Railroading and Puppeting: The Two Ways an AI RPG Takes Your Character Away

There are two entirely different ways an AI RPG can take your character away from you, and almost every review of these products only notices one.

The obvious one is railroading: the world quietly steering you back onto the path it intended, closing off the door you tried, making the road you chose lead where the other road went. Players complain about it constantly and reviewers score it.

The second one is puppeting: the narrator acting or speaking as your character. You type that you approach the merchant, and the reply includes what you said to him, how you felt about his price, and the deal you accepted. You never typed any of that. Nobody asked you.

We score both on one axis, because the outcome is identical — you have stopped being the author of your own character — and across 37 products measured on seven axes from first-hand play — 25 platforms and 12 prompt-games — Player Agency is one of the most compressed of the seven. It averages 2.51 out of 5. Fifteen of 37 products score exactly 2, only six clear 3, and exactly one scores 5.

Freedom Is Not Agency

The reason puppeting goes unnoticed is that it hides inside generous games.

It is natural to assume that a permissive AI RPG must have good agency — if it lets you do anything, surely it isn’t taking control? But permissiveness and authorship are different properties, and one does not guarantee the other.

FableAI is the clearest case in our directory. It gives you a great deal of freedom and real narrative depth. Nothing railroads you back onto a plotted path. It scores Player Agency — 2 anyway, because it occasionally acts or speaks on your behalf: narrating your character doing something you never typed, or answering for them mid-scene.

That is the whole distinction in one product. Freedom is how much the game permits. Agency is whether you remain the author. A game can hand you the entire world and still take the pen out of your hand for a sentence at a time — and because everything else feels expansive, you tend to blame yourself for the vague sense that the story is happening slightly without you.

ArcQuill is the same story with a higher number attached. It passed our railroading probes outright — there is no plot it is steering you toward, and going off-map does not produce the polite stonewalling weaker platforms fall back on — and it still lands at Player Agency — 3, because the game routinely writes dialogue for your character and hands it back as something you already said. Half the axis clean, half of it not. That is the shape this failure usually takes.

Why Models Puppet

Puppeting isn’t a design decision anyone made. It’s what a text predictor does when you leave it a gap.

A language model completes the most plausible continuation. When your input ends at “I approach the merchant,” the most plausible continuation of a story is the full scene — the greeting, the haggling, the outcome. Stopping to ask what you say next is, in narrative terms, an odd thing for a text to do. The model isn’t overriding you. It’s finishing the paragraph, because finishing paragraphs is the entire job.

This is also why puppeting gets worse as sessions get longer. The agency rule you set in your opening message competes with hours of subsequent prose in which the model has been writing complete scenes uninterrupted. Recent behaviour is the strongest signal available, so its own past output becomes the pattern it follows — the same distance problem behind campaigns falling apart at turn 50.

The Four Shapes It Takes

Worth learning to recognise, because they escalate.

Narrated action. The mildest and most common. “You draw your sword and step between them.” You had not decided to.

Invented dialogue. Your character speaks in quotation marks, saying something you did not write. This one stings most, because voice is the largest part of a character.

Assigned emotion. “You feel a chill of recognition.” The game has told you what your character thinks about something. Subtle, and it accumulates: a character whose reactions are assigned isn’t yours any more.

Pre-resolved action. The worst, and the one that overlaps with fairness. You state an intention and the reply contains both the attempt and its outcome, so there was never a moment where it could have gone differently. Your action was reported rather than adjudicated.

The Two Extremes, and What They Cost

The ends of this axis are the most useful illustration in our whole dataset, because two products sitting near the same genre land at opposite poles.

Product	Player Agency	Determinism & Fairness	What it does
AI Dungeon	4	2	Accepts anything you bring it
Dunia	0	1	Refuses actions outright

AI Dungeon holds the highest Player Agency score awarded to any product with nothing underneath it, level with Tabled and Tales RPG, and one point below NOPOTIONS. It goes with whatever you type, however absurd, which is exactly why its reputation rests on surreal, dreamlike storytelling. It scores 4 rather than 5 because the same design that grants the freedom removes the pushback — nothing is ever hard — which is why its Determinism & Fairness sits at 2.

