THE BABBAGE PARADOX IN TESTING

In 1834, Charles Babbage complained to Tennyson about his poetry: “In your otherwise beautiful poem, you wrote ‘Every moment dies a man, Every moment one is born.’ This is mathematically incorrect. It should be ‘Every moment dies a man, Every moment 1 1/16 is born.'” Tennyson kept his original line.

This captures our current Gen AI testing dilemma perfectly. Babbage needed explicit precision (1.0625 births per death). Tennyson needed implicit truth (the rhythm of life and death). Both operated in incompatible modes of understanding.

Modern test generation shows this same divide. Gen AI returns statistically comprehensive coverage: 200 variations of amount validation, boundary checks, format verification. Missing entirely: that specific customer who always uploads CSVs with BOM markers, the race condition that only surfaces under 500ms latency, the legacy field everyone knows to leave empty, but nobody documented why.

Human testers compress experience like Tennyson compressed truth. Sparse data, massive insight. But making Gen AI useful requires translating this compression into Babbage-like precision. We decompose our intuition into atomic operations, define negative boundaries, spell out what we unconsciously know. The J-curve hits hard: productivity drops as we externalise decades of implicit knowledge into explicit instructions.

We’re building a Difference Engine that requires us to mechanise our expertise. Every unstated assumption becomes a rule. Every intuition needs documentation. Every “you know how it gets weird on Mondays” becomes a precise specification. The poetic becomes procedural.
The deeper challenge: how do juniors develop senior intuition when AI handles the grunt work? Previously, juniors learned by seeing patterns across hundreds of manual tests, slowly building that compression ability. Now they must learn differently—by understanding why the AI missed that BOM marker issue, by questioning what the generated tests assume, by recognising when sparse signals matter more than comprehensive coverage.

They need to learn both languages: Babbage’s precision for instructing machines, Tennyson’s compression for knowing what actually breaks systems. And perhaps a third: the meta-language of critique—how to interrogate what Gen AI produces, not just consume it. Because in the end, the future of testing may depend less on what Gen AI can generate, and more on what humans can still discern.