Expertise in Democracy
Modeling the conflict between equal-say and expertise in group decisions
We all work in groups, in our jobs, our schools, and with our friends and families. Our ancestors did the same, and though technology has changed, some rules for organizing groups have become common sense. Here are the two I want to focus on:
- “One person, one vote.” The ideal set of norms, laws, and codes for a group of people counts each person’s perspective about equally. We commonly associate this idea with democracy, and it’s present in lots of political systems.
- “Trust the experts.” Each person’s influence over a given area should be proportional to their knowledge and skill in that area. Expertise is valued all over society: teachers are expected to have degrees in their fields. Plumbers and lawyers are granted privileges because of the licenses they’ve earned.
Most people are likely to agree with both, though maybe in separate contexts — they’d say politics should be about equal say, while markets should be meritocratic, for instance. I argue that an ideal governance scheme has both, but there’s a problem: the two rules are incompatible. If experts matter more, not everyone has equal say. If everyone gets one vote, expertise is wasted.
Resolving such a fundamental dissonance is daunting, but the solutions are approachable, and I hope you’ll find my suggestions both intriguing and useful.
Why?
Since it may not be plain, I’d like to start by justifying time spent reading (and writing) an essay about resolving the disharmony: if our goal is “the greatest happiness of the greatest number,” fixing the contradiction has huge upside.¹
“One person, one vote” is an effective way to hold policy-making accountable to group satisfaction. Since each person has about the same potential for happiness, each represents a same-sized piece in the happiness jigsaw. Giving each one a vote codes that structure into legislation accurately, a feat of information we can’t do any other way.
To know if someone is content, we have no better test than asking them. There is no happiness pill that grants nirvana regardless of our personal stories. If one day such a pill exists — pop one and live satisfied forever, no matter how horrific your past — individual perspectives could become irrelevant to collective happiness. For now, “one person, one vote,” or at least some technique drawing information from each group member, is the only way.
On the other hand, most of us regularly defer to experts, and with good reason: a group that doesn’t listen to its experts struggles with complex problems and solves all its problems less efficiently.
Lots of issues don’t require unusual experience or skill to solve, and an equal-say-only organization should do fine with them. Half the city burned down? Everyone knows to rebuild it out of fire-proof materials. Many dilemmas are much more vexing, unfortunately. Manufacturing jobs are disappearing. Prisons are full, but when we release people, they commit more crimes. Wrestling these problems, the best solutions aren’t obvious, so we turn to experts on manufacturing, jobs, prisons, and crime to help.²
Because manufacturing experts are a small minority of the group, their voices need amplification to move the needle. If the manufacturing expert is just one vote in a thousand, the group might just keep banging its head against the wall — the jobs won’t come back. The Monty Hall Dilemma is a problem of this type.³
Experts are also quicker and less wasteful problem solvers. Not only will a construction crew build a house faster than a bunch of laypeople, they’ll waste less material and energy on mistakes and extraneous steps. This increased efficiency isn’t trivial. Misused company resources can endanger employees’ jobs. Squandered tax money can mean hungry people not fed or children not educated.
Amplifying expert voices is a big upgrade for group problem solving, but weighing each voice equally is key to the greatest happiness for the greatest number. Both rules are worth employing, and so an attempt at resolving their contradiction is a worthy pursuit.
A Past Solution
As is, the two statements compete for the same variable, influence per group member:
best legislative scheme = { influence per group member: equal, influence per group member: proportional to expertise} # Doesn't work!
To let them coexist, we can split the scheme itself in two. The United States Congress is a nice parallel. Its designers wanted 1) each state in the union to have equal legislative representation, and 2) each state to have representation proportional to its population. Just like in our paradox, these things couldn’t both be true, so they split the legislature in two and gave each piece one of the rules. In the Senate, each state has exactly two senators. In the House of Representatives, each state has a number of representatives proportional to its population.
U.S. legislative branch = { house of representatives: { influence per state: proportional to state population }, senate: { influence per state: equal }} # This works?
The US congress solution kind of works, but after a few centuries, we know it’s also disposed toward gridlock. The split into two houses successfully avoids the paradox, but it also traps the two houses in an eternal struggle for control. How many bills make it through one house and fail in the other? That may be working as intended, but it’s also incredibly slow, and people suffer while we dither.
