WorldAI First Look: 24 Hours in a Free AI D&D Game With Honest Dice and One Moral Compass

Access disclosure: we played WorldAI on an ordinary free account, the same one anyone can open today. Nothing was arranged with the developer, no extra credits or special access came with it, and there is nothing to buy. This first look is independent and unpaid.

This is a hands-on first look at a game its developer still calls in development, not a rated review. We don’t score pre-release software. A full rated review, against our public methodology, follows at launch.

WorldAI is a free, browser-based D&D 5e game in which an AI narrates the story while a server-side engine rolls the dice and keeps your items and stats. The full name is WorldArchitect.AI; the app calls itself WorldAI, and so will we. It is built and run by one developer, Jeffrey Lee-Chan, and it is open to anyone with a Google account or an email address.

We spent 24 hours in it. The short version: it is already an enjoyable game, and our story held together from the first scene to the last. It also has four clear problems — memory that loses detail, slow turns, dice that come out too often, and a world where everyone agrees on whether you were right. None of them stopped us from enjoying the campaign. All four are worth knowing about before you start one.

What It Actually Is

The pitch is on the project’s own about page: a persistent D&D 5e game where “dice rolls are evaluated server-side so the rules are not left to the LLM.” That sentence is the design. The language model writes the scenes and speaks for the characters. It does not decide whether your attack hit, and it does not keep your inventory in its head. An engine does both, and the rolls can be audited after the fact.

You set up a campaign in a short wizard: pick a setting from the built-in list or describe your own in a sentence or two, then build a character yourself or let the game generate one. From there the loop is the familiar one — the game master describes, you act, the system rolls if a roll is needed, the game master tells you what happened. The project keeps a public wiki that works as a player manual, and it is far more thorough than most platforms’ documentation — it includes a design specification of 75 “player stories” describing how the game is meant to behave, which is unusual and which we lean on below. Two caveats on how we use it: a specification says what the game should do, not what it does, and where the two differ, this first look reports what we saw.

According to the wiki, there is no single model playing game master. Each action is routed to one of a dozen or more specialised AI agents — combat, dialogue, level-up, faction management and so on — with the engine underneath holding the state. The same documentation describes:

• A living world. NPCs pursue their own agendas between your actions, and the world advances every few turns or every in-game day; rival factions even take their turns while you are logged off.
• An optional faction mode, switched on when you create a campaign: run an organisation with gold, members, territory and intel, fighting rival factions through a battle simulator.
• God-mode directives: standing instructions you write to shape how the narration sounds for the rest of the campaign — tone, voice, pacing, themes, things to avoid.
• Housekeeping players will like: exporting a campaign’s story as PDF, DOCX or plain text, and duplicating a campaign to try a different branch without losing the original.

Models. The default model is Gemini Flash. There is an opt-in “spicy” mode, off by default, which switches the game to Grok 4.3 for less filtered play. The wiki also lists OpenAI, Anthropic and MiniMax as providers and describes bring-your-own-key support. We played on the defaults.

Price. Free, and at the moment there is no way to pay. More on that below.

The Good News First: It Held Together

Plenty of AI RPGs we test in development show cracks early. A plot thread resets, a character forgets who they are, the rules start bending, or a bug ends the session. WorldAI did not do that to us.

We played for 24 hours and finished with a story we enjoyed, one that never broke. We hit a couple of bugs along the way, and none of them was the kind that ruins a session. For a game its own developer does not yet consider finished, that is a strong result, and it is the thing to remember when you read the criticism below.

The engine is a big part of why. Items and stats live on the server rather than in the conversation, so they do not depend on what the model happens to remember. That is the architecture we describe in why AI RPG mechanics fall apart: the rules only last if they live somewhere the model cannot erode. WorldAI puts them there.

Memory: The Key Facts Survive, the Detail Does Not

The story itself is a different matter from the stats, because the story does live in the model’s context — and a long campaign does not fit in any context.

WorldAI’s wiki describes how it handles that. As a campaign grows, older entries are summarised “so the most important events stay in context,” and the context is compacted automatically when it fills past a threshold. Over a long campaign, that means summaries of events get summarised again. Every pass keeps the load-bearing facts and drops some of the rest.

That matches what we saw. The key facts persisted for the whole 24 hours. Some of the narrative did not: detail that was there in the early scenes was thinner or gone later on. You will not lose the plot. You will lose some of the texture that made the plot yours. Every compression scheme makes that trade; WorldAI’s memory is a bit lossy, not broken. (For why every model does this, and what the better fixes look like, see why AI RPGs forget.)

There is one question we want to test properly, and we want to be clear that it is a question, not a finding. Suppose you kill a character early in a campaign. The engine holds items and stats, but a death is also a story fact, and story facts are the ones that pass through the summaries. If the death does not survive the compression, could that character walk back into the story later, alive? Reading the documentation, we think it is possible. We did not see it happen, and we could be wrong.

The documentation pulls in both directions. The design specification says a companion’s death should be permanent, with no quest or ritual to undo it — so the intent is clearly that the dead stay dead. It also says defeated NPCs are cleaned out of the world state as a campaign runs, so the world does not grow bloated. That is sensible housekeeping, and it raises the obvious question: once a dead character has been cleaned out of the state and summarised out of the story, what is left to remember that they died? Our full review runs a fixed set of memory tests, and this will be one of them.

One Verdict for the Whole World

This is the finding that surprised us most, and it takes a little explaining.

In a good tabletop game, and in the best AI RPGs, each character judges what you do through their own values. Burn down a smuggler’s warehouse and the magistrate thanks you, the smugglers want you dead, and the street kid who slept in that warehouse has an opinion of their own. The same act lands differently depending on who is watching, and that difference is what makes a world feel like it is full of people.

