AI Productivity Paradox

Responsible AI use costs verification time, but perceived productivity rewards skipping scrutiny — design harness gates that keep review mandatory.

Bach & Bolton’s productivity paradox: responsible AI use takes considerable time and effort to verify, yet “productive” AI must save time — so organizations are rewarded for using AI irresponsibly. Management pressure to believe the tool works shrinks tolerance for nuanced, hard-to-reproduce failure reports. Refuting fluent wrong output is often an order of magnitude harder than producing it.

For agent harness design: do not optimize only for fewer human minutes per task. Budget mandatory review paths (human approval gates, genai use safety modes) and automated offline prompt evaluation so “faster” cannot silently delete oracles. Prefer me-first prompt collaboration when humans steer agents, to resist anchoring on the first fluent draft.

---

Sources

• Taking Testing Seriously (Bach & Bolton), ch. 7