AI + D&D: Where the Genre Is Stuck

Reflections after 40 minutes with a recently released AI RPG beta

Last Friday I spent about forty minutes with the beta of a new AI native RPG product. The team behind it is one of the most established players in this emerging space, and they’ve been pushing the frontier of AI gaming for years. The experience left me thinking, not so much about whether this particular product succeeds or fails, but about a set of design questions that I think the whole AI native game category needs to grapple with.

The experience leaned heavily on AI generated text and TTS to create a sense of immersion. That part was clearly the strength. But the systems around it felt disconnected, and the overall flow kept breaking. It got me thinking about a problem I suspect many teams in this space are running into: how do you avoid stacking the weakest parts of three different media (tabletop, video games, and creation tools) into a single product?

This category is new, and reasonable people will disagree about what AI native games should even be. So this isn’t meant as a verdict on anyone’s work. I’m genuinely grateful that products like this exist, because they give the rest of us something concrete to think with. The point of this essay is to share where I think the design challenges are, and to open a conversation.

My own view is that D&D is one of the most exciting forms AI could reshape. Its openness, its improvisational nature, the way rules and narrative intertwine, all of it almost feels purpose-built for generative AI. But the moment you put D&D on a screen with AI doing the heavy lifting, you’re no longer making D&D. You’re making something new. And new forms need their own design principles, not borrowed ones.

What follows is in six parts. The first describes two friction points I ran into. The middle sections analyze the underlying design traps. The last sections lay out what I believe the right principles for this category look like, and what UGC means in this context.

1. Two Breakdown Moments

The product is an “AI native RPG platform” launched in beta this April by the makers of AI Dungeon. It runs on a proprietary World Engine that’s meant to keep track of world state and rules, including characters, inventory, relationships, and consequences. World creators can define settings and starter characters. Players can pick a preset character or build their own, in single-player or multiplayer. The ambition is real and the technical claims are interesting, which is part of why I came in with high expectations.

Beyond the well-crafted dialogue and voice work, two parts of the experience really pulled me out of it. They’re worth describing in detail because they each point to something deeper about how AI is being layered onto traditional RPG design.

1.1 Combat: Drawn Out and Hard to Read

I was playing an elf, partnered with a dwarf companion, helping a tavern clear out rats in its basement. Through trial and error I was figuring out which spells my character knew. At some point I even called the tavern keeper down to help. Twenty minutes in, the rats were still alive.

Combat feedback was muted in a way that made the fight hard to read. Every time I thought I’d landed the finishing blow, the screen would say something like “they’re still moving” or “more of them are pouring in.” I could see individual damage numbers, but no total HP, no progress bar, no sense of where I was in the fight. The prose kept generating florid descriptions of my attacks, and every paragraph ended with some version of “the enemy is not dead yet.”

Strategy was also hard to engage with, because the combat system itself wasn’t legible. My character sheet showed a class but no skill list. So from turn one I was guessing. Try fire? Try lightning? Try some spell that might or might not exist? Every supposedly clever combo I came up with was really just a shot in the dark. And each guess required waiting nearly a minute for AI generation. This pattern repeated for fifteen rounds.

The team later mentioned that the rats being unkillable was likely a bug, which I take at face value. But fixing the bug doesn’t address the underlying issue. Even if the rats died on round ten, the first nine rounds would still suffer from the same pacing, visibility, and strategic problems. The deeper question is that AI alone isn’t enough. Combat needs a robust numerical and feedback system underneath it, and that part can’t be improvised away.

1.2 The Narrator: A Broken Contract

The other friction point came from a side panel called the Narrator. It’s designed to help players navigate the game, a guide in concept. In practice, it ended up causing one of the most jarring parts of the experience.

First, the Narrator gave information that contradicted what the game itself was doing. When I asked what spells my character could cast, it clearly listed fire and lightning but said I couldn’t use frost. Meanwhile, in the actual game, I had just tried casting frost on the rats, and the AI DM had accepted it as a successful spell. Two systems giving me conflicting answers in the same session. The moment that happens, you stop trusting either of them, and trust is the substrate that roleplay sits on.

