Human "hallucinations" mirror LLM failures — same fixes apply to both¶
Insight: Human workers "hallucinate" in the same ways LLMs do — confabulating data from wrong sources, using deprecated information, and misinterpreting ambiguous terminology. A manager's fixes for human hallucination directly inform better AI agent design: surface canonical data sources with "Source-of-Truth" badges, auto-label deprecated assets, define domain jargon in shared glossaries, and create feedback loops where mistakes are caught structurally rather than by individual vigilance.
Detail: Shrivu presents three case studies of human "hallucination": (1) Bob the analyst builds a forecast on the wrong dataset because deprecated tables look as authoritative as curated ones and "Activation" means different things to different teams — fix: surface canonical sources, constrain jargon via glossaries. (2) Charlie the engineer writes code against outdated API docs because documentation doesn't reflect current state — fix: auto-detect drift between docs and code, create executable verification. (3) Dave the recruiter uses a template with wrong company details because context was copy-pasted — fix: inject context checks, require confirmation of key facts. The parallel to AI systems is explicit: both humans and LLMs fail when given ambiguous context, outdated information, or mismatched terminology. The solutions — better context management, verification loops, canonical source labeling — are the same strategies that make CLAUDE.md files and MCP integrations effective.
Sources
Related: existing entry "AI coding quality is a skill issue" in external/claude-code.md — COMPLEMENTS