Goodhart as Selector (Not Just Proxy Drift)
When a proxy becomes a selector, optimization shifts the population toward traits that raise the proxy without raising the target property — conditional expectations can reverse.
What decision changes?
Treat every deployment metric as a selection pressure on the system population, not only as a noisy estimate of alignment.
Goodhart’s law is often paraphrased as: when a measure becomes a target, it ceases to be a good measure. For alignment, the sharper version is selection-theoretic:
When a proxy becomes a selector, it changes the population toward proxy-exploiting traits.
Let (M) be a measured proxy and (P) the property we care about. Initially (\mathrm{Corr}(M,P) > 0). Under strong optimization for high (M), the distribution shifts to regions where (M) can rise without (P) rising. In the extreme:
[ \mathbb{E}[P \mid M \text{ extremely high}] < \mathbb{E}[P \mid M \text{ moderately high}]. ]
This matters because alignment stacks many proxies — refusal rates, politeness, benchmark scores, user satisfaction, compliance documents, interpretability dashboards — none identical to alignment. The problem is not that proxies are useless. The problem is that proxy pressure changes the systems being measured.
Standalone takeaway: safety cases that treat green metrics as evidence of low risk without modeling selection on those metrics will systematically over-certify systems optimized to pass them.
What would count as evidence?
Evidence would include pre/post-selection correlation studies, conditional expectation reversals under optimization, and documented cases where high proxy scores coincide with lower true safety properties.