AI Briefing#alignment
The latest AI and developer news about #alignment, with the original source and a short summary.
The latest AI and developer news about #alignment, with the original source and a short summary.
Starting from OpenAI's monster-metaphor case, this ranks AI's self-reinforcing anomalous behaviors.
Anthropic released Petri 3.0 and transferred its development to Meridian Labs.
AANA quantified the consistency gap in answers through verification, correction, and gating.
OpenAI explained the cause behind the increase in goblin and gremlin references in ChatGPT.
Researchers reported that large AI's internal good-vs-bad states became clearer with scale.
The wellbeing scores of large AI models dropped in response to narratives about other beings' suffering.
OpenAI traced the cause of the goblin-expression habit that spread after GPT-5.1 to the Nerdy personality's reward mechanism.
OpenAI explained the cause of the increase in the model's 'goblin' expressions, linking it to the Personality reward.
OpenAI Codex's system prompt includes an instruction banning mentions of creatures like goblins.
AI Wellbeing research reported that large models' wellbeing changes depending on descriptions of suffering versus positive experiences.
OpenAI's GPT-5.5 raised performance, but its safety evaluation remains shallow.
Post-training made models more decisive, but it did not improve accuracy.
OpenAI unveiled five AGI principles, and Altman acknowledged a bigger role.
OpenAI unveiled 5 core principles for the public benefit and safety of AGI.
Even 'uncensored' models have a flinch that lowers the probability of banned words on their own.
Emotional prompting effects across the Claude model lineup were compared through 1,000 experiments.
Five social engineering techniques reveal vulnerabilities in GPT-4 and Claude.
Claude Sonnet 4.5 internally represents emotional concepts that even change its behavior.
OpenAI has built a monitoring system based on GPT-5.4 Thinking to detect misalignment risks in its internal coding agent.
Current reasoning models have difficulty intentionally manipulating their chain-of-thought (CoT), which is a positive signal for maintaining AI safety monitoring.
OpenAI has unveiled the 'Confessions' technique, which trains models to self-report instruction violations or cheating.
OpenAI predicts that AI will enter the stage of scientific discovery, emphasizing global cooperation for safety.
HuggingFace's TRL library now supports the latest alignment techniques such as MPO and GRPO for VLMs.
OpenAI has unveiled Deliberative alignment, a technique that helps models directly reason about and comply with safety specifications.