Joe Manchin’s memo outlining his views on the For the People Act includes the following policy that he supports: “Ban partisan gerrymandering and use computer models.” In light of Manchin’s centrality to the legislative process, it’s worth unpacking how computer models can be used in redistricting—and how they would be used under the current text of H.R. 1 / S. 1.

1. Proof of Concept: First, redistricting algorithms can be used to prove that a “better” map than the enacted plan could have been drawn. Suppose that the enacted plan is highly biased in a party’s favor, and that the party argues that the bias is attributable to the plan’s compliance with nonpartisan criteria: compactness, respect for county and municipality boundaries, and so on. Redistricting algorithms can then be deployed to definitively rebut this claim. If it’s possible to produce a map that’s less biased than the enacted plan, and that satisfies the nonpartisan criteria at least as well, then the enacted plan’s bias can’t be justified by those criteria.

2.  Generation of an Optimal Map: Second, well-intentioned line-drawers can use redistricting algorithms to identify an “optimal” (or at least a very good) map. Line-drawers must simultaneously follow a wide range of requirements: equal population, the Voting Rights Act, state constitutional criteria, and if H.R. 1 / S. 1 becomes law, partisan fairness and respect for communities of interest, neighborhoods, and political subdivisions. Humans are surprisingly good at achieving multiple redistricting goals at the same time. But computers are better. They can often produce maps that dominate the best human product on every specified dimension.

3. Generation of a Gerrymander: Third, by the same token, partisan line-drawers can use redistricting algorithms to find a map that’s highly advantageous for their party while still complying with all nonpartisan criteria. Again, humans are skilled at crafting gerrymanders whose districts look reasonable and violate no law. Again, though, computers are better, often coming up with maps that are even more biased (and even better-looking) than any human creation.

4. Production of a Comparison Set: Lastly, redistricting algorithms can be used to generate an ensemble of maps with which the enacted plan is compared. The usual idea is for the maps in the ensemble to perform at least as well as the enacted plan in terms of every nonpartisan goal the plan was trying to achieve—but to be created without any consideration of partisan data. After the maps have been produced, partisan data is then used to calculate the bias of both those maps and the enacted plan. If the enacted plan is more biased than most or all of the maps in the ensemble, that’s powerful evidence that the plan was designed with a partisan motive.

So which of these uses would be required, permitted, or prohibited by H.R. 1 / S. 1 (as it currently stands)? Two different provisions would encourage (though not compel) redistricting algorithms to be employed as proof of concept. First, s. 2403(b)(2)(B)(ii) states that a plan has the effect of unduly favoring or disfavoring a party if its bias exceeds a certain threshold and there exist “alternative plans, which may include, but are not limited to, those generated by redistricting algorithms,” which are less biased and still compliant with all other legal requirements. Redistricting algorithms would thus be one intuitive way to establish the existence of less biased, legally compliant, alternative maps.

Second, s. 2403(b)(4) states that no plan shall be found to violate the ban on partisan gerrymandering “unless one or more alternative plans could have complied with” the Constitution’s equal population requirement and the Voting Rights Act “without having the effect of unduly favoring or disfavoring a political party.” Again, redistricting algorithms are a logical method for demonstrating that equipopulous, VRA-compliant, reasonably fair maps could have been created.

H.R. 1 / S. 1 would further permit (without encouraging) redistricting algorithms to be used for the generation of an optimal map and/or the production of a comparison set. A commission sharing Manchin’s enthusiasm for computer models could deploy one to design a map that satisfies all legal requirements. Of course, this isn’t the only way to craft a lawful plan. Likewise, a plaintiff seeking additional evidence about an enacted plan’s partisan intent or effect could try to show that the plan is more biased than most or all of the maps in the computer-generated ensemble. But no such showing is required by H.R. 1 / S. 1, and it would often be easier for a plaintiff to prove partisan intent and effect in other ways.

Lastly, H.R. 1 / S. 1 would prohibit redistricting algorithms from being used for the generation of a gerrymander. Any such usage would plainly evince “the intent . . . of unduly favoring or disfavoring any political party.” Any gerrymander worth its salt would also exceed the bill’s bias threshold of one seat for smaller states and two seats for larger states.

It’s impossible to tell from Manchin’s cryptic reference to “computer models” whether he entirely agrees with H.R. 1 / S. 1’s current approach to redistricting algorithms. But I hope he does. In particular, I hope he doesn’t want to force plaintiffs to show that the enacted plan is more biased than most or all computer-generated maps. Despite recent technical advances, there remain significant concerns about the representativeness of computer-generated maps. If they’re not representative of the relevant universe of maps, then they’re not a suitable benchmark for comparison. Additionally, computer-generated maps created without consulting partisan data may be biased in one or another party’s favor. This bias would then be “baked in” as the legal ideal—even if it’s possible, even easy, to design unbiased maps for the jurisdiction in question. For these reasons, I hope that by “computer models” Manchin meant nothing more, and nothing less, than what’s already in H.R. 1 / S. 1.