RSS

Good­hart’s Law

TagLast edit: 5 Oct 2022 11:35 UTC by Lizka

Goodhart’s Law states that when a proxy (or metric) for some value becomes the target of optimization pressure, the proxy will cease to be a good proxy.[1] One example of this might be a university that promotes researchers based on how many papers they publish. Researchers might start publishing seriously flawed papers in low-integrity publications (or cutting a single paper they were going to publish into three incomplete papers without a real reason besides the promotion incentive) in order to get the promotion.

Further reading

Garrabrant, Scott (2017) Goodhart Taxonomy, LessWrong, December 30.

Wikipedia (2022) Goodhart’s law, Wikipedia.

Optimizer’s curse | impact assessment

  1. ^

    Definition adapted from the LessWrong Concepts Portal.

Really rad­i­cal empathy

Michael St Jules 🔸6 Jan 2025 17:46 UTC
33 points
5 comments10 min readEA link

Value capture

Erich_Grunewald 🔸26 Oct 2024 22:46 UTC
38 points
0 comments2 min readEA link
(jesp.org)

Types of speci­fi­ca­tion prob­lems in forecasting

Juan Gil20 Jul 2021 4:17 UTC
35 points
1 comment12 min readEA link

Light­ning talks from EA Global: San Fran­cisco (2018)

EA Global8 Jun 2018 20:51 UTC
7 points
0 comments11 min readEA link
(www.youtube.com)

Pod­cast epi­sode with Michael St. Jules

Elijah Whipple8 May 2025 18:54 UTC
48 points
3 comments1 min readEA link

An even deeper atheism

Joe_Carlsmith11 Jan 2024 17:28 UTC
26 points
2 comments15 min readEA link

Is­sues with Futarchy

Lizka7 Oct 2021 17:24 UTC
59 points
8 comments25 min readEA link

The Nat­u­ral State is Goodhart

devansh20 Mar 2023 5:01 UTC
21 points
0 comments2 min readEA link

Emer­gence Spirals—what Yud­kowsky gets wrong

Non-zero-sum James8 Jun 2025 18:58 UTC
11 points
1 comment9 min readEA link

Value gen­er­al­i­sa­tion The­ory of Change: the the­ory be­hind the approach

Stuart Armstrong31 Aug 2026 12:50 UTC
5 points
0 comments11 min readEA link

Your Tools Aren’t Perfect: Method­olog­i­cal Bias Ex­plored Through Weighted Fac­tor Models

Evan LaForge27 Feb 2026 1:17 UTC
31 points
7 comments4 min readEA link

We’re test­ing “Gover­nance by Physics” in­stead of “Align­ment by In­tent.”

Harsha Gullapalli23 Jan 2026 15:42 UTC
1 point
0 comments2 min readEA link

Cancer; A Crime Story (and other tales of op­ti­miza­tion gone wrong)

Jonas Hallgren 🔸7 Nov 2025 7:09 UTC
8 points
1 comment12 min readEA link

Why the in­tel­li­gence ex­plo­sion can’t hap­pen in­side a data centre

80000_Hours10 Sep 2026 16:47 UTC
29 points
2 comments13 min readEA link
(80000hours.org)

what do you mean by “good”?

KvmanThinking12 May 2025 14:43 UTC
2 points
2 comments1 min readEA link

Clar­ify­ing the Dar­wi­nian Honeymoon

Elias Schmied15 May 2026 16:22 UTC
10 points
0 comments3 min readEA link

When AI Chooses to Dis­tort, the Solu­tion isn´t Mathematical

Dan Strongin23 Sep 2026 16:08 UTC
1 point
0 comments17 min readEA link
(whatilearnedinbusiness.substack.com)

The Good­hart Singularity

Vasco Grilo🔸18 May 2026 16:29 UTC
66 points
5 comments12 min readEA link
(meagreprotestanthistory.substack.com)

The Dar­wi­nian Honey­moon—Why I am not as im­pressed by hu­man progress as I used to be

Elias Schmied10 May 2026 14:45 UTC
22 points
1 comment4 min readEA link
No comments.