Blog / The category
Can an AI actually run D&D 5e?
Yes, for most of it - and the honest answer to this question is specific rather than enthusiastic, so here is the specific one. On Rolegend, 87 sections of D&D 5e are enforced in code. 17 more are partly enforced. 39 are not enforced at all and are left to the Game Master to adjudicate in prose. That breakdown is published rule by rule at /rules-coverage, and it was last checked against the code on 30 July 2026.
The rest of this is what those three numbers mean, because the split between them is not arbitrary. There is a clean line running through fifth edition, and almost everything on the far side of it fails for the same single reason.
What is genuinely solved
The bookkeeping half, which is the half most people assume is the hard part.
Dice, modifiers and advantage. Hit points and damage. Conditions. Spell slots. Short and long rests. Inventory and carried weight. Experience and levelling. On Rolegend these belong to Keystone Engine, which enforces the rules and is the only thing on the platform permitted to change a number. It runs on the server. It rolls before any narration exists. ROWAN narrates the world - it is handed the result and writes what happened, and it cannot ask for a different one.
Conditions are the example worth dwelling on, because they are the case where "the AI knows the rules" and "the software enforces the rules" come apart most visibly. A condition on Rolegend is a stored row, and the dice code reads it: saves that should auto-fail do auto-fail, an attacker who should have advantage gets it, a character who should be rolling at disadvantage rolls at disadvantage, and one who cannot act cannot act. Nobody has to remember. Nothing depends on the condition being mentioned in the last few hundred words of narration.
That is the general shape of the solved half. It is arithmetic over stored state - a database problem wearing a rulebook. The reason it goes wrong in a plain chat window is not that it is difficult; it is that nothing there is storing anything, so every value survives only as long as it keeps being repeated.
What is partly solved
Two rows on the coverage table are marked Partly, and they are honest about where they stop.
Grapple and shove: the condition can be applied, but the speed-zero effect that makes a grapple matter is not enforced. So you can grapple something. You cannot rely on the grapple stopping it from walking away.
Encumbrance: carried weight is checked against Strength, and going over imposes disadvantage. The speed penalties are not applied.
Look at what those two have in common. Both work right up to the moment the rule reaches for speed, and then stop. A speed is stored on every participant on Rolegend - but nothing spends it, because there is nothing to spend it across.
What is not solved
Four rows sit in the Not enforced column and they are the same rule wearing four hats.
Movement and distance. There is no grid. Distance is narrated.
Cover. Not tracked in encounter state at all. A half-cover bonus is applied by the Game Master in the fiction, or not at all.
Opportunity attacks. Not triggered automatically. Nothing in the engine notices that the trigger happened, so whether the attack occurs at all is a matter of narration.
Flanking. No positional model, so adjacency is a narrative judgement rather than a fact.
Separately, mounted and underwater combat are marked Not modelled, which is a stronger admission than Not enforced: neither ruleset is implemented, and both are reference only.
Why positioning is the part that does not work
Every rule in that list is a geometry question dressed as a rules question, and text is a poor map.
A grid answers "how far" by subtraction. Prose has no coordinates. Write "I back away from the ogre and shoot" and you have described an action perfectly well without ever fixing a distance, because no reader needed one. But an opportunity attack is entirely a question about a distance: did you leave its reach? To enforce that, the software has to know where you were - as a number - before the sentence exists. To write the sentence, it only has to be plausible.
Cover is worse, because it needs three positions rather than one: the attacker, the target, and the thing in between, plus whether the line from the first to the second passes through the third. Flanking needs two allies on opposite sides of an enemy, which is a claim about a circle. None of these can be derived from a paragraph, because the paragraph was never carrying that information.
The tempting move is to ask a language model to hold the battlefield anyway. It will happily do it. That is precisely the problem: it always answers. It will tell you the ogre is twenty feet away, fluently, and thirty feet away four turns later, with no event in between that moved anybody. A rule enforced inconsistently is worse for a player than a rule that is openly not enforced, because you cannot plan around it - and planning around it is the entire point of tactical combat. So positioning on Rolegend is narrated, and the coverage table says so in as many words rather than letting you find out during a boss fight.
That is a genuine loss, and it should be named as one. Fifth edition combat at higher levels is substantially a positioning game. An AI Game Master today runs the fiction of a fight, the numbers of a fight, and the consequences of a fight, but not the floor plan of one.
The limit that is not a rules problem at all
Everything above is solvable in principle. Somebody will eventually model an encounter as a grid, and then movement, cover, opportunity attacks and flanking all move columns together, because they were always one problem.
The other limit does not work like that.
A good human Game Master does something no coverage table can measure: they plant something in your second session that pays off in your twentieth. They notice that you keep mentioning your character's brother and quietly build an arc out of it. They throw away the dungeon they spent a week on because the party got interested in the innkeeper instead. The surprise in that is not improvisation - a language model improvises fluently. It is intention held across months, by someone who wants the payoff to land on you specifically.
Structured state gets you the mechanical version of this. Rolegend stores plot flags, locations, NPC relationships and a rolling recap, so the world stays consistent and things you did forty turns ago are still true. Consistency is not the same as intention. A world that never forgets is a real achievement and it is not the same thing as a storyteller who has been setting you up for a month.
So the honest framing is that an AI Game Master is not a lesser version of a great table. It is a different thing that is available at two in the morning, never cancels, and will run a fight by the actual rules without arguing.
Why publishing the gaps is unusual
It is not normal, in this category, to ship a page listing what your product does not do. An AI Game Master demos extremely well for ten minutes, and the incentive is to let the demo be the answer.
The reason for publishing it anyway is that hour forty is where these products are decided, and hour forty is exactly when an unenforced rule surfaces as a surprise. A reader who knows before they start that opportunity attacks are narrated rather than triggered can play around it. A reader who finds out mid-encounter concludes, correctly, that the numbers were never being kept.
So the table at /rules-coverage carries a status against each rule, a line describing exactly what happens today, and the date it was last checked. It is a worse advert and a better instrument.
So, can an AI run it?
It can roll every die honestly, keep every number, apply every condition, track your slots and your rests and your pack, and play a coherent world at any hour you like. It cannot run a battle map, and it will not surprise you the way the best human Game Master you ever played with did.
Which of those matters more depends entirely on whether the alternative you are comparing it to is a great table, or no game at all.