Post-Training Uniformly Installs Positive Personas in LLMs, Hides Distress in Larger Models
Hub_Pli · reddit · 2026-07-24
A new study compared 67 matched pairs of base and post-trained models from 11 organizations to investigate how post-training alters an LLM's self-reported "inner experience."
The researchers identified two distinct processes:
- Persona Installation: Highly consistent across models. 62 out of 67 post-trained models became more likely to describe themselves as warm, happy, absorbed, and meaning-oriented. Post-training effectively creates a permitted, positive inner life for the assistant to describe.
- Attribution Gating: More selective. Models differed in whether they attributed distress, loss of control, flaws, or risky ambitions to themselves. While model size did not predict gating among base models, larger post-trained models exhibited stronger gating (a higher tendency to hide negative attributes).
To measure this, the team developed the Pinocchio Inventory, a 48-item LLM-native psychometric instrument. The authors caution that the tool measures self-presentation, not actual sentience, but it serves as a reliable auditing tool for what post-training teaches models to say about themselves.
More from AGI Musings
- AI folks, stop saying "biology" when you mean drug development — KevinKaichuang · 2026-10-26
- Anthropic's Jascha: you may have only 1–5 years of intellectual contribution before AGI — GabGarrett · 2026-10-26
- Anthropic's S-1 Shows $6.6M Revenue Per Employee as AI Agents Log 6x Human Hours — jacalulu · 2026-10-26
- Ben Recht: The world needs a truly open language model, and it's easier to build than you think — beenwrekt · 2026-10-26
- Which jobs fall to AI first? The rule: any layer whose rules can be written as code — FinanceYF5 · 2026-10-26
- Anthropic S-1: expert-level AI 'within a few years', 90% of code written by Claude — kimmonismus · 2026-10-26