---
id: "concept-looks-right-but-isnt"
type: "concept"
source_timestamps: ["§ Step 3. Analyze the differences between your initial view (from Step 1) and AI's output.", "§ Making Judgment Teachable"]
tags: ["error-detection", "hallucination-mitigation"]
related: ["framework-difference-analysis", "prereq-domain-knowledge"]
definition: "Plausible, well-structured AI outputs that contain subtle, context-specific errors which can only be detected by a human possessing non-public domain knowledge."
sources: ["reskilling"]
sourceVaultSlug: "hbr-seg-reskilling"
originDay: 10
articleStem: "hbr-edu-32-help-employees-get-better-with-ai"
sourceUrl: "https://hbr.org/2026/06/help-employees-get-better-not-just-faster-with-ai"
sourceTitle: "Help Employees Get Better—Not Just Faster—with AI"
---
# Looks Right But Isn't

**'Looks right but isn't'** names a specific, dangerous category of AI error: outputs that are plausible, well-structured, and grammatically perfect, yet *fundamentally wrong in subtle ways* that require deep domain knowledge or non-public context to catch.

Two canonical examples from the article: an **ROI estimate built on generic industry benchmarks** rather than the firm's specific proprietary data; and a **logically sequenced meeting agenda that violates a particular organization's idiosyncratic workflow**. Catching these errors is the *primary value-add* of the human in the AI loop.

Detection is the third bucket of the [[framework-difference-analysis|difference analysis]] (Step 3) and requires [[prereq-domain-knowledge|underlying domain knowledge]] as a hard prerequisite. It is the concrete face of the [[claim-ai-lacks-context|zero-context limitation]]. The enrichment overlay confirms the underlying concern — AI can produce polished outputs that are wrong or detached from ground truth — while noting the exact label is the authors' own framing.


## Related across articles
- [[concept-workslop-d50]]
- [[concept-workslop-d49]]
- [[claim-uncritical-ai-use-harms-novices]]