We had our AMA session with Dutch attorney Anton Ekker.
Here is one exchange from it:
That depends on what we mean by “a good prediction”, @antonekker. If we are happy with being “good” (outperforming randomness) at the aggregate level, we might need very little data. For example, in predicting the outcome of football matches, the simplest model “the home team always wins” does (a little) better than random. Hal Varian (Google’s chief economist) a few years ago went on record saying “if you have 99% correlation, who cares about causation”, or something like that. But this extra performance only applies to predicting a whole lot of football matches (the population), while being useless if you are trying to predict one match in particular.
I think @katejsim worries that prejudices outperform randomness. If you don’t care about fairness and the rights of the individual , you could indeed predict that the poorer neighbors would have more social welfare fraud than rich ones. But this would come at the expense of treating poorer individuals fairly, and, unlike with football matches, it would end up reinforcing the conditions that force those people to apply for welfare in the first place.
Interesting point.
Taking the possible consequences for citizens in account, the predictions should actually be much better than just ‘good’. If 2% procent of the outcomes are wrong, this is already effecting a large number of people.
This raises the question if decisions by government about fraud can ever be left to algorithms alone. Maybe, human interference should be mandatory.