For months, OpenAI didn’t know why its model was summoning goblins. This is not a bug report. This is an audit problem that affects anyone who believes that regulation can be based on self-reporting by the labs.
May 8, 2026
4 min reading time
OpenAI explained on April 29, 2026 why GPT-5.1 to 5.5 summoned goblins. The root cause lay in a reward signal for a personality feature that propagated unnoticed throughout the entire RLHF and SFT feedback loop, only becoming apparent after months and with the help of internal analyzes and external community reports. Anyone who wants to understand the technical mechanics can find it in This week’s reward hacking article. The real question of the AI ruling is different: What does it mean for external control and regulation if labs remain in the dark for months about their own model?
OpenAI’s honest answer is rarer than it seems. Labs rarely publicly admit that they don’t understand why their model behaves this way. That doesn’t make the case any more harmless. It makes it more structurally relevant.
📬 Newsletters
Do you like this article?
Every Friday: The AI verdict — exclusively for members. Subscribe for free.
The EU is postponing the high-risk AI deadlines by 16 months — opening a loophole that companies can specifically exploit. A judgment on regulatory self-dismantling.
May 1, 2026
5 mins
Unit 42 shows: LLM guardrails are not security borders. And AI in malware is real — but often more theater than threat. The verdict.
Apr 24, 2026
3 mins
Germany is sending AI agents into authorities – 18 pilot projects, a federal ministry, and a foundation that is not yet finished. What this means for the entire DACH region.
Apr 17, 2026
3 mins
AI analysis, not AI hype.
Every Friday: The AI verdict — exclusively for members. Subscribe for free.