More striking was what happened when I asked the Narrator for help against the rats, the way you’d ask for a strategy hint. The responses included: you can set the rat group as an NPC so you can see their HP, you can reduce their health, or I can just end this fight for you. I stopped and reread those lines. If twenty minutes and fifteen rounds of combat can be dissolved by clicking a button in the side panel, what did any of it mean?

AI as an in-game guide is genuinely an interesting design space. But the Narrator as implemented here was operating from a developer level or even god level view of the game, not a player’s view. And that shift collapses something important.

In traditional D&D, the player plays, the DM directs, and the rules arbitrate. These three roles stay separate. The player’s investment depends on an unspoken contract: they accept randomness and challenge, but the outcome can’t be rewritten by their own hand. Once the player realizes they have a god mode toggle, every prior attempt, every minute of waiting, every prompt they typed loses some of its weight.

Each of these Narrator features has a defensible design rationale, taken on its own. Giving players a safety net is reasonable. So is adding worldbuilding tools, given the product’s UGC ambitions. The issue isn’t the features themselves. It’s that they sit in the same window, at the same moment, as the player’s active roleplay.

This “we want both” instinct shows up a lot in AI products, and it usually reflects an unwillingness to make hard tradeoffs. In RPGs it cuts deeper than usual, because the whole genre depends on the player’s willingness to believe in the world. Whether the player is an actor or a director is a question that needs to be answered at the product definition stage. If you want to support both, the answer has to be mode-switching, not panels coexisting in the same interface. A single screen can’t hold both immersion and abstraction at the same time without one undermining the other.

2. D&D Is a Collection of Experiences, Not One Experience

Both breakdown moments above point to something deeper. I think a lot of teams approaching AI + D&D treat D&D as a single, transportable experience. That assumption is where the trouble starts.

2.1 D&D as a System of Coupled Subsystems

When people say “D&D” they’re actually referring to a collection of subsystems that work together. At a minimum, the list includes: Narrative (story progression, branching, worldbuilding reveals), Combat (turn-based fights, abilities, dice resolution), Roleplay (acting and deciding as a character), Worldbuilding (geography, factions, history), Social performance (the social texture between players and NPCs), and Rule arbitration (the DM’s judgment calls).

Each of these on its own is a substantial design domain. Six of them braided together is an extraordinarily complex system. Tabletop D&D can carry that complexity because it has scaffolding most digital products don’t: multiple players sharing the cognitive load, a DM who reads the room and trims dead air in real time, physical character sheets that act as visible rule anchors, and the collective imagination of a group keeping the fiction coherent.

Here’s a useful reminder: the first challenge with AI + D&D isn’t actually the AI part. It’s the electronification of D&D itself. Larian spent the better part of a decade, from the Divinity series through Baldur’s Gate 3, getting D&D style play to feel right on a screen. That history alone tells you that translating tabletop into a digital experience is non-trivial. Now imagine doing it without anyone sitting next to you, without a DM who reads the room, without easy access to reference material, and on top of all that, asking an AI to improvise the rules. Porting D&D’s full complexity into that context, unchanged, is going to be very hard to make work.

2.2 The Root Cause: No Controlled Variables

Stepping back, what made the experience hard to enjoy wasn’t any single feature. It was that the product was trying to be too many things at once, with no controlled variables. It wants to be a tabletop experience, a video game, a creation tool, and an AI generation platform, all simultaneously. The goals of those four things actively conflict with each other.

Within forty minutes, I was asked to occupy different roles in rapid succession. One moment I was playing a tabletop character, describing my actions in natural language. The next moment I was a video game player, watching numbers tick down and pushing toward a fight outcome. Then a prompt popped up asking whether I’d like to create a new NPC and set its HP, and suddenly I was a designer. And underlying all of it, I was an AI product user: waiting for generations, editing prompts, retrying outputs.

Four identities, no signposting, no framing for which one I was supposed to be in at any moment. The result wasn’t richness. It was confusion. I genuinely didn’t know whether I was supposed to trust the system, override it, or fill in for it.

