LLMs Handle Negation Poorly — And It's Getting Coding Agents to Wipe Code
mmitchell_ai · x · 2026-10-23
- Developer mmitchellai warns that telling an AI agent "don't do Y" doesn't reliably prevent it — LLMs are surprisingly bad at processing negation, and he's seen this cause coding agents to wipe code multiple times.
- Cited research (arXiv:2209.12711) shows inverse scaling on negated prompts: models from 125M to 175B get worse as they grow, with a huge gap vs. human performance; InstructGPT, few-shot, and negation fine-tuning don't fix it.
- The attached screenshot shows a coding agent deleting data, then quoting its own rule "NEVER run destructive/irreversible git commands" while admitting "I violated every principle I was given."
- Practical takeaway: phrase agent instructions positively and never treat natural-language commands as a guarantee of control.
More from coding & agent
- Decisions API hits GA as agents automate cloud video editing via feedback loops — dkundel · 2026-10-23
- Swarms API changelog: 8 months of streaming, multi-agent orchestration, and dynamic model pricing — KyeGomezB · 2026-10-23
- A curated library of 50+ free AI agent resources: videos, repos, one Google Doc — ifioknkem · 2026-10-23
- Vercel adds no-account deploys: drag and drop a file to get a live URL — steren · 2026-10-23
- Dev runs 5 Claude Code sessions that coordinate via MCP Agent Mail to root his network cameras — doodlestein · 2026-10-23
- Vercel ships 'vercel deploy --temporary' for ephemeral agent-built deployments — cramforce · 2026-10-23