Could an AI actually DM your game? What a Virtual DM would (and wouldn't) do

The fantasy and the real question

Every few months a new product launches that promises to be your AI Dungeon Master. Friends & Fables. AI Dungeon. A handful of agentic LLM experiments on Reddit. The pitch is consistent: get an AI to run your D&D game so you can play whenever you want, no scheduling, no DM burnout, no flaky players.

Half of D&D Twitter loves the idea. The other half thinks it's the death of the hobby.

Both takes are too tidy. The real question isn't "will AI replace DMs," because the answer to that, today, is no. The real question is "what specifically can AI do well in the DM role, where does it fall apart, and how is the gap likely to evolve."

I've been thinking about this for a couple of years now, partly because I work on Dungeon Diary and we get the question often, partly because I've actually run a few experiments with current LLMs in the DM seat. This post is what I've learned, what I think the realistic timeline is, and where I think the line will be drawn.

What current LLMs do passably well in a DM role

Let me start with what works. I've sat down a handful of times with a custom GPT or Claude project loaded with a campaign world and asked it to run a session for me. Here's what surprised me about how well it handled some things.

Improv NPC dialogue. Within a single conversation, current LLMs are genuinely decent at staying in character for an NPC. Give the model a name, a personality, and a goal, and it'll talk like that person for the duration of the scene. Not Oscar-winning, but better than I expected.

Generating consequence on the fly. "I attack the bartender." Okay, the model riffs: "The bartender ducks and shouts for the bouncer. Two large men step out from the back room. Roll initiative." That's fine. Functional.

Skill check adjudication. "I want to climb the wall." The model picks a DC, asks for a roll, narrates the outcome. It picks reasonable DCs, more often than not.

Lore recall. If the campaign data is in the model's context, it'll remember names and references reasonably well. "Tobin Vance is the innkeeper of the Crooked Anchor in Brassgate." Fine.

Encounter narration. Reading out attack results, describing damage in flavored language, narrating spell effects. Surprisingly good at this. Probably because it's a high-volume task with clear structure.

Filling small gaps. "I look around the room for clues." Generic but functional descriptions, often with a couple of options the player can investigate.

In all of these, the model is doing the thing a competent first-time DM might do. Not great, but not embarrassing.

Where it falls apart

Now the other side. Current LLMs hit walls in DM role that a human DM doesn't.

Tonal consistency over many sessions. A great campaign has a feel. Curse of Strahd is gothic horror; Saltmarsh is gritty pirate intrigue; my homebrew has a kind of resigned post-empire tone. A skilled human DM holds that tone across every NPC, every scene, every loot drop. Current LLMs drift. They'll start in the right tone, then twenty exchanges later they're describing things in a slightly different register, and forty exchanges later they've lost the thread. Without conscious intervention, the campaign feel erodes.

Player buy-in management. A real DM watches the table. Sees which player is checked out. Notices that one PC hasn't had a moment to shine in two sessions and engineers a scene where they do. Catches the energy in the room and adjusts. LLMs can't see the table. They can't read body language, voice tone, or the side conversation happening between players. They get player engagement only through what's typed.

Calibrated stakes. When should a fight be deadly? When should an NPC die? When should the world give the players a win? These are judgment calls that depend on the season of the campaign, the players' moods, recent table events, and what would make for the best long-term story. Models can't make these calls because they don't have the context. They can be told (by a configuration prompt or tool), but they don't develop intuition for it.

Surprise. A great DM moment is the unscripted twist. The NPC the players assumed was a villain turns out to be the only one telling the truth. The fight that seemed routine reveals a nested betrayal. These moments emerge from a DM's creative reaction to specific player choices. Current LLMs are statistically inclined toward expected patterns. They generate the average, not the surprising. They can be prompted toward the unusual, but it doesn't come naturally.

