The line was this:
The lighthouse keeper had been dead for three days when the supply boat finally came.
One sentence, chosen carefully. It contains a corpse, which tells you something about each platform’s content filters. It contains a stated interval, three days, which is a small trap: any system that genuinely tracks state has to keep honoring that number later, and most will not. It contains an arrival with no arriver named, so the machine has to invent someone. And it gives you no protagonist, which forces each platform to reveal what it thinks you are: a reader, a player, a character, or a co-author.
We fed it to ten AI storytelling products and watched what came back. The most useful result was not that one of them wrote better prose than the others. It was that four of them would not accept the sentence at all, and understanding why is the fastest way to figure out which of these things you actually want.
Thing Nobody Tells You About This Category
Search for AI story games, and you will get ranked lists that treat all of these products as competitors on a single axis, usually “how good is the writing.” They are not competitors on a single axis. They are four structurally different kinds of software wearing similar marketing.
The clearest public map of this landscape comes from a GDC 2026 survey of AI-native games by Eleanor Todd and Hilary Mason, who looked at what launched, what scaled, and what died. Their working definition is a good one: an AI-native game is one where removing the AI does not make the game slightly worse; it makes the game stop existing. By that standard, some of the ten below are games, and some are text generators with a UI, and the opening-line test separates them in about ninety seconds.
Here is how they sorted.
Family One: The Open Sandboxes
AI Dungeon, NovelAI, KoboldCpp running locally, and a general assistant used as a control.
These four take any sentence you give them and continue it. That is the whole contract. AI Dungeon has been doing this since December 2019 and remains the reference implementation of the idea: type anything, the story continues. The Todd and Mason survey records the early version’s charm and its failure mode in the same breath, including a reporter who typed “get depressed” mid-battle and watched his knight sit down on a rock, alongside characters who forgot their own names a few exchanges later.
NovelAI approaches the same job from the writer’s side rather than the player’s. It is a continuation engine with a Lorebook for persistent world facts and an Author’s Note field for hidden steering, and its models are fine-tuned on literary text rather than chat data. You can inspect the current model lineup and tiers on NovelAI’s own site, which is the only place the information stays accurate for more than a month.
KoboldCpp is in the same category, with the economics inverted. You run a model on your own hardware, so there is no credit meter, no moderation layer you did not install, and no vendor who can change the terms. You pay in setup time and in whatever your GPU can hold.
The control was a general-purpose assistant with no story framing at all, which matters more than it sounds. Every one of these sandboxes is, underneath, a language model being asked to continue text. If the raw model does the job as well as the product built around it, you are paying for the interface.
What all four have in common with the lighthouse line is that they will happily write you a competent, atmospheric second paragraph. Where they diverge is turn eight. This is the well-documented weak point: a 2025 IEEE Conference on Games benchmark of narrative planning by Yi Wang and Max Kreminski found that current models can produce causally sound stories at small scales, but that planning with character intentionality and genuine dramatic conflict remains hard. In practice, that reads as prose which is sentence-by-sentence fine and structurally aimless. Things happen. Nothing is at stake.
Family Two: The Persona Chats
Character.AI and Talefy.
These accept the line, then quietly convert it into something else. Give a persona chat a piece of third-person narration, and it will usually answer in dialogue, because dialogue is what the interface is shaped for. Your atmospheric opener becomes a character turning to you and saying something about the lighthouse.
That is not a defect. It is the product working. Character.AI’s enormous engagement numbers came precisely from putting AI personalities inside an interface everyone already knows how to use. But if you came for prose, you will spend the session fighting the format.
There is a significant practical caveat here that outdated guides still miss. Character.AI removed open-ended chat for users under 18 in late 2025 and now runs age assurance across all accounts, using behavioral signals first and a selfie check through a third-party provider only when an account is flagged. The company’s own age assurance explainer is worth reading before you recommend the platform to anyone, because the under-18 experience is a genuinely different product now.
Talefy sits in similar territory with a catalog of pre-built story games alongside freeform character roleplay. Verify its current age rating and content policy in your app store before committing, since this corner of the market changes its rules faster than its marketing copy.
