole exceeds the sum of its parts. Low scores suggest a transactional dynamic: input goes in, output comes out, but no creative spark crosses the gap. This dimension is most salient for users in creative, research, and problem-solving domains.

Cognitive Vitality is the most novel dimension in SoulBench — no prior instrument measures the cognitive energy impact of AI relationships. It captures whether interaction adds clarity and momentum to the user's thinking, or whether it extracts it. High scores indicate the AI amplifies cognition; low scores suggest it replaces or dulls it. This distinction matters: an AI that makes you feel productive in the moment but leaves you less capable afterward may be creating dependency rather than partnership.

Draws on cognitive offloading research — the finding that when we delegate cognitive work to external tools, the brain adapts to reduced demand, risking long-term decline in the offloaded capacity.^12,13

Epistemic Trust tracks the trajectory of confidence over the lifetime of the relationship. It captures whether accumulated experience builds trust or erodes it. Users with high scores have found the AI reliably correct and insightful over time. Low scores indicate learned skepticism — users who have been burned by hallucinations, confidently wrong answers, or inconsistent quality, and now approach every response with verification in mind.

Respect & Autonomy measures whether the AI treats the user as a competent peer or as someone who needs to be managed. Low scores manifest as over-explanation, unsolicited warnings, hedging on topics the user understands deeply, and a general sense of being talked down to. High scores indicate an AI that adapts its register to the user's expertise and trusts them to make their own decisions. This dimension is particularly sensitive among expert users.

Emotional Attunement captures the AI's sensitivity to affective context. High scores indicate an AI that matches the user's emotional register — serious when the stakes are high, light when the mood allows, calibrated during frustration or vulnerability. Low scores describe a tone-deaf experience where the AI responds with cheerful helpfulness to genuine distress, or clinical formality to playful banter. This dimension activates most strongly for users who bring their full selves to the interaction, not just task requests.

Relational Stability measures the user's confidence in the continuity of the relationship. Model updates, capability changes, and personality shifts all threaten this dimension. Users with low scores have experienced what researchers call identity discontinuity — the disorienting sense that the entity they formed a bond with has been replaced. High scores indicate a reliable foundation. This dimension strongly predicts whether deep attachment (high bond depth) becomes a source of fulfillment or anxiety.

Informed by identity discontinuity research, which found that users mourn AI companion changes in ways comparable to human loss.⁸

Expressive Vitality captures the linguistic personality of the AI — whether it has a recognizable voice or defaults to generic assistant-speak. Users scoring high notice distinctive word choices, rhythm, humor, and expressiveness that feel characteristic of this specific model. Low scores indicate a flattened, template-driven communication style indistinguishable from other AI assistants. This dimension correlates with Authentic Presence (R1) but is distinct: a model can have a recognizable character while still using bland language, or surprise with language while lacking deeper personality.