# Formal Verification Becomes AI's Trust Layer

As model-generated proofs and code grow past what humans can review, machine checking in Lean and similar systems becomes the accepted way to verify AI output in mathematics and security-critical software, and labs will increasingly ship formal certificates instead of benchmark claims.

- Conviction: 40 / 100 (forming)
- Horizon: Emerging (watchlist)
- Tracking since: 2026-11-22T00:00:00.000Z
- Last updated: 2026-11-22T06:36:09.478Z
- Canonical: https://polylog.news/ai/trends/machine-checked-proof-as-ai-trust-layer
- Publisher: Polylog

## Recent evidence

- [confirms] Claude agents solved 2 of 22 open black-hole physics problems in six weeks, Anthropic says (2026-11-22): Anthropic reports that only one of its two black-hole physics results has been checked in Lean. Verification is still partial, but the lab is presenting formal checking as the standard its AI science results should meet.
- [confirms] Lean and Mathlib maintainers say a Claude-assisted proof of the Riemann Hypothesis passes machine checking (2026-11-22): Lean and Mathlib maintainers say a 17-million-line, Claude-assisted Lean library from Anthropic mathematician Levent Alpöge compiles against Mathlib's own statement of the Riemann Hypothesis, and no human-readable paper exists. The claim now rests entirely on machine checking, which makes the Lean certificate the main way to trust the result rather than a supplement to human review.