WorldAI did not work that way for us. When we did something that sat in a grey area, the game seemed to decide, once, whether we had been right or wrong — and then every character took that side. Characters whose personalities should have pulled them in opposite directions agreed with each other. If the game decided we had done wrong, we were penalised for it, even when the act was defensible and even when some of the people in the scene should plainly have approved.

This matters beyond fairness. When good characters and evil characters agree about what is moral and what is not, good and evil stop being about values. They become roles: costumes the characters wear, not beliefs that make them act. A villain who shares the hero’s moral scorecard is not really a villain. That villain is the hero’s opinion in a different outfit.

Our 24 hours cannot show where the verdict comes from, and we are describing what we saw, not what the code does. What makes the finding stand out is that WorldAI’s own design says it should not happen. The specification says companions should “approve or disapprove based on their values.” It gives companions a personality type and a D&D alignment precisely as “a quick shorthand for ‘how does this character react in this situation?’” And it says an individual relationship should take precedence over general faction or public reputation. That is exactly the world we described two paragraphs up — the magistrate, the smugglers, the street kid. It is the design we wanted, written down. In our campaign, the reactions converged anyway.

If you play it, the lever the wiki offers is a god-mode directive: “companion rules” are one of its listed categories. We did not test whether a directive such as “each character judges the player by their own values, not the narrator’s” breaks the consensus, so treat that as something to try, not a fix we can vouch for.

It is also a new version of a failure we have written about before. In why AI RPG characters collapse in a party we list the ways a cast of characters flattens into one — usually into one voice. WorldAI’s characters did have different voices. What they shared was a single set of morals, which is a quieter version of the same problem and, in some ways, a deeper one. It also touches the question at the heart of player agency: if the world has already decided what your choice means, how much of the choice was yours?

Honest Dice, Called Too Often

WorldAI’s dice are the good kind. They are rolled on the server, not invented by the narrator, and the wiki goes further than most platforms in explaining how. The model is not allowed to write dice results into the story; it has to request every roll through a server-side tool, and it sees the result only after the roll is made. Each roll is logged with its modifier and result, can be reviewed in the campaign’s roll history, and — according to the specification — carries a seed and commitment so it can be verified afterwards. In a category where plenty of games are narration wearing dice, that is a real strength.

The problem is not whether the rolls are honest. It is how often the game asks for one.

Our character commanded an army. When we summoned our lieutenants to hear their report or to give them orders, the game asked for a roll. That makes little sense. A commander calling their own officers to their tent is not an uncertain act. There is no interesting way for it to fail. The tabletop rule of thumb is to roll only when the outcome is in doubt and failure would change something; WorldAI rolls much more eagerly than that. Per the documentation, it is the system that decides what to roll for your action, not you, and the wiki is explicit that a god-mode directive cannot switch the dice off. So this one cannot be fixed from the player’s side; it needs tuning from the developer’s.

It is a small thing in any single moment and a real thing across 24 hours. Every unneeded roll adds a pause, and every unneeded roll puts ordinary actions at the mercy of chance. It also cheapens the rolls that matter. When everything is a check, nothing feels like one. Tabled showed us the same lesson from another direction: honest dice are one part of fairness, and the story around them is the other part.

The Wait Between Turns

Latency is the other thing you will notice in your first session. The time between sending an action and reading the reply is on the long side. We cannot see inside the system, so we will not name a cause. What the documentation does make clear is that a turn here is more than one chat reply: an action is routed to a specialist agent, rolls go through the server, and the state is updated before you read anything.

It does not break the game. It does change the rhythm. A combat round or a tense conversation loses some of its pull when each exchange comes with a wait, and it adds to the cost of the extra rolls above. If you are used to near-instant replies, expect to adjust.

Free, For Now

WorldAI costs nothing, and at the moment there is no way to pay for it even if you wanted to. The developer considers the game still in development, so no paid plan has been launched. The project’s wiki already describes a future free tier with limits and a paid tier for full access; neither exists today.

The honest caveat is that “free” here depends on the servers. The game is free while the load stays manageable, and that may change as more people come to try it. If a paid plan appears, our directory entry will carry the table.

For now, that makes it one of the easiest games in this category to try: no card, no credits, no trial clock.

Who Should Try It Now

Try it if you want a D&D 5e campaign with real rules underneath and you want to start tonight without paying anything. It is especially worth a look if you have been burned by platforms whose dice and inventory drift. Those are the parts WorldAI is built to get right.

Wait if the thing you care about most is a world of characters with distinct values who argue with each other about what you did, or if slow turns will pull you out of the story. Those are the parts that still need work.

What Happens Next

WorldAI is a promising project that needs more time. The foundation is right — honest dice, an engine that holds the numbers, and a story that stayed coherent for 24 hours. The rough edges are specific, and none of them looks like a wall: call for fewer rolls, shorten the wait, and let characters judge the player by their own values instead of one shared verdict.

We will keep an eye on it through development and publish a full, rated review when it launches — scored on the same seven axes as everything else in our directory, memory test included. Until then, our directory of AI RPG platforms covers everything you can play today.

Disclaimer: WorldArchitect.AI (“WorldAI”) is an independent project by Jeffrey Lee-Chan. Arcanum is an independent publication and is not affiliated with, endorsed by, or sponsored by WorldArchitect.AI. Dungeons & Dragons and D&D are trademarks of Wizards of the Coast LLC; Arcanum is unaffiliated with Wizards of the Coast or Hasbro, and “D&D 5e” is used nominatively to describe rules compatibility. Gemini is a trademark of Google LLC and Grok of xAI. Features and prices described are those of a game in development and may change.