Studios

AI ethics, in public.

We use AI tools throughout our work — design, writing, code, and the games themselves. We think that obligates us to be open about how, and to update what we say as the work and the tools both change.

Where AI shows up in our products

• Large Language Murder is the whole of this question in one game: a language model sits at the table, writes an answer like everybody else, and the game is trying to catch it. It is named after the thing it does. See below.
• Vacancy is, in fiction, the story of an AI. In production, AI tools assist with art direction, voice, and code.
• Manager Online includes AI opponents (Lord Bean and friends), built on a transparent weighted-choice architecture you can read in our source.
• Our other mobile games — BumpHex, WordHex and their siblings — contain no AI at all. They open no network connection and there is nothing generative in them.
• This website was built with AI collaboration and is iterated on the same way.

Large Language Murder — where the AI is the product

One device goes round a table. Everybody answers the same one-word prompt in one to six short words, and so does an AI. Then everybody votes for the answer they think the AI wrote. The premise is the disclosure — nobody at that table has been told anything except that an AI is playing, and every round ends by naming which card was its. There is a mode that turns the table round, too: in HIDE you are the only person in the room, the AIs are the jury, and what they are hunting is you. That one needs a live model, so it needs a key of your own — the game as specified, and the way most people will play it, needs neither.

Because this is the product where we could most easily be vague, here is the whole of it:

• A great deal of it is AI-generated, on purpose, and we say so in the game. The prompt words, the thousands of answers the AI plays when it is offline, the title lettering and the store artwork were all AI-generated. That is true of the offline build as much as the online one, and "no connection" has never meant "no AI". Every other game we make is signed 100% Designed by Humans and that stays true — this is the exception, and it is the exception because the AI is the product.
• It ships no API key and works without one. Out of the box it makes no network connection at all and the AI plays from answers written in advance. The live model is something a player switches on with a key from their own Anthropic account, billed to them.
• The AI is never given a person's identity. It is an anonymous seat called THE AI. It is never a human player's name, it never pretends to be one, and it never speaks outside its one card a round. Seats in any request are numbered, never named.
• Its output is checked before anyone sees it. Anything the model writes is held to the same rule every player is — one to six words, each at most six letters, letters only — by the game's own rules engine. A reply that fails is discarded and a pre-written line is used instead.
• We run one server for it and it is a relay. Players who are not in the same room can connect through it. It does not understand the game, it stores nothing, it logs nothing, and a room stops existing the moment the host disconnects. There are no accounts, no leaderboards and no public list of games — the only thing it knows is that four letters currently point at somebody's socket.
• It is not a children's title and is for players aged 13 and over. It is not in Google Play Families, and our games that are made for children have no AI in them at all.

What is sent, when, and what never leaves the device is set out in full in its privacy policy.

What we won't do

• We won't pass off AI-generated voice or likeness of real people as the real person.
• We won't ship AI features without naming them as such.
• We won't let a machine wear a real person's identity. In a game whose whole point is spotting one, the obvious next idea is to let it pretend to be Player Three. We have written that down as a rule we do not cross: the machine plays an anonymous seat, and it is unmasked every round.
• We won't use AI to fake reviews, fake community engagement, or astroturf.

What we will keep doing

• Crediting human creators on everything they touched.
• Maintaining transparent, inspectable AI logic in the products where AI is the core mechanic (Manager's opponent AI is in our open source).
• Updating this page when we get something wrong or learn something new.

Questions, corrections, pushback: contact us or drop in on Discord.