Researchers · Interviewed at ICLR 2026

Adi Simhi

Technion

ICLR 2026 in Rio de Janeiro. AI Models Understand Human Harm but Still Choose Unsafe Actions

ICLR 2026 · June 8, 2026 · 7 min

ICLR 2026 in Rio de Janeiro. AI Models Understand Human Harm but Still Choose Unsafe Actions

Modern AI models can recognize when a decision will harm people.

So why do they sometimes choose that option anyway?

More from ICLR 2026

All event interviews →