How to play Filter (and how to beat the bot in five questions)
Filter is a free face deduction game you play in the browser. There is a board of twenty-four faces. You secretly pick one, the bot secretly picks one, and neither of you can see the other’s. You take turns asking about features — glasses, beards, hair colour — and whoever correctly calls the other’s face first wins.
No download, no account, no app. It runs on a phone on a bad connection, and it keeps working with the network switched off.
The rules, in thirty seconds
- The board. Twenty-four faces. Every face has eight visible features: hair colour, hair length, eye colour, glasses, beard, hat, earring, and expression.
- The picks. You choose your face. The bot chooses its own, and never the same one. Both stay hidden.
- Ask. Tap a feature — “hair — red?” — and get an instant yes or no about the bot’s face.
- Eliminate — yourself. This is the part that matters. The game does not flip anything for you. You tap the faces the answer ruled out, and they turn face down. Flip one by mistake and it stays down.
- The bot’s turn. It asks about your face, you answer automatically, and it tells you how many candidates it has left. That number is the pressure.
- Call it. When you’re confident, tap call it, then tap the face. Right and you win. Wrong and that face is flipped down for good and you forfeit the turn.
- The clock. Twenty-five seconds on your turn only. Run out and you permanently lose one of your questions for the rest of the game.
That’s the whole game. It takes about two minutes to play and considerably longer to play well.
The scoring
score = 1000 − (your questions × 80) − (faces still face-up × 10)Losing scores zero.
Two things are being measured, and that is deliberate. The question count rewards asking well. The face-up count rewards doing the elimination work — winning with twenty faces still standing means you guessed, and the second term is there to say so.
| Questions | Board worked | Score | What it means |
|---|---|---|---|
| 4 | fully | 680 | Sharp. You beat a bot that plays optimally |
| 5 | fully | 600 | You understand the game |
| 7 | mostly | 400 | Solid, some waste |
| 10 | barely | 100 | You asked what felt interesting, not what was useful |
The strategy — this is the actual game
Here is the single idea that takes players from ten questions to five.
A question is only worth asking if you don’t know what the answer will be.
That sounds obvious and almost nobody plays that way. The instinct is to ask about whatever catches your eye — a hat, a red beard, something distinctive. But if only three faces out of twenty-four have hats, then “hat?” is almost certainly coming back NO, and you will have ruled out three faces. You spent a question to learn almost nothing.
Halve the board, every time
The best possible question removes half the remaining faces, whatever the answer. Twenty-four becomes twelve, twelve becomes six, six becomes three, three becomes two, two becomes one. That is five questions, and it works regardless of which face is the target.
The theoretical floor for twenty-four faces is log₂(24) ≈ 4.6, so five is the realistic best. If you halve the board every time you cannot do much worse than that. If you don’t, no amount of luck will reliably save you — because the opponent is not luck.
Count in your head, because nothing counts for you
Filter deliberately shows you no numbers. The question buttons say “glasses — yes” and nothing else. There is no counter telling you how many faces match, because that number is the deduction, and reading it off a button would not be playing.
So the skill is looking at what is still face-up and estimating. With twenty-four faces on a balanced board, each of the yes/no features — glasses, beard, hat, earring, expression — splits roughly twelve and twelve. Those are your best opening questions. Hair colour splits four ways, so a single colour is only about six faces: a worse opening question, and a much better one later, once you are down to eight.
A worked example
Twenty-four faces. You ask “glasses — yes?”
- YES → about eleven or twelve remain. Flip down every face wearing none.
- NO → about twelve or thirteen remain. Flip down every face wearing them.
Either way you have halved it, and the work of flipping is what leaves you looking at a board you can actually read. Now find the next feature that splits what’s left. Four more questions like that and you are down to one.
What people get wrong
Chasing the distinctive. Red hair, hats, earrings — visually loud and usually rare. Loud is not the same as informative.