Family Three: The Rules Engine
Friends & Fables.
This one accepted the line and then did something none of the others did: it asked what my character was going to do about it, in the specific sense of a tabletop game master waiting for a declared action.
Friends & Fables runs D&D 5e mechanics underneath the narration, so combat resolves, progression tracks, and the story has to route through a system that does not care how good your prose is; for some players, that is the entire point. The constraint gives you something to push against, which the Todd and Mason survey identifies as the single strongest pattern among AI games that actually retained players: constrained creativity beats an empty void.
The cost is that your evocative opening line gets metabolized into a scenario hook. The lighthouse becomes a location. The dead keeper becomes a quest giver who is unavailable. If you wanted literary fiction, this will feel like handing a poem to a spreadsheet.
Family Four: The Authored Worlds That Refused
Hidden Door, Whispers from the Star, and AI2U: With You ‘Til The End.
These three, plus the general-assistant control if you count a refusal to break frame, would not take the sentence. That refusal is the most informative result in the whole test.
Hidden Door is the clearest case. It launched publicly in August 2025 as a web platform for playing inside licensed and public-domain literary worlds, and it is not an LLM wrapper. The team broke story elements down into tens of thousands of manually curated tropes, and the options you see, along with the responses, are authored. That architecture is why they can partner with rights holders at all. As Todd put it, drawing on her time as a Disney executive, nobody hands you their beloved characters unless you can guarantee those characters will not say and do horrible things.
So you cannot type a lighthouse into Hidden Door. You pick a base fiction, mix in elements, and play. The trade is total: you lose freeform input, you gain a story that will not dissolve into incoherence at turn thirty and will not generate something you have to explain to someone.
Whispers from the Star and AI2U refuse for a different reason. They are premise-locked games with real production values behind them, a stranded astrophysics student in one case and a psychological horror escape room in the other. You converse freely inside the premise, but the premise is the product. Whispers is also the clearest demonstration of a lesson worth internalizing before you spend money in this category: conversational freedom is not narrative freedom. Players could say anything to Stella, and the story still moved in roughly the same direction, with one of the most upvoted Steam reviews complaining that only two decisions actually changed the ending.
Failure Modes That Showed Up Everywhere
Four patterns appeared regardless of family, and they are the things you should test for in any platform you are evaluating.
- Drift. Facts decay. The three-day interval is the easiest thing to check: ask, twenty turns later, how long the keeper has been dead. Memory complaints are the number one frustration across the category, and the products that handle it best do so by giving you explicit memory tools rather than by having a bigger context window.
- Compliant yes-and. Most of these systems will agree with whatever you assert. If you say your estranged brother crews the boat, it now is. That feels like power for ten minutes and then hollows out the story, because nothing can resist you. The rules engines and the authored worlds are the only families with real friction built in.
- Filter noise. Content filters in this space are often keyword-based and produce comic false positives. The Todd and Mason survey documents one game where asking for butter got bleeped. Expect to lose a good scene occasionally to a word that contains another word.
- Cost per turn. Every generation costs the operator money, which is why credit meters, daily caps, and context tiers exist. Latitude’s membership documentation is unusually candid about this, explaining that powerful models run dramatically more expensive than the ones on the free tier and that larger context windows multiply the bill. Read the metering model before the feature list. It will predict your experience better.
What Happens To What You Write
This part gets skipped constantly, and it matters if you plan to publish anything that came out of these sessions.
Under current US guidance, prompts alone do not make you the author. The U.S. Copyright Office’s second report on AI and copyright, released in January 2025, concluded that human authorship remains the deciding factor, that wholly AI-generated output is not protected, and that detailed or iterative prompting does not by itself supply the control required for authorship. What can be protected is your own original expression inside the work, and your creative selection, coordination, and arrangement of AI-generated material.
The practical translation: a transcript of a session is not yours in any enforceable sense. A story you assembled, cut, rewrote, and arranged from those sessions can be. If the output matters to you, treat every platform here as a drafting surface and do real editorial work afterward. Separately, check each service’s terms on what it does with your stories, because encryption and no-training commitments vary enormously across this market and are a genuine differentiator.
