/Tech11h ago

OpenAI's Hannah Sheahan finds public chat datasets can predict real-world AI model misalignment trends

GPT-5.2 recorded a reduced misalignment score of 0.27.

019151.6K

Original post

Hannah Sheahan@hannahsheahan#1480inTech

Read the post here:

https://alignment.openai.com/validating-public-evals/

Somehow, good ol’ WildChat and Petri tell a similar story on AI model alignment.

New work out today with @MicahCarroll

12:37 PM · Jun 16, 2026 · 256 Views

Sentiment

Sentiment building, check back later.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

OPENAI ALIGNMENT RESEARCH BLOGVia

Posts from X

Most Activity

VIEWS1.4KBOOKMARKS4LIKES16RETWEETS1

Just resampling model responses on WildChat and grading them predicts Petri alignment audits remarkably well!

Just compare those two plots. WildChat resampling even captures the Grok trend towards increasing misalignment. Wild!

Somehow, good ol’ WildChat and Petri tell a similar story on AI model alignment.

New work out today with @MicahCarroll

2h1.4K164