Can we start over? Ferrytale's use of ClubFloyd transcripts.

Technical Development
1

TL;DR I’m sorry for launching Ferrytale in a way that upset some authors within the IF community. I’m open to significant changes moving forward if game transcript opt-out doesn’t suffice. I think I can make it right and will continue to focus on what I care most about - immersion and accessibility in IF through AI-enhancement, for authors and players who want it. Hopefully I can do that from inside the community instead of as an outsider.

If you choose to reply, please first read my post in full.

My name’s Alex Keybl - my background’s in software engineering and immersive mediums like VR. I’m used to moving fast and breaking things for the sake of progress, but I also want to fix mistakes when they happen.

On Friday I released github dot com / akeybl / ferrytale v0.1.0 (MIT License) with a mostly-unseen post on socials and a Discord message on an IF discord server I found on this forum (in the #ai-chat channel).

Ferrytale is an open source prototype meant to make Interactive Fiction more immersive for existing IF enjoyers, and more accessible for players who have never enjoyed IF. It uses IF transcripts (specifically from ClubFloyd) to allow a player to do whatever they want within the constraints of an existing story - both actions and dialog. It remains grounded in the author’s original writing and story arc, of course with whatever current limitations the underlying LLM (Gemini Flash) and voice models have. It’s also unlikely to know any story paths that weren’t explored by ClubFloyd, unless trained into the LLM from other transcripts/walkthroughs.

I’m an outsider to the IF community, but not to the medium. While I’m not an author, over the years I’ve really enjoyed a few games and I’ve always wanted to improve upon what I see are the biggest barriers to accessibility and immersion:

  • text input/output: no solution for folks who enjoy audiobooks more than books or don’t want to type a lot to enjoy IF
  • slow pacing: repetitious outputs and feeling like you’re playing “find the verb/noun”
  • feeling lost: not knowing what to do next without reading a walkthrough full of spoilers

Yesterday I was banned from the above Discord because of the general concept of AI-enhanced IF and how I chose to launch it. That’s unfortunate, but I’d still like to have the opportunity to explain my thinking, and have further discussion with authors about how to implement a resolution.

My mistakes were twofold - not giving authors a heads up before release and assuming that almost all authors who were OK with ClubFloyd transcripts of their games being available online wouldn’t mind seeing them included in a for-fun project. I think I misjudged the response of some authors and over-focused on the whimsy of the project. I saw it as a love letter in the same way a VR mod for Zelda BotW is a love letter and attempts to expand on the original game’s functionality.

My intent was never malicious, to “steal” authors work, make money off of their writing, depress sales of their work, compete with them, or disregard their opinions about their works. My intent was to explore immersion, accessibility, and create joy. I thought that the project being open source, not for profit, and allowing authors to request a transcript be taken down (same as how ClubFloyd does it) would cover my bases for such a small project. As promised I’ve already removed 3 transcripts from the catalog at the request of those authors.

What I’m still grappling with is whether to blow away the transcript catalog links entirely and move to opt-in instead of opt-out. It’ll add manual steps (copy/pasting) for playing works whose authors don’t say one way or the other, but I’m OK with that if it’s actually the majority opinion amongst authors I can get in touch with. It was not my intent to upset a segment of the community - just to make it possible for players to try the prototype.

I’m planning to move forward with growing the concept as a whole, for new IF collaborations that will be hosted and accessible to all. To be clear, these ClubFloyd transcripts will never be available in a fleshed out “product” like a widely available hosted site or app regardless of how things play out. The Ferrytale CLI is built to be bring your own computer, API keys, and $.

Basically I’m sorry, and I’d like to start over with the community if possible. I personally believe that there’s tons of uninitiated kids and adults who will enjoy the medium for the first time if somebody succeeds in AI-enhanced voice-driven IF, and I’d love to continue that conversation in another thread. The fact that I’m an outsider offers me a different perspective but is also a reason the launch upset some authors.

2

Remove all transcripts for I Am Prey.

