Reliable AI Design: Say "I'm Not Sure" When Evidence Is Unclear
goyalshaliniuk · x · 2026-09-25
The author lays out a key principle for reliable AI systems: they shouldn't always try to pick a winner, but instead follow a flow of Detect → Compare → Verify → Resolve → Communicate uncertainty.
The highlight is step 5, "surface the conflict": when there isn't enough evidence to determine which source is correct, don't hide the disagreement. Tell the user the sources disagree, then show:
- What each source says
- Where they differ
- Which evidence is available
- What remains uncertain
This makes the AI transparent instead of overconfident — when evidence is unclear, "I'm not sure" can be more useful than a confidently wrong answer.
Related event: Handling Conflicting Sources in AI: Rank, Verify, and Admit Uncertainty(2 posts)→
More from coding & agent
- Cloudflare open-sources Agentic Inbox: a self-hosted AI email client on Workers, 8k stars — tom_doerr · 2026-09-25
- Free MCP Community Connect conference heads to Bengaluru with focus on production-grade MCP — lee_stott · 2026-09-25
- Building an AI platform: model API calls are just 2% of the system, says Adobe's Suhas Suresha — Al_Grigor · 2026-09-25
- Opus 5.5 can now drive Blender to make claymations from a single prompt, and the skill is open-sourced — sidahuj · 2026-09-25
- First working n8n WhatsApp AI receptionist for a dental clinic, with two snags — vxdant23 · 2026-09-25
- Every PR Opus 5.5 opens passes review with 'no notes' from agent Astra — banteg · 2026-09-25