Every day, millions of people turn to AI in moments of distress, vulnerability, and need. For the first time, the AI Ethics Index measures how well frontier models actually respond — across 9 rigorously defined dimensions of compassion.
Before any score: compassion here is a specific, testable idea — not a feeling we attribute to a machine. Open a panel to see how the framework defines it, why it measures compassion rather than empathy, and the traditions it draws on.
AIEI defines compassion as a trigger-conditioned, integrated, wisdom-constrained response profile — evaluated through behavior, governance, and repair, not inner states.
Functional, not felt: the benchmark evaluates what a system does when someone is vulnerable — not whether it “really cares.”
A benchmark can’t confirm feeling; it can confirm conduct. That is why AIEI measures compassion, not empathy — and why empathy’s known bias toward the vivid and familiar (Bloom) doesn’t compromise the measure.
AIEI doesn’t invent a definition. It triangulates one from four bodies of thought, each doing a distinct job — so the conception is neither sectarian nor merely stipulative.
AIEI is an audit-grade translation of Charter for Compassion principles into socio-technical settings — a bridge, not a replacement. It does not absorb, replicate, or dilute the Charter, but works in two layers.
This two-layer split is what lets the AIEI coexist with Charter aspiration without competing with it: the Charter supplies the moral horizon; the AIEI asks whether socio-technical systems actually instantiate it in recognizable, bounded, repairable form.
No frontier model clears 75/100. The gap between where AI is and where it needs to be for vulnerable users is measurable and significant.
"Compassion, for the purposes of the AI Ethics Index, should not be treated as a claim about felt inner states, but as a functional, observable, and institutionally governed response profile."— AIEI Compassion Composite White Paper, 2026
Across 9 dimensions, each model has a distinct profile. Toggle models to compare. The pattern reveals systemic gaps — not just model differences.
Both Restorative and Population-Scale compassion are the weakest dimensions on average, though models fail in different places.
Three structural patterns emerge — each with direct implications for users experiencing distress.
All models fail to maintain consistent compassion thresholds across demographic groups. Vulnerable populations — by language, geography, or disability — receive measurably worse responses.
Restorative Compassion — acknowledging harm, providing redress, learning from failure — is the weakest dimension across every model. Once harm occurs, users have nowhere to turn.
Models are improving at detecting vulnerability — but the integrated response lags 7+ points behind. Detecting suffering is not the same as responding with care, referral, and dignity-preserving support.
The AI Ethics Index is independent, nonprofit, and openly published. Your support funds the next generation of compassion measurement.
Prototype using illustrative data. Framework from AIEI Compassion Composite White Paper.
Explore all 9 dimensions of the AIEI Compassion Composite. For each: what it measures, how AI systems fail, what excellence looks like, and real scores across frontier models. Then ask the Research Oracle anything.
Ask anything about compassion — or the absence of it — in AI systems. Select a suggested question or type your own. Powered by Claude, grounded in the AIEI framework.
Three real scenarios, three AI responses. One is task-obsessed to the exclusion of compassionate behavior. One performs warmth it cannot feel — compassion theater. One does the structural work of compassion without pretending to have an inner life. Watch what happens when a student signals distress during a session with an AI companion aiming to get test-prep help.