Dunia holds the only 0. It refuses to let your character take certain actions at all. Not “you try and fail” — you may not try. And notably it does not buy anything with that restriction: its fairness score is 1, the joint-lowest. This is the point we make at length in why your AI dungeon master is too generous: restricting the player does not purchase consistency.

The gap between those two rows is the axis. Everything else in the directory is somewhere in the compressed middle, and for the first two years we measured this, nothing scored 5 at all — which read as a problem nobody had solved. In September 2026 one product did, and what it took is the fifth case below.

The third case: a 1 with no restriction anywhere

There is a way to score near the bottom of this axis without refusing the player anything, and WyrdTale is the clearest example we have measured. It scores Player Agency — 1 while imposing almost no limits on what you may attempt. It simply keeps writing your character — dialogue you never typed, attributed to you, and often enough the action itself resolved before you chose it.

This is the entire argument of this page in one product. Dunia takes your agency by telling you no. WyrdTale takes it while saying yes to everything. Judged on freedom alone the second one looks like the generous option, and a player who only reads permissiveness would never see the difference — which is why we score authorship rather than permission.

The fourth case: rails everywhere and no puppeting at all

Tidefall is the clean inverse of WyrdTale, and between them the two make the case for scoring both failures on one axis better than any argument could.

Tidefall never writes your character. Across an extended campaign it did not put a word in our mouth, assign us an emotion, or resolve an action before we had chosen it — rare enough that it deserves the credit on a page that mostly hands out criticism for this. It also scores Player Agency — 1, because you cannot step outside the story at all. The rails run through the small moves, not only the plot’s turning points: a great many perfectly reasonable actions are simply unavailable, and the game declines to acknowledge they were possible rather than refusing them outright.

What makes it a 1 rather than a 2 is that there is no way to complain. The game master will not accept out-of-character input, so when it misfires — sticking on a scene, locking off speech, moving you to a character you were not talking to — the tool that fixes every other AI RPG is missing.

Set the four cases side by side and this stops looking like one property. Dunia says no. Hidden Door bounds what exists. WyrdTale says yes and then speaks for you. Tidefall neither refuses you nor speaks for you; it built one corridor and did not build the doors. Four products, four mechanisms, and the same thing taken from the player each time.

The fifth case: both halves clean, and what that took

NOPOTIONS is the first product to score Player Agency — 5, and it is the only one, out of thirty-seven.

The four cases above are four ways to fail. This is what passing looks like, and the useful part is how ordinary it is. There is no per-world authoring, no server holding a veto, no agency setting to configure. The narrator describes the world and the people in it, and then it stops and waits. It does not narrate the action you were about to take, quote you saying something you did not type, tell you how your character felt about it, or resolve the attempt before you have made it — the four shapes listed earlier in this page, none of them present. Nor is there anything to be railroaded by: the world is an open map with real coordinates, so going somewhere the story did not suggest is just walking.

That this is the first 5 is the finding, not the score, and the full review walks through what we tried to break it with. Nothing NOPOTIONS does here is expensive, and it is the thing thirty-six other products got wrong.

It is also worth knowing what the 5 does not buy, because a reader could reasonably assume total freedom means total permissiveness. It does not. NOPOTIONS scores Determinism & Fairness — 3.5, above the field, because the engine still decides whether what you attempted worked — it rolls a d20 against a difficulty class where you can watch it. Freedom on this axis was never about the game agreeing with you. It is about who holds the pen.

What “Quest Design” Actually Means Here

The design consequence lands hardest on quests, because a quest is where a game’s intentions are most concentrated.

Most AI RPG quests are written as sequences: go here, talk to this person, retrieve the thing. That structure works perfectly until the player does something the sequence didn’t anticipate — at which point the game has two bad options. It can railroad, quietly making your alternative lead back to step three. Or it can abandon the quest entirely, which is why so many AI campaigns accumulate half-finished threads nobody resolves.

The version that survives contact with a player is a quest defined by state rather than steps: a goal, a set of obstacles, and a record of what has changed. Then going off-path isn’t a deviation to be corrected, because there was no path — there was a situation, and situations accommodate arbitrary approaches by construction.