Keep GET CORN if you want; it’s a funny poison to have in your dataset.

Move to opt-in.

You can start by moving to opt-in.

4 Likes
3

Remove all transcripts for I Am Prey.

Done. I’ll continue to monitor this thread for opt-out requests.

You can start by moving to opt-in.

Thanks yeah. I know the response of most people who reached out to me on Monday were against. I’m just gauging if that’s the case for most or the few that have reached out.

Having worked in open source previously, I know that there can be a vocal minority (which opt-out would serve). If it’s a vocal majority opt-in is the obvious answer.

4

Is there a handy list of games that have transcripts included? Without people having to sign up.

In any case, if you have included my Napier’s Cache - which was played through by Club Floyd - please remove it.

I am opposed to AI’s use in IF, for multiple reasons, including ethical and environmental. And do not give permission for my game content to be used in this way.

I also recommend that you consider removing all transcripts without permission. This should be opt in, not opt out.

Thanks.

6 Likes
5

I don’t know what happened on Discord, but to quote another post of mine, this forum has recently seen a big increase in posts from people who are only here to self-promote their custom AI-written games or systems but don’t participate in the larger community. There was also a terrible AI game that got submitted to last year’s IFComp and gave a lot of people a bad taste in their mouths when it comes to this kind of thing. Starting off by promoting an AI-based system will give many a negative impression.

You seem well-intentioned. That said, AI is extremely controversial on this forum and in this community, which you can see by browsing the ai tag. A lot of people in this community are inherently against AI. Some have seen it having a debilitating impact on the creative industry, which includes their friends. Also, a common argument against AI is that it amounts to plagiarism, so I imagine a lot of people really didn’t want transcripts of their games, involving writing they made, to be fed into an AI system, even if it’s open source.

I would say the anti-AI contingent is a large part of the community - see the IFComp polls on AI for evidence - and so moving to opt-in is the only way to not have a lot of people here strongly oppose you. Even then, a lot of forum members avoid AI stuff out of principle.

In general, though people with a technical background or career are often pro-AI or AI-neutral nowadays, which includes a lot of open source people, many other people with a background in the arts are very much against it, and I’d say this forum leans much closer to the second group than the first.

To add onto this, I think if you came up with an audio harness (something like an audio interpreter) based on the actual game files themselves that doesn’t use AI to read or transform the text, there may be more interest. What’s the point of basing it on a transcript if it can only see exactly what happens in the transcript? Can you really say the result is similar to the original game?

5 Likes
6

Is there a handy list of games that have transcripts included? Without people having to sign up.

I unfortunately can’t post direct links with this new account. github dot com / akeybl / ferrytale / blob / main / catalog / clubfloyd.json

In any case, if you have included my Napier’s Cache - which was played through by Club Floyd - please remove it.

I did and it is now removed.

I am opposed to AI’s use in IF, for multiple reasons, including ethical and environmental.

I appreciate your thinking here. It’s clear a subset of creators and creatives share your perspective, and I’m trying to understand the split. I’m curious to know if there’s any possibility for a future where AI isn’t attempting to replace but rather truly augment, and your environmental concerns are addressed. I’m admittedly a bit of a techno-optimist so I personally see a path forward.

7

Thanks!

Here’s the link to the list of all the games for other folks who may want to check.

Did you get permission from the ClubFloyd folks to use this transcript collection in this way? That’s an additional step in addition to clearing it with authors.

2 Likes
8

And you also have my Bad Beer in there I see. Please remove that too. Thanks.

3 Likes
9

You should make this opt-in. Full stop.

I shouldn’t have to ask.

5 Likes
10

I would say the anti-AI contingent is a large part of the community - see the IFComp polls on AI for evidence - and so moving to opt-in is the only way to not have a lot of people here strongly oppose you. Even then, a lot of forum members avoid AI stuff out of principle.

Thank you - this survey is indeed eye opening and does seem to serve as consensus in the community on the subject. Based on that irrefutable data I’ve decided to move to opt-in. I appreciate all of your level-headed responses.

