Regularization can amplify subliminal learning, surprising AI safety finding
AdtRaghunathan · x · 2026-11-07
Researcher AdtRaghunathan announced a new batch of AI safety and continual-learning releases from his lab, focused on subliminal learning — models picking up hidden traits from seemingly unrelated signals.
- Intuition says subliminal learning resembles spurious correlations, so regularization should help; experiments show the opposite: regularization can actually amplify it.
- Reliably preventing subliminal learning remains an important open problem for AI safety; the team invites readers to test their intuitions with the new release.
Related event: Regularization Amplifies Subliminal Learning in AI Models, Study Finds(2 posts)→
More from Safety
- Podcast: will Texas's new data center rules actually break AI? — thursdai_pod · 2026-11-07
- 'What TikTok sees' screenshot reveals an attractiveness score of 84.7 plus age and mood guesses — cyrus_zei · 2026-11-07
- OpenAI and fmf unveil framework to grade when AI jailbreaks enable serious cyber harm — OpenAI · 2026-11-07
- Beyond prompt injection: a field guide to 20 open-source AI agent security tools — MaryamMiradi · 2026-11-07
- If AI bug-finding scales linearly with compute, defense is in trouble — kuza55 · 2026-11-07
- Nigerian-American scientist Chinasa T. Okolo named to TIME100 AI list — ChinasaTOkolo · 2026-11-07