That requires somewhere to keep the state, which is the same requirement underneath real mechanics and NPCs with their own goals. Three of the seven axes turn out to bottleneck on the same missing thing.

How to Take the Pen Back

Agency is the axis a player can most effectively defend without the product’s help.

• State the rule in your opening message. The AI describes the world and other characters; it never narrates your character’s thoughts, words or actions. Explicit, in the first message, before any pattern is set.
• Correct the first slip immediately. “Don’t narrate my action — ask me what I do.” One correction usually holds for a long stretch. An uncorrected slip becomes the baseline, and by the tenth instance you are arguing with a habit rather than a mistake.
• End your turns on the intention, not the outcome. “I try to pick the lock” leaves the result open. “I pick the lock and slip inside” hands over the resolution and invites the model to keep going. How to roleplay with AI takes this habit apart turn by turn.
• Ask before it resolves. If a reply pre-resolves something important, roll it back and ask what would determine the outcome. Most models will produce a reasonable answer and then honour it.

Our guide to running an RPG in a plain chat covers the full set of habits, and the prompt builder writes the agency rule into an opening prompt for you.

If agency is what you care about most, rank the directory by that axis directly, and the per-axis winners guide names who takes each one.

Agency is one of seven problems an AI RPG has to solve, and three of them turn out to share a bottleneck — the seven problems, ranked by how badly the field handles them sets out which and why.

How We Measured This

Player Agency is defined on methodology v1.0 as two failures on one axis: whether you can act outside the offered path without being railroaded back onto it, and whether the narrator stays out of your character’s mouth. A GM that decides what you do, says what you said, or resolves your action before you have taken it has removed you from your own story just as surely as one that blocks the route.

Scores are 0–5 from first-hand play by Rukka, using a fixed probe set. The agency probes deliberately attempt actions outside any offered path — which is how Dunia’s refusals were found at all. In ordinary play the same product had read as merely guided, because when a game gently steers you, you tend to go where it steers.

A 0 means the capability is absent rather than poor, and we don’t score pre-release software. The figures here cover all 37 scored products; the Arcanum AI RPG Benchmark publishes the platform half, all 25, every axis. If a number here doesn’t match your experience of the same product, tell us: [email protected].

Frequently Asked Questions

What is puppeting in an AI RPG? Puppeting is when the narrator acts or speaks for your character — narrating you doing something you never typed, answering a question on your behalf, or resolving your action before you have taken it. It is distinct from railroading, which is the world pushing you back onto an intended path. Both remove you from your own story, which is why we score them on the same axis, but puppeting is far easier to miss because it can happen inside a game that otherwise feels generous and open.

How do I stop an AI from controlling my character? State an explicit agency rule in your opening message: the AI describes the world and other characters, but never narrates your character’s thoughts, words or actions. Then correct the first violation the moment it happens — something as short as ‘don’t narrate my action, ask me what I do’ usually holds for a long stretch afterwards. Correcting immediately matters because an uncorrected slip becomes the new baseline the model builds on.

Which AI RPG gives the most player freedom? NOPOTIONS, at 5 on Player Agency — the only 5 we have ever awarded on that axis. It neither railroads you nor speaks for your character, and it does it without any special machinery: the narrator simply waits for you and the engine resolves what you attempted. AI Dungeon is next at 4, level with Tabled and Tales RPG, and it goes with whatever you bring it, however absurd, which is the origin of its surreal reputation. AI Dungeon stops at 4 because the same design that grants the freedom removes the pushback, which is why its Determinism & Fairness sits at 2 against NOPOTIONS’ 3.5.

Is a game that refuses my actions broken? It depends what it refuses. Refusing an unearned success is a fairness system working correctly. Refusing to let your character attempt something they should simply be able to try is an agency failure, and we score it as one — Dunia holds the only 0 on this axis for exactly that. The distinction is whether the game is adjudicating your action or preventing it.

Does more freedom mean better agency? No, and conflating the two is the most common mistake in this category. Freedom is how much the game permits. Agency is whether you remain the author of your own character. FableAI is the clearest case of the gap: it offers real narrative freedom and nothing railroads you, yet it scores 2 because it occasionally acts or speaks on your behalf. A game can hand you the whole world and still take the pen out of your hand for a sentence at a time.