To add onto this, I think if you came up with an audio harness (something like an audio interpreter) based on the actual game files themselves that doesn’t use AI to read or transform the text, there may be more interest. What’s the point of basing it on a transcript if it can only see exactly what happens in the transcript? Can you really say the result is similar to the original game?

Yes, I’ve considered this previously. Would still require some sort of “known good” inputs to translate to (likely extracted from transcripts) but of course it’s possible to make output be as the author intended. It won’t fix all of the issues (repetitious outputs as one of them) but it would definitely make IF easier to play. Worth considering.

11

Forgot to address this - similar yes, complete no. IF that is purposely created for AI would require thoughtful dumping of game logic into a human readable format. Totally doable given large/cheap AI context windows in 2026, and obviously superior to this incomplete version. It’s why I originally called it an “interpreter” in the README and specifically called out the deficiency in the first couple of paragraphs.

12

IF is an extremely unprofitable niche, so in this community you’ll find a lot of people who prioritize creativity and artistry over profit and are sinking many hours into passion projects instead of trying to make money off it. There’s a general perception that the AI companies want to steal all that work, grind it up, and sell it back to us to line their own pockets.

Not while large language models are mainly controlled by large for-profit corporations. That’s just the nature of capitalism. If a resource exists (and to the AI companies, all creative work is a resource), there’s an incentive for someone to seize control of it and start charging for access.

This one seems more plausible, but I don’t expect it to happen in the near future.

I’m also a techno-optimist, but I don’t think people in this community specifically are interested in automating away creative writing. We’re not, as a whole, doing this because we see writing as a necessary evil or a chore that needs to be optimized; we’re doing this because writing is what we want to be doing. So any proposal that goes against that is unlikely to be popular.

I want technology that does the dishes so that I can spend my time writing, not technology that does the writing so I can spend my time on the dishes.

The difference is that a VR mod isn’t replacing the original graphics made by the artists, it’s presenting those original graphics in a new way. I’d say a VR mod is closer to a text-to-speech and dictation system, and I’m wholeheartedly in favor of that for IF! What I’m not in favor of is having an AI do the writing for me.

I want people to enjoy my IF, but I want them to enjoy my IF—the words I wrote, not the words a LLM thinks I should have written.

12 Likes
13

Thank you for reconsidering. I’ve reversed your ban from the discord server in question as a reciprocal act of good faith. That obviously does not obligate you to rejoin the server if you no longer have any interest.

4 Likes
14

For many game formats it’s possible to automatically extract the “dictionary”—the complete set of words the game can understand. It shouldn’t be hard to derive a plausible set of commands from that. A big part of parser IF is deliberate combinatorial explosion (if “take” works on one small object, it probably works on every small object, and so does “drop”, “put”, and so on) so the inputs from a transcript represent a small fraction of the things a player might want to try.

2 Likes
15

For many game formats it’s possible to automatically extract the “dictionary”—the complete set of words the game can understand. It shouldn’t be hard to derive a plausible set of commands from that.

I’ll keep thinking about this. I see glulx-strings exists for the purpose of dumping strings from files within a gblorb at least. It seems folks are comfortable with input translation (and voice input).

I’m wondering where people land on voice output by a single voice, and voice output with character voices. Ferrytale currently uses full transcripts to design appropriate voices for each character, with no other previous knowledge of the games.

I’m also curious where people land on the concept of transcript/walkthrough usage for the purposes of out-of-story guidance without spoilers when requested by the player.

16

My personal take is that I’m not interested in AI being used to replace humans for creative purposes—text processing is okay, text generation is not. Deciding which character is speaking each line is not, for me, the kind of creativity I’m concerned about. (It’s also a task that can often be handled by a local model, which helps with the environmental issues.)

Nor is trying to match someone’s voice to a valid command. Dictation software is a wonderful thing for accessibility and it’s been using neural networks since long before “AI” became a buzzword. My concerns are specifically with generating new game text.

