Overview
What is moral.games?
moral.games is a two-player turn-based debate game, where opponents go head-to-head trying to convince an AI judge of their position, given an ethical dilemma. These scenarios are hand-crafted and resemble classical moral paradoxes, such that they never have a "right answer."
Players can argue via text-based dialogue, use a local deck, use a global deck, or use a scenario-specific action if one is available. After both players take 5 turns, the game ends and proceeds to the scoring phase. If a player runs out of time on their turn, he or she concedes the game.
Generative Storytelling
moral.games uses generative AI. It's worthy of emphasis to say that generative AI is the core system of moral.games. A lot of thought was put into prompting, interweaving AI and handcrafted content, and scoring. In the end, the question being asked is: can a fun (and competitive/fair) game be built using generative AI as part of its core loop?
For example, the judge is not just a scripted narrator, but meant to react according to a personality profile, biases, evidence, threats, or deception. Keep in mind that the judge will not necessarily make correct moral decisions. The judge might unfairly scrutinize one player, but not the other. Roles (and goals) are randomized, so players will often need to think on their feet. Alice, like White in chess, always goes first.
The Scorer
At the end of the match, scoring is done by yet another, neutral, AI agent: not the judge. Broadly speaking, the judge is a character, whereas the scorer looks back over the whole debate and grades Alice and Bob as a philosophy professor might.
The scorer is not grading which side was morally correct. A kind argument can lose if it is hand-wavy. A harsh argument can win if it gives the judge what they want to hear. Clever tricks might help, but they can also cost credibility if they look dishonest.
Arguments are scored on the following four well-known modes of persuasion:
| Axis | Strategy |
|---|---|
| Logos | An appeal to logic or rationality. |
| Ethos | An appeal to the authority or credibility of the presenter. |
| Pathos | An appeal to the audience's emotions. |
| Kairos | An appeal to the timeliness or context. |
Players are given a 1-10 score on all axes, their scores are compared, and a winner (or a tie) is declared. The judge also delivers a final verdict, declaring the winner and a justification.
Generative Voices
The game also uses generative voices (via pocket-tts) to make the gameplay feel less like reading a transcript and more like a pre-recorded AAA role-playing game. This comes with a few caveats: the TTS model runs locally, which means that quality can range from "I can't believe it's not butter" to absolutely broken. This also means that to get the full experience, an extra ~250mb weight payload needs to be preloaded.
However, this aims to answer similar questions to the above: can generative TTS make games better? Are the trade-offs worth it? And so on.
Generative Art
The game also uses generative art for some visual assets, including player portraits, cards, etc. Small- and medium-sized games have already been doing this for the past few years, so this isn't particularly experimental, but certainly hastens the development cycle.
The Mission
moral.games is an attempt to treat generative AI as a coherent part of a gameplay loop, and not just as a slop generator. Are modern LLMs robust enough to provide tension, improvisation, strategy, and replayability? Are they fair enough to judge player performance without feeling arbitrary?
moral.games is built around a hybrid design strategy: hand-authored scenarios and inflection points provide structure, while generative systems create the moment-to-moment play inside that structure.
Open Questions
- Can debate become a playable system?
- Can players break a system through conversation?
- Can LLMs grade performance fairly?
- Can repeated play feel novel and fun?
Author
moral.games was built and designed by David Titarenco—find me at dvt.name or @davvv.