Another way to look at this system’s problems it to ask: does the US legislature really solve the paradox? It’s progress, but it’s not good enough. Is influence per state equal? Despite avoiding the contradiction, no. In the Senate, states have equal influence, but in the sum of both houses, they don’t. Having equal say half the time isn’t a good substitute for having it all the time. Is influence proportional to state population? Again, it isn’t. Representation is proportional in the House but not the Senate. A half is not a whole.
Solution 2.0
Some better options are out there. Instead of two houses competing for the same role, we can create two layers responsible for distinct parts of the decision-making process. Members all have equal say in one layer, while experts hold sway over the other. I suggest this deconstruction, though others may be valid:
- Layer 1 decides which topics to discuss
- Layer 2 decides what to do on each topic
Choosing the topics of discussion, the problems that the group must solve, requires general knowledge of all possible topics. Only the entire group, or nearly all of it, has that knowledge, so it makes some sense to divide power over the first layer evenly between all group members. If the paradox is to be solved, that leaves the second layer to favor expert opinions, and luckily, it’s a fit. Designing complete solutions to defined problems is exactly what experts should be better at. Our new model looks like this:
best legislative scheme = { layer 1: { influence per group member: equal, controls: legislative direction }, layer 2: { influence per group member: proportional to expertise, controls: detailed legislative design }}detailed legislative design = function(legislative direction, context)
In this scheme, the scope of legislation is decided by “one person, one vote,” and the details are ironed out by experts in each topic. To stop the experts going rogue, some norm, code, or law requires the detailed policies they design to be derived from the general directions of the first layer, with some inevitable wiggle room.
Such a system would be a better solution. Influence per group member isn’t equal through the whole scheme, but each person does have equal say in which problems are legislated, controlling the entire system’s goals. While expert opinions don’t get special treatment through out, they’re put first when it counts, in the search for good answers to difficult-but-defined questions. It’s not perfect, but I think this second attempt clearly outdoes the first.
Modeling the Solutions
I’ve programmed a crude model of the paradox to compare the two-layered scheme with other common solutions. Here’s the set up:
Say 100 people live in a neighborhood called “Oak Grove.” They’ve got all sorts of ideas and issues with their neighborhood: one wants to repave the streets, another wants to plan a music festival, a third thinks Oak Grove Elementary should hire more teachers, and so on. Each person has their own preferred order for these ideas, but if you asked everyone for the one thing Oak Grove should change, you might find a few ideas cropping up more than the others. Oak Grove residents also have particular skills: each is expert at their profession and maybe a couple other topics. Can we implement an organizing scheme affordably address Oak Grove’s biggest issues?
We simulate the scenario by creating 100 agents, the Oak Grove residents or whatever group you like, and 100 problems, the list of issues they’d like to solve. To model the diversity of opinions and skills, the agents are given a preference value and an expertise value for each problem, representing how much they want the problem addressed and how good they’d be at solving it. Since Oak Grove residents know each other, we know everyone’s expertise values, but we don’t know their preferences: we’ll have to ask to find those out. There are two possible steps, which we can repeat as many times as we want:
- Ask an agent for their highest unknown preference and its associated problem.
- Ask an agent or agents to attempt solving a problem. Each agent votes for a correct or incorrect solution depending on their expertise. Ask a group with lots of experts in the problem, and they’ll probably solve it. Ask one with few experts and you’ll have to get lucky.
The challenge is: solve the agents’ 10 most preferred problems in the fewest steps and the fewest total problems solved. Solving the most preferred problems parallels the group goal of achieving maximum satisfaction, and doing it efficiently mirrors resource restraints. Oak Grove isn’t made of money or time, so we’ve got to keep costs modest.
I ran trials with three algorithms:
- one that discovers preferences like the first, but asks all the agents to solve the top problems, ignoring experts.
- one that favors experts, asking only experts to solve all the problems indiscriminately;
- Last, a two-layered approach based on the paradox scheme, which first attempts to discover the most preferred problems treating all agents equally, then asks only experts to solve its hypothesized top problems.
