Development

Models are trained to describe themselves faithfully

development2026-06-19

Claim

MIT's Self-CTRL trains models to describe themselves faithfully, and OpenAI reported that rewarding honesty and humility produced broad alignment that held under adversarial fine-tuning.

Reporting date: 2026-06-19 (not necessarily the event date)

Sources

Attribution and reporting linked by this record; links do not imply independent verification.

Welcome to June 19, 2026 (theinnermostloop.substack.com)

Knowledge graphopen in browser →

Reported in

Welcome to June 19, 2026

Actor

Evidence for

Supersedes

Filtering bad rollouts fails because traits distill from the teacher

Referenced by

Cite or correct this record

Citation formatCitation text

Models are trained to describe themselves faithfully
Innermost Loop KB — secondary record
Reporting date: 2026-06-19
KB permalink: https://nicholsn.github.io/innermost-loop-kb/developments/2026-06-19-models-trained-to-describe-themselves-faithfully
Claim: MIT's Self-CTRL trains models to describe themselves faithfully, and OpenAI reported that rewarding honesty and humility produced broad alignment that held under adversarial fine-tuning.
Attributed sources:
- Welcome to June 19, 2026 — https://theinnermostloop.substack.com/p/welcome-to-june-19-2026
This citation does not imply independent verification of linked-source claims.

Suggest a correction — opens a GitHub draft for you to review and submit.