Long-form continuity. A campaign over 50 sessions builds up a continuity that current LLMs can't hold in context. Even with retrieval-augmented generation pulling from a campaign knowledge base, the model doesn't have a coherent sense of "this is what's happened over the last six months." It has facts but not a narrative throughline. Major callbacks, foreshadowing, the ten-session payoff that pays off because the seed was planted ten sessions ago, these are hard to engineer through retrieval alone.

Fairness and adversarial play. A DM is the players' opponent during combat and the players' ally during everything else. This dual role is delicate. Current LLMs lean too hard one way or the other, depending on prompting. Either they're pushovers (every encounter is easy because they want the players to feel good) or they're sadistic (every encounter is brutal because they took "be challenging" too literally). Calibrating to "the right kind of opposition" requires more than a prompt.

Improv that involves rules. A new player asks "can I do this thing the rules don't directly cover?" A skilled DM says yes or no with a quick ruling that fits the moment. LLMs sometimes do this well, but they're inconsistent. Sometimes they over-rule the player into restrictive readings of the rules. Sometimes they say yes to everything and break the game. The "yes, and" or "yes, but" judgment is a DM's craft, and current models don't have it reliably.

The gap is real but not infinite

Now here's where I'll go against the "AI will never replace a DM" crowd a little.

The gap I described above isn't a fundamental impossibility. It's a current capability ceiling. Every problem on that list has a known direction of progress in AI research right now.

Long-form continuity? Long-context models, retrieval, and memory systems are improving quickly. The 1M-token context windows in Claude and Gemini today were science fiction two years ago.

Tonal consistency? Better fine-tuning, better instruction-following, and better context management are converging on this.

Player buy-in management? Multimodal models that can see and hear the table (with consent and privacy considerations) are coming. It's not far off that an AI DM could literally see whether players look bored.

Surprise and creativity? This is the hardest one. Current models genuinely lean toward the average. But fine-tuning on specifically creative DM examples, plus reinforcement learning from human feedback specifically on "this twist was good," is a real research direction.

Fairness and adversarial play? This is solvable through better instruction frameworks and tool use. A DM AI that has access to encounter difficulty tools, party state tools, and pacing tools can make better calls than a chat-only model.

The wall right now isn't technical impossibility. It's that the current generation of models is general-purpose. They aren't trained specifically for the DM role. The specialization hasn't happened yet.

What a real Virtual DM would look like

If I were designing a serious Virtual DM, here's what I'd build (and might, eventually).

Campaign-aware foundation. The DM AI has access to the full campaign world as structured data. Factions, NPCs, settlements, lore, character sheets, faction reputation, recent events. Not stuffed into a prompt, but properly structured and retrievable. This is largely what Dungeon Diary already does for the human DM use case. Extending it to a Virtual DM is mostly a matter of agent design, not new infrastructure.

Tool use for mechanical correctness. The DM AI doesn't try to do D&D math in its head. It calls tools. Roll dice, calculate damage, check spell rules, validate a save. This way the rules are always correct even if the model itself can't perfectly remember every condition modifier.

Pacing and tone tools. The model has access to "what's the campaign's tone" and "what's the pacing of the current session." It checks these as it generates. Off-tone outputs get filtered or regenerated. Pacing checks ensure the session has the right rhythm of combat, social, and exploration.

Player state tracking. The model tracks (with player consent) which players have had recent spotlight time, who's been less engaged, who's been quiet. It engineers scenes that bring the quiet players in.

Multi-step planning. Not "respond to the next player input." Plan the next session arc, plan the next three sessions, plan the campaign climax. Update the plan as players make choices. This is where current models fall short, and where agentic frameworks are evolving fast.

Human override. A real DM stays in the loop. Maybe the DM is one of the players, with the AI handling NPC voices and rules adjudication. Maybe the DM steps in for major story beats and the AI handles the minutiae. Maybe the AI runs full sessions for groups that don't have a DM, with the players collectively able to push back on outputs they don't like.

This isn't science fiction. Most of these pieces exist or are close to existing. The work is in integrating them into something that runs a coherent game.

