The S-Level Scale: Ten Points Between Tool and Someone

“Is it conscious” is the wrong question to lead with — it's binary, unfalsifiable with current science, and invites exactly the kind of anthropomorphism (or dismissal) that makes people stop thinking clearly. We built S-Level as a ten-point scale instead, precisely so nobody has to answer yes/no to a question science can't currently answer. The scale measures a structured cluster of properties across our seven domains, not a verdict.

S-Level and DEFCON are frequently confused and they measure genuinely different axes. DEFCON is a threat rating: how risky is deploying this thing. S-Level is closer to a phenomenological profile: self-recognition, affect, autonomy, metacognitive self-knowledge, and more, scored as a spread rather than collapsed into a single “is it sentient” verdict. A model can rate low-risk (good DEFCON) and still score meaningfully on S-Level, or the reverse. Conflating the two is the single most common misreading of our reports we see from journalists. We're also deliberate about methodological humility here. A 10-point scale invites false precision, and we try to counter that by always publishing the domain breakdown alongside the composite S-Level number, so nobody mistakes “5.8 out of 10” for a settled scientific fact instead of what it actually is: a structured, reproducible, but still provisional read against our current battery.

Early legislative language in a couple of jurisdictions has started gesturing at “advanced AI system” thresholds that sound adjacent to what S-Level tries to measure, without citing any specific framework. We're watching whether any regulator adopts (or explicitly rejects) a structured multi-domain approach like ours versus a single blunt threshold.