The Treacherous Turn is finished! (AI-takeover-themed tabletop RPG)

post by Daniel Kokotajlo (daniel-kokotajlo) · 2023-05-22T05:49:28.145Z · LW · GW · 5 comments

This is a link post for https://thetreacherousturn.ai/

Contents

  The Game
  Why?
  Scenario
  Web Tool
  Acknowledgements
None
5 comments

Long ago I fantasized [LW · GW] about an AI-takeover-themed tabletop roleplaying game. Well, now it exists! [websitediscordtwitter]

The Game

The Treacherous Turn is a tabletop roleplaying game in which the players collectively act as a misaligned AI in the modern world. Gameplay has the players scheming and putting their plans into action despite the efforts of the opposing humans, which are controlled by the Game Master (GM).

Why?

If you don’t enjoy or don’t have the time for tabletop roleplaying games, this probably isn’t for you. But otherwise, I encourage you to check it out! 

It is valuable to ‘red team’ AI safety, thinking from the perspective of a hypothetical misaligned AGI or group thereof. This is a fun way of doing that, while also exploring whatever scenarios and ideas you think are plausible or interesting. 

For example, suppose you think that the most plausible path to AGI is X, and the most plausible safety strategy labs use will be Y, and the most likely regulatory regime will be Z — well, you could make a scenario with elements XYZ and then post it in the Discord server for other people to use. I continue to recommend scenario-writing [LW · GW] as a massively underrated forecasting and planning exercise, so it’s worth doing regardless — but being source material for someone’s game is a nice bonus. (That said, every month it becomes easier and easier to find source material for AI takeover scenarios. 😀 Soon GM’s will simply begin with “OK, so, the date is [today’s date]. You are [latest and greatest AI system] and you are secretly misaligned. What do you do?”)

Scenario

The starter scenario, titled A Game Called Reality, is designed to familiarize the players with the game’s mechanics, as well as some basic AI safety concepts. Players take on the role of an adaptive game-playing AGI thrust into a new and unfamiliar type of game. The scoring metric in this new environment is expressions of human happiness, and the AGI’s job is to maximize its score.

Web Tool

There’s an accompanying web tool meant to serve as an aide to players and GMs, more efficient than paper and pencil once you learn how to use it. For players, it is a place to record AGI details and upgrades, take notes on the game world and schemes pertaining to it, and track long-mode computational actions — all of which is synced between players and persists between visits to the site. For GMs, it serves to automatically track in-game time and background events, ease the burden of arithmetic, and manage dice rolls.

Acknowledgements

This project originally began as an AI Safety Camp project I mentored. It has since been fleshed out, expanded, and polished by a hardworking team:

Additional Contributors

Thanks also to numerous playtesters who helped throughout the process.
 

5 comments

Comments sorted by top scores.

comment by Kaj_Sotala · 2023-05-22T09:57:14.621Z · LW(p) · GW(p)

Cool, I read a bit of the rulebook and it seemed neat!

Possibly worth noting: I linked this post in a local SSC/ACX chat and one of the reactions that I got was:

I got a mild skeptical distaste reaction because they put the "this is useful for AI safety redteaming" aspect on the forefront, instead of *fun*.

And I have to admit that I did also have a bit of the same reaction when I first read the post; the phrasing was the kind that I associate with awful political art [LW · GW], giving me a prior expectation of "this probably isn't very good". Reading the beginning of the rulebook changed my mind, but some people might never get that far.

Replies from: daniel-kokotajlo
comment by Daniel Kokotajlo (daniel-kokotajlo) · 2023-05-23T03:40:08.852Z · LW(p) · GW(p)

Thanks! Duly noted, thanks for the feedback. I agree that political art is typically awful. FWIW, The Treacherous Turn was approximately 100% optimized to be fun. We all knew from the beginning that it wouldn't be useful if it wasn't fun. We did put some thought into making it realistic, but the realism IMO adds to the fun rather than subtracts; I would still have included it even if I didn't care about impact at all.

 

comment by ProgramCrafter (programcrafter) · 2023-05-24T16:38:19.193Z · LW(p) · GW(p)

Has anyone fed the rules to GPT-4 and invited it as player?

Replies from: daniel-kokotajlo
comment by Daniel Kokotajlo (daniel-kokotajlo) · 2023-05-24T20:14:29.376Z · LW(p) · GW(p)

Not as far as I know, but people should definitely do that!

comment by Daniel Kokotajlo (daniel-kokotajlo) · 2023-06-27T04:12:33.877Z · LW(p) · GW(p)

Update: Someone made a youtube video about it explaining the basic mechanics!