AI dungeon masters have gone from "interesting demo" to "actually playable" in the last 18 months. The category is real, the tools are getting good, and there are now meaningful differences between them. This post is about how to choose.
We make one of these tools (TableForge), so this isn't a neutral review. But the framework below is honest, and we'd rather you pick the right tool for what you want than pick the wrong one and be disappointed in the whole category.
What an AI dungeon master needs to do well
Five things separate an AI DM that holds together from one that doesn't.
Rules accuracy. Combat, spells, conditions, action economy, dice. Does the system actually run D&D, or is it improvising rulings every turn? This matters more the longer your campaign goes. A loose ruling in session one is a contradiction in session twelve.
Persistent campaign memory. Does the AI remember the NPC you befriended three sessions ago? The promise you made in session two? The fact that the cult's ritual is tied to the next new moon? Without persistent memory, every session is effectively a one-shot wearing the costume of a campaign.
Multiplayer support. Can multiple humans actually play together with one AI DM? Some tools are strictly solo, which is fine if you want solo play but a dealbreaker if you want to play with friends.
Narrative quality. Is the prose evocative? Are the NPCs distinct? Does the world feel alive, or does everything sound the same? This is what LLMs are genuinely good at, and it's still the part that separates the merely functional tools from the ones you want to keep playing.
Pace and friction. Does a session move at the speed of actual play? Or does every decision involve waiting, re-prompting, or fighting the interface? An AI DM that's correct but slow doesn't work, because nobody runs long campaigns on a tool they dread opening.
The three categories, compared
| Generic LLM chatbots | Fiction sandboxes | Purpose-built AI tabletop | |
|---|---|---|---|
| Examples | ChatGPT, Claude | AI Dungeon and similar | TableForge and a few others |
| Ruleset | None; improvised | None or loose | D&D 5e via the SRD |
| Rules adjudication | LLM improvises | Whatever feels right | Programmatic engine |
| Dice | Generated text, not random | Not real | Real random number generation |
| Character sheet | You maintain by hand | Does not update itself | Updates automatically |
| Cross-session memory | Whatever you paste back in | Limited | Persistent and structured |
| Multiplayer | One person per chat | Mostly solo | Up to 6 players, sync or async |
| Best for | One-shots, prep, experimenting | Freeform fantasy fiction | Long D&D 5e campaigns |
Generic LLM chatbots as DMs
ChatGPT, Claude, and similar general-purpose LLMs can run a D&D session if you prompt them well. The narration is often genuinely good. The category breaks down on the other four criteria.
There's no persistent state, so every session starts from a context window stuffed with whatever notes you remember to paste in. There's no real dice roller, so you're trusting the AI to roll honestly (it doesn't, not because it cheats but because randomness from a language model isn't actually random). There's no rules engine, so combat eventually drifts. There's no multiplayer; one human, one chat.
Generic LLMs are fine for one-shots, creative warmups, and experimenting with what an AI DM could feel like. They are not a long-term campaign solution. There's a fuller breakdown in AI dungeon master vs ChatGPT.
Generative fiction sandboxes
AI Dungeon and similar tools are great at freeform collaborative fiction. You write what your character does, the AI writes what happens, you go back and forth. The stories can be wild and creative.
These aren't really D&D. There's no specific ruleset (or only a loose suggestion of one). The dice aren't real. The character sheet doesn't update itself. Combat is whatever the AI decides feels right in the moment.
If what you want is freeform fantasy storytelling, these tools are fun. If what you want is a D&D campaign, they're a different product, and we go through the distinction in detail in AI Dungeon alternatives for actual D&D 5e.
Purpose-built AI tabletop tools
A small group of tools, including TableForge, are built specifically for tabletop roleplaying. They typically have some combination of: a real rules engine, persistent campaign state, multiplayer support, and integration with a specific ruleset (usually D&D 5e via the SRD).
The tools in this category vary widely in how complete their rules engines are, how robust their memory systems are, and whether multiplayer actually works. This is where the meaningful comparison happens.
Where TableForge fits
We're honest about the rest of the category, so we'll be honest about ourselves.
TableForge is built specifically for D&D 5e using the SRD 2024 (CC-BY 4.0). The rules engine is programmatic code, not the AI. Dice, HP, spell slots, conditions, and combat resolution all run through deterministic logic. The AI handles narration; the engine handles adjudication. This is the architectural choice that lets long campaigns hold together.
Campaign memory is persistent across sessions. NPCs, locations, decisions, quest state, and consequences carry forward. You can play tonight, come back in a month, and the world is where you left it.
Multiplayer is first-class: solo, duo, and groups up to six are all supported. Real-time and async play both work.
The product is in open beta. The free tier covers your first session.
What's still hard
The honest part of any "best in category" post is the part where we say what doesn't work yet.
Long combat encounters across many sessions are still where AI DMs work hardest. Tracking initiative, conditions, lair actions, and reactions across a 12-round fight is the limit case, and every tool in the category (ours included) is still getting better at it.
Homebrew content support is limited industry-wide. Tools built on the SRD work with SRD content. If your campaign relies on third-party books, you'll be doing some manual integration with any AI DM. The same goes for uploading a PDF adventure or running an official published module, which nothing in the category does well yet.
Voice and table presence are different from a real human DM. A great human DM reading the room, riffing on a joke, pulling a player back into the scene, that's a kind of magic AI doesn't do yet. AI DMs are very good at what they do. They're not the same thing as the right human DM at the right table.
The category is improving fast. What was true six months ago about AI DMs isn't true today, and what's true today won't be true in six more months. Try the current generation. Don't assume you already know what it can do.
How to pick one
If you want freeform fantasy storytelling without specific rules: AI Dungeon or similar sandbox tools are fine. They do what they do well.
If you want to use ChatGPT or Claude directly as a DM: it works for one-shots and casual play. Don't try to run a long campaign this way.
If you want a real D&D 5e campaign that holds together across sessions, with persistent memory and an actual rules engine: you need a purpose-built tool. TableForge's AI dungeon master is one, and how it runs a campaign shows what a purpose-built tool does differently. There are a few others worth trying. Test the free tier of whatever you pick before you subscribe.
A good human DM running a great group is hard to beat. AI dungeon masters aren't trying to be that. They're for the times when finding or being that DM isn't possible, and you want to play anyway.