Because a minority of agents are experts at any given problem, the algorithm that ignores experts, analogous to direct democracy, takes dramatically longer than the other two, so much so that we can dismiss it. The average steps and problems solved for the other two algorithms are shown here:
With 100 agents and 100 problems, the two-layered method is the clear winner in this simulation: it takes about the same number of steps as the experts-only scheme but has to solve a fifth the number of problems, which is a significant cost reduction.
Placing It In the World and Widening the Lens
Some forms of the two-layered approach appear in the wild. All marketplaces have something like it. While a market has no explicit goal, its inferred destination is making participants as happy as possible, since consumers presumably choose what’s best for them and producers try to make what consumers will choose. The first layer can be framed as producers observing consumption habits and distilling a model of what people most prefer. Each consumer has the power to influence that model, but their power is connected to their consumption bandwidth, which is not always evenly distributed. A rich person can buy 10 times what an average one can, but a rich person can’t buy more than 24 hours in the day: monetary wealth varies greatly, but there’s a cap on the variance of our attention budgets.⁴
The second layer comes when producers create new material for consumption. If they’re aiming for popularity, a producer creates something derived from their model of top consumer preferences, just like the experts in our two layer model design specific answers to the general problems we crowdsourced. The better the producer executes this task, which is related to their expertise but also their available resources, the more consumers will reward them. Thus, marketplaces have a “behavioral” version of the scheme, built off consumer behavior and granting influence in proportion to wealth more than the original version.⁵
A second, “cognitive” incarnation lives in feedback surveys and voting schemes used by companies, schools, and governments. The first layer is a survey or ballot, granting each group member the same opportunity to contribute, then compressing that information into a set of concise community preferences. The second layer assumes that community leaders will take action on those preferences, delegating the work of fulfilling them to the most appropriate, expert community members. If the employees ask for faster payroll turnover, leaders ask accounting to work on it. If the mayor got elected on a education-reform platform, she’s expected to develop an education plan with expert help. Things often don’t go so smoothly in practice — most customer feedback surveys seem to either get ignored or lost in bureaucracy, and we’ll let the sleeping political dog lie — but the theory is there, and a working cognitive approach might be less corrupt than the behavioral examples, since being wealthy shouldn’t make your survey answers louder or your vote count twice.
The best way to execute the two-layered scheme is still up for debate. The behavioral version cares about what people do; the cognitive one listens to what they ask for. Which leads to greater happiness? Maybe I set a goal to eat healthy all year, but I only do it for a month. Would I have been better off had I the discipline to accomplish my goal, or does my behavior reveal my true preferences? That’s too big a topic for this essay, but my short answer is: probably a little from both buckets.
There might be ways to combine the behavioral and cognitive approaches. Bounties and crowdfunding are taking markets down that path. Markets usually care only about actions, not plans, but with bounties, people pay experts to create what they consciously request, not just what they consume. Crowdfunding shows creators what we prefer to fund, not just what we’ll buy. Both emphasize planned choices for the future instead of instant gratification, and maybe that design direction can add a little balance to the ways markets make decisions in the future.
Footnotes
¹ If the group’s goal isn’t some version of happiness for its members, these rules aren’t as relevant. A profit-focused company doesn’t give all its employees much of a say in its decisions, and that’s partly because their individual perspectives are hardly relevant to its goal.
² “Expert” here refers to someone with skill and experience in a given area, regardless of how they attained those things. Someone with a biology PhD. is an expert, but someone who’s been living in the woods for 40 years probably is, too.
³ Nicky Case’s The Wisdom and/or the Madness of Crowds is also a great model of how groups can either virally spread information or stubbornly refuse it.
⁴ Wealth can buy more than 24 hours in the day in an indirect sense: I can buy other people’s time, which isn’t that different from extending my own. We see lots of services selling social media followers and views. Whether those views and follows come from people or bots, you are at least nominally buying attention, so the distinction between monetary wealth and attention wealth isn’t quite as clean as my argument would like.
⁵ In marketplaces, wealth has a complex relationship with expertise that’s not explored here. Having expertise can lead to wealth, but a wealthy person also has an easier road to expertise, through education or hiring. This special relation means you can always argue that wealth has been earned by expertise, even if it was earned by a prior generation, but it also creates a dangerous positive feedback loop, where a wealthy person or family buys additional expertise, which turns back into more wealth. For the sake of our group goal, at what point is the increased influence someone is granted by this cycle unacceptably larger than the influence of an average group member?
