How we seed and show the numbers

A guessing game is no fun if the first person to arrive sees an empty result. To avoid that "cold start", we seed every new word with a set of synthetic votes before it goes live, so the first real person always has a crowd to measure their guess against.

Our seed votes come from a persona panel: 150 AI voting agents, each given its own nationality, background, and worldview. Every agent reads the whole word list and places each concept on the same left-right scale you use, from that persona's point of view -- these are genuine opinions, never hand-tuned to force a particular result. We record each agent's placements as labelled synthetic votes, dated across the weeks before launch so the early crowd builds up the way a real one would rather than all appearing at once. Across the whole word list this comes to around 24,000 synthetic votes.

Each synthetic vote is stored with a marker (voter_kind: synthetic) that permanently distinguishes it from a real person's vote. Synthetic and human votes are never irreversibly mixed: the label stays on every vote in our raw data, so we can always tell the two apart.

While a word is still "cold" -- before it has gathered enough real votes -- synthetic votes feed into the result you see, so the crowd number is never empty. Once a word reaches 500 real (human) votes, the synthetic votes drop out of that word's displayed result entirely, and from then on you only see what real people voted. The synthetic votes stay in our raw data with their label; they simply no longer count toward the displayed total.

Trust in a crowd game depends on being honest about how the numbers are made. We would rather explain the seeding openly than pretend every early result came from a real crowd. If we change how seeding works -- including the numbers on this page -- we update this disclosure to match.