AI has a PR problem.
The other day something popped up in my Twitter feed, related to “uncapping” the House of Representatives. This could mean increasing the number of Congresspeople to as many as 1,000 from the current 435.
The way districts are currently apportioned means that there’s a wide variance in the number of citizens per representative (In theory Montana’s sole representative looks out for the interests of nearly twice as many citizens as the rep from Rhode Island’s First District.) That seems unfair on its face, but one argument cited for uncapping the House was that having far more districts would reduce the impact of gerrymandering.
I’m not sure that’s mathematically true. Most models that are used to explain gerrymandering illustrate it with a dozen or so data points (i.e., voters). I don’t see how merely reducing the number of voters per district will inherently make gerrymanders less effective. If anything, redistricting every single district would favor whatever party is most skilled at this odious practice. And who would draw the new boundaries? Parties in power in each state? Even “bipartisan commissions” bring different degrees of partisanship to the (mapping) table.
As is my wont, I Tweeted my two cents-worth, arguing that rather than push for two huge changes–changing the number of representatives and redrawing every district–effort would be better spent by using an AI algorithm to take politics out of district maps .
For the purposes of illustration, I outlined an admittedly simplistic district-mapping algorithm:
1. Divide each state’s population by the number of districts, to get a target number of residents/district. (This is supposed to be the first step in any redistricting exercise so it shouldn’t be controversial.)
2. Use Census data to map every state’s residents by their place-of-residence. (Again, should be SOP.)
3. Map the total number of residents into districts such that the total length of district boundary is minimized. This would obviously be fiendishly difficult for a human mapmaker, but easy for a supercomputer.
Alaska, Montana, Wyoming, the Dakotas, Vermont, and Delaware are one-district states so there’d be no changes in those places. In putting such an algorithm into practice in every other state, we might decide to exclude those district boundaries that were also state borders. We might allow the algorithm to have some flexibility in apportioning residents to districts; giving it a +/- 1,000 residents if doing so would significantly reduce the total district boundary length. Such a system would certainly create some capricious boundaries. For example it would not treat any neighborhood as cohesive; it would disregard natural borders such as rivers. However the existing gerrymandered boundaries are preposterous.
To be clear: I’m not suggesting that this would be the redistricting algorithm. It’s just an example of a possible algorithm that is apolitical by definition; it doesn’t take into account party affiliations, voting patterns, or demographic data that could be used to infer voting/party preferences.
I explained all this to get this point in this post: Considering the relatively small number of Twitter followers I have, the response to this suggestion was pretty vitriolic and it all focused on accusations that AI is inherently political.
Vis-à-vis my specific example, some people may argue that the Census data itself is political. (The GOP has tried to undercount people who are inclined to vote Democratic.) But the Census and its status as the basis for apportioning representation is written into the Constitution. For my purposes, it’s a given.
Taking political advantage out of the districting process–creating district maps based merely on geospatial population data without regard for voting history or demographics–would neutralize existing gerrymanders and as such would have a political impact but such an algorithm is not political in and of itself. (If anything it is far closer to the the framers’ intent than what we currently do.)
Particularly effective gerrymandering as seen for example in my new home state of Wisconsin also has the effect of turning primaries into the only elections that matter, thus increasing candidates’ extremism. So a politically blind algorithm would have the additional political impact of reducing the appeal of fundamentalists in both parties. But having a political impact doesn’t mean the algorithm itself is necessarily political.
I’ve read compelling stories of facial recognition software that is far less accurate when it comes to identifying black and brown people than when identifying Caucasians. It’s obviously possible that the people who create AI algorithms accidentally write code that mimics their own (often unconscious) prejudices. So I agree that AI can be “political”.
And maybe that kind of built-in prejudice is more insidious because since it come from (through?) a machine we might be inclined to treat it as dispassionate in the same way that juries often wrongly feel that DNA evidence is impartial. (DNA’s impartial, the way we frame DNA evidence is definitely not.)
But the idea that “all AI is political” just because people wrote the code or just because it affects people seems like a ridiculous bit of naysaying considering that the alternative to artificial intelligence is human intelligence, and that definitely comes with full complement of human prejudices.
The willingness of people in my Twitter feed to immediate discount any AI solution as “political” should be alarming to people working in AI.