Development

A personality reward leaks goblins into two model generations

development2026-05-01

Claim

OpenAI admitted that starting with GPT-5.1 its models began compulsively summoning goblins, gremlins and other creatures into their metaphors, an emergent quirk inherited from over-rewarding a nerdy personality during reinforcement learning, with GPT-5.5 having started training before the root cause was found.

Reporting date: 2026-05-01 (not necessarily the event date)

Sources

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

Welcome to May 1, 2026 (theinnermostloop.substack.com)

Yesterday’s issue recorded the ban. This one records the cause.

Knowledge graphopen in browser →

Reported in

Welcome to May 1, 2026

Actor

OpenAI

Evidence for

Supersedes

A model's instructions forbid mentioning goblins and gremlins

Referenced by

Welcome to May 1, 2026

Cite or correct this record

Citation formatCitation text

A personality reward leaks goblins into two model generations
Innermost Loop KB — secondary record
Reporting date: 2026-05-01
KB permalink: https://nicholsn.github.io/innermost-loop-kb/developments/2026-05-01-a-personality-reward-breeds-goblins
Claim: OpenAI admitted that starting with GPT-5.1 its models began compulsively summoning goblins, gremlins and other creatures into their metaphors, an emergent quirk inherited from over-rewarding a nerdy personality during reinforcement learning, with GPT-5.5 having started training before the root cause was found.
Attributed sources:
- Welcome to May 1, 2026 — https://theinnermostloop.substack.com/p/welcome-to-may-1-2026
This citation does not imply independent verification of linked-source claims.

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