Nor is trying to decide which part of a transcript is relevant to a player’s current situation, for that matter—but I’m skeptical on how well an LLM can recognize what’s a spoiler and what isn’t.

2 Likes
17

Thank you. That puts this conversation on a better foundation.

That is only true if you’re focused on current-industry LLMs. (“If a multibillion-dollar company wants to sell you hammers, everything looks like a nail.”) There must be an interesting domain of machine-readable formats, and ways to operate on them. They don’t look like ChatGPT though.

On the flip side of an entirely different coin: a great deal of game design is based around hard limits. That is, a lot of gameplay is about finding out what you can’t do. Generative AI is bad at hard boundaries; at best it does probabilistic boundaries.

This means obvious failures for, say, puzzle-games. If you get past a gate in chapter 1 without finding the key from chapter 4, that’s not “immersion”; that’s called “sequence breaking” and you’ve thoroughly bunged up the plot. But it’s also true in a more general sense. Character and setting are very often described through constraint – or through omission. I know AI text tools have gotten good at some stuff, but “figure out what this transcript doesn’t have and be careful not to accidentally add it” is a tall order.

8 Likes
18

I see the issue has been settled already but count me as one more vote for opt-in.

It’s not just this community that is opposed to AI-made creative works. If you read around, it’s the prevailing public opinion in all creative areas. People generally do not want to see AI’s being used to generate creative works. The mainstream AAA gaming community in particular has been very vocal and unkind to game companies that admit to using AI.

This particular line of thought precisely encapsulates the greatest offense of AI-generated works: that they steal and regurgitate existing works. It sounds like you’ve created a tool that copies an author’s work from an intermediate source – a transcript that is already, in a sense, a step removed from the original work – and then outputs a third generation copy. The intermediate work isn’t alive and responsive like the original game was - it’s just a static recording of a piece of interactive art, like a recorded performance. And then your tool takes that static recording and makes a bunch of inferences to make a new different interactive thing that looks like and claims to be like the original thing. You call this making the author’s work more accessible. I call it theft and perversion of the author’s work and frankly I find it deeply offensive.

The fact that you felt that it was ok to do this without speaking to any of the original creators mirrors the move-fast-break-things mindset of the techno-libertarians who unleashed LLMs on society in the first place. And I say this as someone who uses LLMs regularly as a development tool. But I see a bright sharp hard line between development and repurposing creative works.

8 Likes
19

Even if this didn’t involve the perpetually divisive subject of AI, familiarity with any open source community should make it clear that requiring opt-out is almost never acceptable.

6 Likes
20

Thank you for removing the links to the transcripts.

I’m one of the first people who interacted with you on Discord. As part of that conversation, you kept bringing up how this is “progress”, how it’s “inevitable”, and important for accessibility, how IF is in decline, mentioning poor sales of IF games on Steam, and this is what will help bring many more players and financial success to the medium.

Those are a lot of assumptions about what’s important to authors and players, and the capabilities and potential of your tool. It’s also a condescending attitude, similar to the “have fun staying poor” attitude of cryptocurrency “bros”. As a self-proclaimed outsider, is it surprising that the community isn’t receptive to what you’re offering?

What you made is an experimental toy. You’re exploring if something is “interesting”. Don’t jump to imagined future scenarios like “playing interactive fiction while commuting” or “allowing people not familiar with IF to play them”.

Suppose you gave a human who’s a creative professional writer/actor a transcript of a game, like you do with the chatbot, asked them to read it carefully, internalize it, and then interact with you, based only on that transcript. Do you think that’s an interesting/useful/amusing/novel thing to have instead of actually playing the game? I doubt it. But if you think so, maybe work on improving that before adding text-so-speech or hooking it up to all the transcripts you can find online.

And if you think IF has faults you can help with, maybe work with the community on the individual problems, instead of thinking a single prompt to a chatbot is the solution to everything.

5 Likes