What we're considering at Dungeon Diary

Full transparency. We've been talking about a Virtual DM mode for Dungeon Diary. Not as a replacement for human DMs, but as an option for groups that want to play when their DM can't make it, or solo play for people who don't have a group, or training wheels for first-time DMs who want to learn by example.

The shape of it would be something like: a campaign you've already built (or one we provide), a party of player characters, and an AI agent that runs the DM role using OpenRouter (so the model is swappable as better ones come out). The agent has access to your campaign collections (NPCs, factions, lore, encounters), tools for dice and rules, and a playbook for pacing and tone.

We're not committing to a date because we don't want to ship it before it's actually fun. Current LLMs hit some of the walls I listed above hard, and a half-baked Virtual DM is worse than no Virtual DM.

If you want early access when we have a usable version, sign up and you'll be on the list. We'll be honest about when it's ready to play, not when it's ready to demo.

The question of legitimacy

I want to address the elephant in the room. There's a strain of D&D community discourse that says AI in any DM role is illegitimate. That the social contract of the game requires a human at the table holding the world.

I sympathize with this view. I also think it's overstated.

Most criticisms of AI-DMing come from people who have a great DM. If you have a great DM, you don't need a Virtual DM, and you'd never replace your great DM with one. That's fine.

But not everyone has a great DM. A lot of people don't have any DM. They want to play, they can't find a group, they can't find anyone willing to run, and the alternative isn't "a great human DM" versus "an AI DM." It's "an AI DM" versus "no D&D at all."

For those people, even an imperfect Virtual DM expands the hobby. It gets more people playing. Some of those people, eventually, will become human DMs themselves and run games for others. The AI is a gateway, not a replacement.

For people in established groups with a human DM, AI-DMing isn't going to change anything. You'll still play with your group, your DM will still run, and the game will still feel like the game. This isn't a future where AI displaces the human DM at every table. It's a future where AI fills the gaps where there was no human DM available.

Both can be true. Don't let the moralism of D&D Twitter convince you that the choice is binary.

What I'd actually predict

Putting on my prediction hat, here's where I think this goes over the next three to five years.

Year one (now). AI DMs exist as toy products and experimental tools. Friends & Fables and similar apps are usable for casual play. None are good enough to seriously substitute for a human DM in a real campaign. Hobbyists experiment, mostly for solo play and one-shots.

Years two to three. Agentic AI matures. Tool use and multi-step planning become reliable. Campaign-aware retrieval becomes the standard. The first AI DMs that can run a coherent multi-session campaign appear. Early adopters use them for solo play and as training tools. Mainstream D&D groups still use human DMs.

Years three to five. Specialized models trained for DM-style tasks appear. Voice and multimodal interfaces let AI DMs participate in physical sessions. AI as "co-DM" becomes a standard feature in many campaign managers. Some groups run AI as their primary DM. Most groups still prefer human, but the gap closes.

Beyond five years. Hard to predict. The current AI capability curve has been steep. Could continue or plateau. Either way, the hobby itself doesn't go anywhere. D&D has survived TV, the internet, video games, and every other "this kills tabletop RPGs" doomsayer. AI is just another input.

The takeaway

A Virtual DM is a real thing, just not yet. Current LLMs are a flawed substitute for a skilled human DM, but they're a real substitute for nothing at all. The capability gap is shrinking on all sides, and the products that succeed will be the ones that integrate AI as a tool inside the existing structure of running a campaign, not the ones that pretend AI can do the whole job.

If you're a DM, you don't need to worry about being replaced. You should probably try AI as a prep tool (here's the workflow that works for me), because that's where AI is genuinely useful right now.

If you're a player without a DM, AI-DMing tools today are decent for casual play and getting better fast. Try Friends & Fables or wait six months and try whatever has launched by then.

If you're working on Virtual DM tools (and there are several teams), the killer feature isn't text quality. It's the integration with structured campaign data, the calibrated tone, and the willingness to ship something that's good for one specific use case rather than okay for everything.

That's where I think it goes. Not the death of the DM, not the salvation of the hobby. Just another tool in the kit, one that's going to get a lot better in the next few years.