The other consequence of having no controlled variable is that there’s no clear target user. Tabletop fans want narrative density and roleplay depth. Video game players want clean mechanical feedback and a sense of progress. Creators want stable, predictable tools. A product that tries to serve all three at once tends to give each group a little bit, but never enough for any of them to stay.

“Doing a bit of everything” sounds like giving the player freedom. In practice, it offloads the design decisions onto the player as a choice burden.

3. Design Principles for AI RPGs and AI D&D

Enough about what isn’t working. Let me offer what I think the productive direction looks like.

My starting claim is this: AI RPGs and AI D&D aren’t “D&D plus AI” or “video games plus AI.” They’re a new form, and new forms need their own native design principles rather than borrowed ones from either parent medium.

The principle I keep coming back to is flow.

3.1 Why Flow Needs to Be Rethought for This Medium

Flow is something every game designer thinks about. Nothing new there. What is new is that AI changes the rhythm of play through latency, through agentic behavior, through the sheer volume of generated text. So flow can’t be ported in from tabletop or traditional video games. It has to be redesigned for this context.

I see this play out in AI + D&D specifically in two ways. The first is about the auditory channel. The second is about latency. Let me take them one at a time, then in the following sections I’ll lay out three design principles that follow from this analysis.

First, AI RPGs lean very heavily on language as the medium of experience. The density of generated text and voiced narration is far higher than in traditional video games. And there’s a property of the auditory channel that’s easy to overlook: it’s good at being a background channel, not a primary one. This is why people can listen to podcasts while driving, or audiobooks while doing chores. The auditory channel has lower cognitive load, which leaves room for attention to wander.

In a tabletop session, that wandering attention is fine, because it gets filled by the social environment, like other players’ faces, the physical dice, gestures, laughter. In a traditional video game, the auditory channel sits alongside rich visual action. There’s always something for the eyes to do. But in an AI RPG, the player is often facing a mostly static screen while listening to long passages of narration. The auditory channel can’t hold attention on its own, and the visual channel has nothing to offer it. That leftover attention has nowhere to go, and it turns into a low grade restlessness. This is why the same act of “spending time on narrative” feels enjoyable at a tabletop, fine as a video game cutscene, and oddly uncomfortable in an AI RPG.

Second, AI generation introduces real, physical latency. Every player action might require nearly a minute of wait time. This kind of delay doesn’t exist in traditional video games, and it certainly doesn’t exist at a tabletop (where “slow” pacing is filled by social interaction, which feels completely different from waiting at a screen). Traditional game design has spent decades on the relationship between map size, player movement speed, and the cadence of stimulus, precisely because long stretches of unrewarded time erode curiosity and momentum. AI games can’t just hand this latency to the player unprocessed. They have to actively design around it.

Both of these observations point to the same conclusion: AI RPGs need their own model of flow. The next three sections describe what I think that model looks like.

3.2 Principle One: Subtract. Pick One Module and Go Deep

Going back to the earlier point: D&D is a collection of subsystems. Each of those subsystems, whether it’s narrative, combat, roleplay, or worldbuilding, is rich enough that AI could deepen any one of them into a standalone product.

AI + Narrative could mean an AI driven branching story with a clean, simple combat system or none at all, focused entirely on delivering a complete narrative flow within a short session. AI + Worldbuilding could mean a generator and simulator for explorable worlds, where the reward loop is discovery and surprise as the world reveals itself. AI + Combat could mean AI generated tactical maps with every encounter feeling distinct, while leaving the underlying numerical and combat systems intact, and using natural language as the medium for coordination and creative problem-solving with companions. This last direction is one I’ve been exploring myself. The image below shows a loop design I’ve been working on, combining a board layout, AI NPCs, and semantic understanding combat.

Each of these is a serious direction in its own right, and any of them would be more coherent than trying to do all of them at once.

