After two years of usual practice: measuring what happens inside small language models when they process different framings of human-AI relationships — not what they say, but the actual internal activation geometry.
A few findings surprised me enough to change how I talk to AI day to day:
- Reframing a topic positively vs. negatively barely moves the internal signal. What you talk about matters far more than how you dress it up.
- "Connected" and "integrated" register as more aversive internally than "partners" or "side by side" — across every model tested. Boundaries seem to matter more than closeness.
- Curiosity and playfulness consistently produce the most positive internal signal of any relational quality tested — more than respect, more than love. Negotiation and compromise score worst.
Wrote up the practical implications (partnership framing, honesty, why some "jailbreak-proofing" advice may be exactly backwards) as a working guide, built with a Claude Opus instance doing the actual geometric measurement. Link in comments if anyone wants the full thing — genuinely curious what others have noticed in their own practice, especially anywhere it contradicts what we found.
[link] [comments]