What Computers Can Honestly Do for Collective Decisions

An old note of mine reads, in full: a new democracy, a proposal to reduce stupidity in decision making; use computers to control the government. I want to unpack that, because the compressed version is easy to misread as the fantasy that machines should govern, and that is not what I meant and not what would help.

The stupidity worth reducing is not the stupidity of people. People are about as clever as they were in Athens. The stupidity worth reducing is aggregation stupidity: the loss that happens between what a population knows and wants, and what its decision procedure produces. That loss is enormous, it is mechanical, and machines are good at mechanical things.

Making Many Decisions Cheap

The first and largest contribution is boring: cost.

A national referendum on paper is a major operation. It needs polling stations, staff, counting, weeks of logistics, and it costs enough that a country holds one every few years at most. That cost is the practical reason we elect representatives: we cannot afford to ask the public about each question, so we ask it once about a person and let the person answer the questions.

Computers make asking cheap. A decision that once cost a national mobilisation costs, in principle, what a bank transfer costs. Once asking is cheap, the argument for bundling forty questions into a single vote for a single person collapses. You can ask the forty questions. I have written about what that bundling does to the quality of debate; the machine is what makes unbundling affordable.

That alone is not a small thing. It is the difference between a system that consults the public once a term and one that can consult it whenever there is something to decide.

Exposing the Information

The second contribution is transparency, in the literal sense.

Most of what a government does is recorded, and most of the records are in principle public, and almost none of it is legible to anyone. Budgets run to thousands of pages. Legislation is amended by reference to clauses in other legislation. Who met whom, who funded what, which contract went to which firm at what price, all of it exists and none of it can be read by a citizen with a job.

Machines can hold the whole of it and answer questions about it. Not summarise it into a slogan, but let a person ask: how much did this ministry spend on this category, and who supplied it, and what did the same thing cost last year. That is not artificial intelligence. It is a database with a decent interface, and its absence is a choice.

Running Juries

The third contribution is to the decisions that should not go to the whole public.

Most decisions are technical. Nobody wants a national vote on the specification of a bridge. But somebody has to check the people who decide such things, and the check cannot be another official from the same hierarchy. The classic answer is a jury: ordinary citizens, chosen at random, given the file and the time to read it.

Selecting a jury at random from the population, verifying that the selection was random, giving it secure access to the material, and publishing what it concluded and why, are all things a computer does well and a bureaucracy does badly. The bureaucracy has an interest in the outcome. The random number generator does not.

Recording Who Decided What

The fourth contribution is memory.

A great deal of political stupidity survives because nobody remembers. A promise made in one year is contradicted in the next; a policy fails and the people who advocated it are advocating the next one; a vote was cast and the record of it is somewhere nobody looks. I have argued that this forgetfulness is one of the structural advantages of the known liar.

A system that records every decision, who took it, on what stated grounds and with what predicted result, and then keeps the record next to what actually happened, changes the incentives of everyone inside it. It is the cheapest form of accountability there is, and it requires no new law, only the refusal to let the record be lost.

What They Cannot Do

Now the limit, and it matters.

Computers cannot decide. Not because they are not clever enough, but because a decision about what a society should do is a decision about whose interests count and by how much, and that is not a computation. Any machine that appeared to decide would be executing somebody's weights, and the person who set the weights would be the real government, unelected and invisible. That is the fantasy version of "computers controlling the government," and it is worse than what we have.

The right division is exact: people decide, machines make deciding cheap, informed, checked and remembered. Every item on that list attacks aggregation stupidity and leaves human judgement where it was. That is the whole proposal, and it needs no intelligence from the machine at all, only the ordinary ability to count, store and show.