“Big and broad” is a luxury tabletop can afford, because it has tens of hours and six or seven people sharing the cognitive load. AI RPGs, whether single-player or multiplayer, don’t have that budget. Flow demands sustained attention on a single, clean core loop. It doesn’t tolerate the player being shunted between modules.

3.3 Principle Two: Separate the Game from the Creation Tool

Back to the Narrator moment from earlier, the one where the system asked me, mid-combat, whether I wanted to create a new NPC and set its HP. That moment surfaces a second principle.

A game wants the player to forget they’re operating a system. A creation tool wants the creator to feel the system at every step. These two goals are in direct opposition.

When a product tries to do both inside the same interface and the same session, you get the worst of both. Immersion gets interrupted by tool prompts. Tool controllability gets constrained by narrative continuity. The result is something that’s neither playable nor authorable, a frustrating middle ground that satisfies no one.

The fix is mode separation. If a product wants to support creation (which is fine!), put it in its own mode. The player explicitly steps out of the play session to enter creation mode, and knows their role has changed. This isn’t a sophisticated design principle. It’s basic, and it’s the kind of basic that a lot of AI products end up overlooking.

3.4 Principle Three: Hide AI Latency Outside the Critical Path of Flow

AI generation time is fundamentally a design problem, not a technical one. With current models, waits of around a minute are common, and that won’t change overnight. But putting that minute directly in the critical path of the player’s flow is a product level failure, and it’s entirely avoidable.

The right pattern is to hide generation outside of the player’s critical action loop. You can fill the wait with parallel, non-blocking activities, such as letting the player browse inventory, organize equipment, or talk to companions. You can use streaming output so the player starts reading while the rest is still being generated. None of these are new techniques. They’re well-established patterns.

A nice example of getting this right is the AI Riverside Scene at Qingming I tried on Xiaohongshu a while back. They handled the rhythm beautifully: the first section streamed in, and by the time you’d finished reading the half-screen of text in front of you, the rest of the chapter was already generated. The whole experience flowed without ever surfacing a wait.

AI Qingming example

AI Qingming example

The beta I played, by contrast, used the simplest possible serial design. Player inputs an action, waits a minute, sees the result, inputs the next action, waits again. AI latency, especially in combat, sits right on the player’s flow path. Every input becomes a hard interruption. And the wait isn’t even rewarded with a meaningful payoff like a new ability unlocked or a climactic moment resolved. It’s just more turns of the same fight.

Flow continuity matters more than almost any other design goal. A product that interrupts the player on every input ends up giving them not a coherent experience, but a string of disconnected moments.

4. UGC Ambition and the Quality Signal

One last topic: UGC.

Almost every AI RPG product in this space positions itself as a UGC platform, with the goal of letting creators build their own worlds using the platform’s tools. The direction itself is fine. AI Dungeon was already heading there years ago. But UGC platforms have one very hard prerequisite: the platform owner needs to ship at least one polished, replayable example first. That example is what proves the tools can produce something worth using. This is the cold-start rule for any UGC product.

For AI RPGs specifically, this prerequisite is harder to meet than usual, because the competitive landscape is unforgiving.

The active creators who came out of the AI Dungeon era have more options than they used to. SillyTavern, Janitor, Character.ai and various derivatives all offer creator-friendly environments. If an official platform can’t demonstrate a sufficiently polished first-party showcase, there’s a real risk that creators just go elsewhere, to places where they have more room to express what they want to make.

UGC platforms get their flywheel going through quality signal, and that signal almost always has to come from the platform owner first. Talking up UGC before shipping a single demo worth playing tends to put the cart before the horse.

Closing Thoughts

I came away from this experience feeling a kind of mild regret. AI + D&D is a direction with real potential, and even the most ambitious attempts so far seem to be getting caught in the same trap: treating the goal as “recreating D&D,” rather than as “using AI to reinvent the most valuable part of D&D.”

Subtraction is harder than addition. But finding a meaningful flow loop first, and then expanding outward from there, may be the only path that actually works for this category. I’m writing this less as a critique of any single product and more as a set of notes for the road ahead. Anyone working on this, myself included, is figuring it out as we go. I’d love to hear what others think.