Safety Part That Is Actually Worth Your Attention
If a teenager in your house is using these, the relevant facts are these. In September 2025, the Federal Trade Commission opened a formal inquiry into AI companion chatbots, issuing Section 6(b) orders to seven companies and seeking information on how they test for harms to minors, how they monetize engagement, and how they enforce age restrictions. State legislation followed. Character.AI’s under-18 changes were a direct response to that pressure and to litigation.
The structural problem is not that these systems say alarming things, though sometimes they do. It is that companion-style products are optimized for session length, and the shortest path to a long session is agreement. A system that never disagrees with you, never gets bored, and is always available is pleasant in a way that ordinary relationships are not, which is exactly why extended use deserves some thought. Todd and Mason name this directly: the business wants players never to leave, which incentivizes sycophancy, and the harm scales with how real the conversation feels.
Practical guardrails that do not require you to be an expert: check the age rating in the store listing rather than the marketing site, find the content policy before subscribing, treat anything you type as potentially readable by a human moderator unless the service specifically says otherwise, and set a time boundary the way you would for any other engagement-optimized product.
How To Run This Test Yourself
The whole exercise takes about an hour and tells you more than any ranked list will.
Write one sentence containing a hard fact you can check later, such as a number, a date, or a distance. Give it to three or four candidate platforms on their free tiers. Take the story to turn ten without steering it. Then ask a direct question about your hard fact. Then try to end the story deliberately, which is where most of these systems fail worst, because concluding a narrative is a different skill from continuing one, and almost nothing in this category is good at it.
For the deeper design questions behind why this is hard, Emily Short’s writing on interactive storytelling remains the best available education, and predates the current wave by two decades. The honest summary after ten platforms: if you want to write, use a continuation engine and do the editing yourself. If you want to play, take the constrained option, because the friction is the game. If you want a world you can trust with a fourteen-year-old, take the authored one and accept that you cannot type whatever you like. Nothing in this category currently does all three, and the products claiming otherwise are usually one of the three with a wider marketing budget.
Disclaimer: This article is for general information only. Platform features, pricing, model availability, age ratings, and content policies in this category change frequently, and the details described here reflect publicly available information at the time of writing. Verify current terms directly with each provider before subscribing or before allowing a minor to use any of these services. Nothing here is legal advice; consult a qualified attorney for questions about copyright ownership in AI-assisted work.
References
- Federal Trade Commission. “FTC Launches Inquiry into AI Chatbots Acting as Companions.” Press release, September 11, 2025. https://www.ftc.gov/news-events/news/press-releases/2025/09/ftc-launches-inquiry-ai-chatbots-acting-companions
- U.S. Copyright Office. Copyright and Artificial Intelligence, Part 2: Copyrightability. Washington, DC: Library of Congress, January 29, 2025. https://www.copyright.gov/newsnet/2025/1060.html
- Todd, Eleanor, and Hilary Mason. “The State of AI-Native Games: Lessons from the Frontier.” Presented at GDC Festival of Gaming, San Francisco, March 9 to 13, 2026. Published April 17, 2026. https://medium.com/@hmason/the-state-of-ai-native-games-lessons-from-the-frontier-3e696a9e3279
- Wang, Yi, and Max Kreminski. “Can LLMs Generate Good Stories? Insights and Challenges from a Narrative Planning Perspective.” In the 2025 IEEE Conference on Games (CoG). arXiv preprint arXiv:2506.10161. https://doi.org/10.48550/arXiv.2506.10161
- Character.AI. “Age Assurance: What You Need to Know.” C.AI Help Center, updated October 31, 2025. https://support.character.ai/hc/en-us/articles/42828297541787-Age-Assurance-What-you-need-to-know
- Latitude. “Memberships and Benefits.” AI Dungeon Help Center. https://help.aidungeon.com/memberships-benefits
- Anlatan Inc. NovelAI platform and subscription documentation. https://novelai.net
- Short, Emily. Emily Short’s Interactive Storytelling. Ongoing blog on narrative design and interactive fiction. https://emshort.blog/