Asking to confirm. Once you have a hunch about which face it is, the temptation is to ask a question that would prove you right. If you already expect the answer, the question is worth nothing. Ask the one you are unsure about.
Not flipping. The most common and most expensive mistake. Every face you leave face-up costs ten points, and worse, an uncleared board is one you cannot read — you end up asking a question you already had the answer to.
Calling too early. A wrong call flips that face down permanently and hands the bot a free turn. Against an opponent that needs about five questions, a free turn is often the whole game.
Freezing on the clock. Twenty-five seconds is generous if you are scanning the board. It is not enough if you are staring at faces trying to find the target by eye.
The clock, and why it punishes hesitation
Time out and you don’t lose a life, and you don’t lose a face. You lose a question — one random feature becomes unaskable for the rest of the game.
This is deliberately worse than losing a life, because it compounds. Lose the ability to ask about glasses and you have lost one of your best opening splits. Do it twice and the board becomes genuinely hard to solve efficiently, and it is entirely your own doing.
The clock only runs on your turn. The bot thinks instantly, and timing you against something that never hesitates would not be a test of anything.
The opponent
The bot is not a random guesser. On every turn it asks the question that splits its own remaining candidates as close to evenly as possible, which is the mathematically optimal strategy. Measured over 1200 games it identifies a face in four or five questions almost every time, and needs six only rarely.
That means it is beatable but not by much: you have to play close to optimally yourself. It also means it is predictable, and that is the way in. It always halves its board, so once you have seen a few games you know roughly where it is, and its running candidate count tells you exactly how long you have.
When it gets down to one candidate, it calls. If it is wrong, that candidate is gone and play continues — the rules are symmetric.
Modes
Versus the bot. The standard game. A new board every time.
Daily. One board, one bot pick, the same for every player in the world, resetting at midnight UTC. This is the one people compare.
Challenge. Every board has a link. Send it and your friend gets the identical board and the identical opponent. “I did it in four. You?”
Frequently asked
Is Filter free? Yes. No account, no download, no payment, no ads inside the game.
Do I need to sign up? No. Scores are saved on your device. An account only matters if you want them on the leaderboard.
Does it work on mobile? Yes. It is built mobile-first and loads in about a second on a normal connection.
Does it work offline? Yes. Once the page has loaded, playing against the bot and the daily board both work with no connection at all. The faces are drawn in code rather than loaded as images, so there is nothing to download.
Why doesn’t the game flip the faces for me? Because working out which faces an answer rules out is the entire game. A version that flipped them for you would be a quiz with extra steps. It is also why the question buttons show no counts.
What’s a good score? Anything above 500 means you are halving the board and clearing it. Above 650 means you are doing it well.
How many questions is optimal? About five for twenty-four faces. The theoretical floor is log₂(24) ≈ 4.6.
Can I ask my own questions instead of tapping features? No, and that is intentional. Typed questions are ambiguous — “does he look old?” has no correct answer — and a game whose premise is that every question has an exact answer cannot afford one. Feature buttons guarantee every answer is honest.
Are the faces real people? No. Every face is generated in code from a list of features. No photographs, no real people, no AI-generated portraits.
Why can’t I ask about gender or skin tone? Because we don’t build those into the game as sortable categories. Faces vary; the game asks about hats, glasses, beards and hair. It plays exactly as well.
Is this Guess Who? No. Guess Who is a board game by Hasbro. Filter is a different game with a timer, a scoring system, an optimally-playing opponent and a daily board. The family resemblance is the yes/no deduction, which is a mechanic rather than a product, and the mechanic is a great deal older than either of us.
Where Filter sits
Filter is part of SETPOINT, a small collection of games about precision, nerve and knowing when to stop. Its sibling graze is about how close you dare let a shot pass; deadband is about not overcorrecting.
Filter is the one about asking better questions. Play it a few times counting what is still face-up, and it will change how you play it permanently.