Why it matters: As AI churns out code faster than anyone can review it, the real risk shifts from buggy code to a precisely proven wrong idea.
The Details:
Formal verification used to split work into two halves: a quick specification and a brutal proof. AI has flipped that ratio. Machines can now grind through proofs, so the specification, the exact statement of what a system must always do, is the part that still demands human judgment. A fast proof of the wrong property gives false confidence, which is worse than no proof at all.
Teams rarely know what "correct" means until they try to write it down. A rule like "orders can't be charged twice" falls apart under questions about retries, time windows, or subscription rebills. The fix is low-tech: draft the rule in plain English, then attack it at a whiteboard until every "what about" case forces a new clause. AI earns its keep here by generating the adversarial scenario humans run out of energy to invent.
Property-based testing tools, such as Hypothesis, fast-check, or QuickCheck, offer a gentle entry point. Instead of checking one input, you state a rule that must hold for every input and let the tool search for a counterexample. If writing that rule feels harder than writing a normal test, that discomfort is the specification work arriving in disguise.
Heavier tools like TLA+ and Lean only pay off once plain English stops surprising you, meaning the open questions are about mechanics, not intent. AI can then translate tightened language into a formal model and turn counterexamples back into readable sentences, since an unreadable counterexample just gets ignored.
Bottom Line: The job hasn't vanished, it moved up a level: writing the two or three sentences that state what must always be true is now the work that matters most.
Enjoy this article?
Listen to the Claude Code Conversations radio show or join the community.