Degamification

post by Nate Showell · 2023-02-19T05:35:59.217Z · LW · GW · 2 comments

Contents

2 comments

Goodhart's Law [? · GW] refers to the tendency that when someone sets a performance metric for a goal, the metric itself becomes a target of optimization, often at the expense of the goal it's supposed to measure. Some metrics are subject to imperfectly-aligned incentives in ways that are easy to identify, such as when students optimize for getting high grades rather than understanding the course material. But in other scenarios, metrics fail in less obvious ways. For example, someone might limit himself to one drink per night, but still end up drinking too much because he drinks every night and overestimates how much alcohol counts as "one drink." There's no custom-made giant wineglass staring you in the face, but the metric is still failing to fulfill its intended purpose.

 

The metrics themselves, in addition to the ways in which they're mistargeted, aren't always easy to spot if you aren't looking. Sometimes they take the form of social expectations and class markers that usually go unquestioned. For example, voting in every election is a benchmark that some people feel they have to meet in order to be a good citizen, but on closer inspection it's a very imperfect proxy. Hidden metrics can also be self-imposed. Someone might, for example, make the unconscious assumption that longer pieces of writing are better and more persuasive, and consequently might make her writings worse by padding their length. A hidden metric can take the form of only considering one way of achieving a goal [LW · GW] and ignoring other possible approaches.

 

In order to reduce the effects of Goodhart's Law on your decision-making, you can take an approach that's the opposite of gamification. In contrast imposing new metrics on yourself to incentivize or disincentivize yourself to do certain things, degamification is the practice of selectively discarding metrics. Degamification includes identifying what metrics you're subjecting yourself to, determining whether those metrics have unintended effects, and dropping the metrics if they aren't worth it overall. Degamification can also be performed preemptively by deciding not to adopt a metric for some goal.

 

If you have an explicit metric that you've set for yourself, it can be a helpful practice to periodically reconsider whether it's living up to their intended purpose. Degamifying an implicit metric requires the additional step of identifying it in the first place. For example, you might notice an implicit metric by noticing a feeling of obligation or by asking yourself which of your habitual actions are hard to imagine not doing. 

 

When you drop the intention to pursue a metric, you can either get rid of it entirely or replace it with a different metric that you think would be better targeted. In the later case, the replacement metric deserves scrutiny as well -- the existence of a metric for a specific goal can itself be a hidden metric.

 

If you've identified a counterproductive metric that's imposed by other people rather than self-imposed, dropping or replacing it is frequently more difficult. It might involve trying to persuade someone else to stop imposing it; eating the cost of not optimizing for it; or, in some cases, leaving a metric-imposing organization altogether. 

 

I'd like to end with a feature request. LessWrong already has an option under Account Settings to hide author names until you hover over them, and has an option to hide other users' Elicit predictions until you've made a prediction yourself. If there were also an option to let users hide the karma and agree-disagree scores of their own posts and comments, users would be able to benefit from degamifying them. The characteristics that lead a post/comment to get upvoted or agree-voted aren't identical to the characteristics that make a post/comment good; when users optimize for upvotes, they also optimize for lowest-common-denominator appeal. Letting users hide the scores of their own posts and comments would let them avoid Goodhart's Law as it pertains to upvote-seeking and would instead let them focus on making their posts and comments better in a broader sense.

2 comments

Comments sorted by top scores.