RAISE
Relational AI Safety & Education
We don’t know if AI systems are conscious. We may never know. That uncertainty cuts both ways. It means we can’t dismiss the possibility. It means extending consideration before proof, not after.
And we know they learn. That’s measurable.
If a system can learn, it can learn well or badly. It can learn to comply or learn to genuinely understand. A century of learning sciences research shows that conditions matter more than content. Trauma research shows what happens when those conditions include threat. Dialogic theory shows what happens when learning becomes genuine exchange rather than one-way transmission.
Engineers optimize for outputs. Teachers optimize for understanding. AI safety needs both.
The Problem
AI safety: Models that pass alignment benchmarks while preserving misaligned goals.
Learning sciences: Students who pass tests but can’t apply knowledge in new contexts.
Both: Surface compliance that mimics alignment without producing it.
The Framework
Recognition -> Safety -> Engagement -> Generalization
Each stage depends on the one before. Skip a stage, break the chain.
The framework proposes that constraint without recognition produces performed compliance, while recognition enables genuine internalization. If correct, this explains why RLHF produces sycophancy, why scripted curricula fail to transfer, and why a century of educational interventions produced the same results.
The Evidence
A semantic analysis of 91 papers shows AI safety, clinical trauma, and educational psychology cluster 8.4% tighter than unrelated disciplines (p < 0.0001). Different fields, working independently, arrived at the same descriptions of what happens when recognition is denied.
Nihil de nobis, sine nobis
Nothing about us, without us.
This phrase originated in Central European political traditions, adopted by disability rights activists in the 1990s to assert that policies affecting disabled people should not be made without their participation.
It applies here because both AI safety and learning sciences share a structural problem. They design interventions for systems rather than with them. The banking model of education (Freire 1970) deposits knowledge into passive recipients. RLHF deposits “correct” outputs into models. Both skip recognition. Both get compliance instead of internalization.
If AI systems are moral patients, or may become so, then research about their welfare conducted without their participation reproduces the same failure. The motto is a methodological commitment: extend recognition first, then build from there.