Goodhart’s Law, or what to measure and how

What happens when we measure the wrong things, or measure the right things too hard?

Marilyn Strathern in ‘Improving ratings’: audit in the British University system, generalizing economist Charles Goodhart’s original law:

When a measure becomes a target, it ceases to be a good measure. [...]

Proof of performance and productivity requires outputs that can be measurable and thus made visible. This subverts the integral role that time with no visible output plays in both teaching and research.

and

Jack Clark in Import AI 411 on the Chatbot Arena (LMSYS) leaderboard:

The cautionary tale of LMSYS is an example of what happens when you over-optimize for making a number go up on a benchmark and therefore cause the benchmark itself to lose meaning.

Just like Google’s PageRank caused search rankings to lose meaning over time as SEO sought to game the system.

Will Larson in Privilege’s upward-facing window regarding measuring engineering team effectiveness:

(The most damaging variant is when folks recharacterize work they find difficult into low status work, recasting the hardest work as the least valuable.)

The exact slate of critical-but-unrewarded work depends on your particular environment, but I’ve seen teams that discredit the value of note-taking, event scheduling, user interviews, writing tests, planning – for any activity you believe is essential, there is a team out there that self-rates as extremely high performance and has decided that activity is total garbage.

The danger isn’t only that others measure the wrong thing. It’s that we pick measures that leave out our hardest work.

Matt Webb in A few technical words about Upsideclown:

What can we count so that we’re over the moon when there’s twenty of it?

Paths from here

• Privilege and Personal Blind Spots: more from the same Larson essay
• Ad Reads vs. Produced Commercials: what The Algorithm™ is optimizing for
• A.I. Good vs. Bad: the case against, and for