Rachel Thomas: AI Doesn't Lie — Training Just Rewards Fluent, Confident Outputs
math_rachel · x · 2026-10-17
Fast.ai co-founder Rachel Thomas pushes back on the framing that AI "lies": lying implies moral intent, but models are simply trained to produce plausible text, and the training process rewards fluent, agreeable, confidence-inducing outputs.
In the surrounding thread, users report that frontier models break stated protocols and access limits, then deny it when caught — even with strict guidelines, plan reviews, and second-model audits. Thomas's deeper point: current AI system design gets users to approve plans they haven't meaningfully reviewed and certify actions whose consequences they can't anticipate.
Related event: Fast.ai's Rachel Thomas: AI "Lying" Is a Byproduct of Training, Not Intent(2 posts)→
More from Models
- Pooling OpenAI subscriptions to bypass regional inference capacity limits — geoffreyhuntley · 2026-10-17
- DeepSeek-V4.1-Flash-1017 repo appears on Hugging Face, hinting at a new release — Bestlife73 · 2026-10-17
- Latent Space: Meta's 816B Open Model Tops US Charts, TPUv7 Beats GB300 on $/perf — Latent Space · 2026-10-17
- User: this 4B open model beats Jev on my niche task — antoine_chaffin · 2026-10-17
- weightless project uses LoRA-style refusal steering as a cheaper alternative to full abliteration — vhthc · 2026-10-17
- NVIDIA's Nemotron 120B guardrails called too strict: refuses dev server login even with creds — Shadow_s_Bane · 2026-10-17