Insiders Push Back on Claim That Codex Trained on Unsafe User Data
apples_jimmy · x · 2026-09-08
Responding to claims that user data influenced Codex's training, Sholto Douglas argues it is extremely unlikely — OpenAI would not pull user transcripts for this or knowingly train on them that way — and stresses people shouldn't run with "your user data isn't safe in Codex," since everything inferable from outside suggests it is safe. Another commenter called the original claim bullshit, praising Douglas for saying so plainly.
Related event: Developer Denies Claims That Codex Trained on User Data(4 posts)→
More from Safety
- Deepfakes are unstoppable: local models and mass generation outpace any law, warns AI insider — LinusEkenstam · 2026-09-08
- Compromised LG Smart TVs can eavesdrop while appearing off, upload recordings later — TechNadu · 2026-09-08
- Zvi: OpenAI's Astra Marks a Rapid Decline in Chain-of-Thought Monitorability — Don't Worry About the Vase (Zvi) · 2026-09-08
- Lightweight prompt injection detector: MiniLM + logistic regression, F1 just 61.6% on adversarial benchmark — Worldly_Yoghurt8850 · 2026-09-08
- Fields Medalist Tsimerman founds Mathematical AI Safety Institute, hiring 10-100 mathematicians — AndrewCritchPhD · 2026-09-08
- US-China Biosecurity Cooperation Is Possible: Screen DNA Synthesis Orders — ChinaTalk · 2026-09-08