It's Hard to Eval Is a Product Smell

9 hours ago

The statement 'our product is hard to eval' is a product smell, indicating potential usability issues for both developers and users.
Designing products for ease of verification should prioritize user trust and reduce the need for users to redo work to verify AI outputs.
Example 1: AI data agents should provide checkable artifacts, like interactive notebooks with progressive disclosure, to allow verification through trusted sources, metric definitions, and sanity checks.
Example小孩2: An AI tool for physical education lesson plans should anchor generated plans in vetted, existing plans, showing diffs and reasons for changes to reduce cognitive load during review.
Example 3: A workers' compensation medical report tool should act as a research assistant, surfacing evidence, contradictions, and gaps for doctors to verify before assembling final reports.
General principles include understanding user verification needs, using provenance and progressive disclosure, and breaking workflows into smaller, verifiable units.
Designing for verification aligns with established principles like needfinding and sensemaking, and is crucial in the AI era where verification is a bottleneck.

Hasty Briefsbeta