Johannes Heidecke
Johannes Heidecke is a researcher formerly serving as Head of Safety Systems at OpenAI, where he oversaw the alignment, safety, and evaluation of large language models. He co-authored the GPT-4 Technical Report and contributed to system cards for frontier models including GPT-4o, OpenAI o1, and GPT-5. His research includes the Instruction Hierarchy framework for training models to prioritize privileged instructions, Deliberative Alignment for reasoning-based safety, and evaluation benchmarks such as HealthBench and PaperBench. He departed OpenAI in 2026 amid a reorganization that integrated safety and research teams under a single leader.
“Head of safety systems Johannes Heidecke... have all recently departed.”
Source→“The author bolds two specific names: Johannes Heidecke (head of safety) and Chloé Bakalar (head of ethics). Listing them separately and in bold amid a broader wave of departures implies these are not routine liquidity-driven exits — they carry qualitatively different implications for OpenAI's internal culture as it approaches IPO scale.”
Source→AI-extracted from podcast / newsletter / paper summaries. May contain errors.