3 ms·
> I don’t think you can introduce a recommendation algorithm without it having a negative effect on the content it’s supposed to aid discovery of. This feels l
by dgritsko 5y ago
> I don’t think you can introduce a recommendation algorithm without it having a negative effect on the content it’s supposed to aid discovery of.
This feels like it ought to be a corollary to Goodhart's Law ("When a measure becomes a target, it ceases to be a good measure"), or perhaps a specialized application thereof.
https://en.wikipedia.org/wiki/Goodhart%27s_law https://en.wikipedia.org/wiki/Goodhart%27s_law
- tunesmith 5y agoIt's just not always true though, like in the case of "proper scoring rules". So it would seem at least theoretically possible to create a recommendation algorithm such that trying to "gamify it" would only increase the utility of its output.
- avivo 5y agoThis is sorta true. But it turns out there that there are always targets, even if they aren't explicit (is there a name for this law?). So whether you have a recommendation system or a chronological feed or whatever other way you want to display info—there is always that implicit Goodhartish: measure → target → behavior change. You can't escape it. The only question is how you want to harness it—how you can bring out the best in people (with their consent ideally!) and mitigate the negative impacts of Goodharting. (Shameless plug, I'm working on this, e.g. https://techpolicy.press/can-algorithmic-recommendation-systems-be-good-for-democracy/ https://techpolicy.press/can-algorithmic-recommendation-syst... )
- daniel-cussen 5y agoThere are exceptions to Goodhart's law, which are figures of merit. The size of transistors was a good target for a very long time, like 60 years. The height of basketball players. Muscle mass, profile, or volume. The profit margins of a company. Scores on a standardized math test.