AI rule for Spring Thing: How to make a rule that is enforceable and fair?

Competitions Spring Thing
62

By the way, for complete transparency I have used AI before to help troubleshoot automated email sending with smtp on the Spring Thing website as well as setting up Node variables on the console when working on Bisquixe.

I’ve also used it to translate toki pona for the game anpa ma.

That’s why I haven’t included banning ai code or translation because it would be hypocritical to do so. My arguments here are based entirely on the aesthetic enjoyment of ai games.

I recognize that this against the general public position and it would be nice to hide it but I want all cards on the table.

8 Likes
63

Re: detector accuracy, this depends if you are willing to pay money. I have access to a paid AI detector that has a 1/10,000 false positive rate

The free ones are pretty bad.

Basically the issue is you can’t detect AI without running through an equivalent amount of AI, which costs money (the reason OpenAI is working right now is bleeding ridiculous amounts of money, they lose money for every subscriber, and there isn’t nearly as much speculative venture captial being dropped on stopping AI)

I think all this is likely moot here – it should be obvious in this use case if something has that kind of text, and as already pointed out, this is a case where community norms out to be enough.

2 Likes
64

Wow, what a spectrum of stances on this! I think the “don’t like it, don’t read it” argument, however, carries sort of an in-your-face tone. Any group is self-policing, even if only by shared mores/tastes. It appears a minority feel so strongly against any rules at all pertaining to AI. I suspect that games produced by people liberally using AI aren’t going to appeal to the majority. I don’t have any direct evidence for thinking that: just the fact that for me, it’s the human touch – the little glimmers of personality, foibles, and endearing eccentricities that give an int-fic game its soul and evoke appreciation in me.

Dunno if we are ever going to be taken over by AI Overlords in the future. If so, and if you’re reading this and I’m still around, please don’t send a terminator after me…

1 Like
65

I’ve said before that I’m (1) in favor of voluntary AI labelling and (2) against bans and mandatory labelling. That’s unpopular so I won’t go further.

Anyway … at this point, it’s clear that the IF community is overwhelmingly against AI content, aside from some pro-AI pockets. It’s time to stop asking whether that’s the case.

The community needs to figure out whether they want rules based on (a) convenience or (b) moral/ideological opposition to AI.

Various organizers need to do the same. Is a given policy based on (a) convenience, (b) moral/ideological opposition, or (c) audience pressure.

So an ideal policy would start with “Spring Thing holds that AI-generated content is …” rather than just saying that the policy is being set at the organizers’ discretion.

That puts more accountability on you, but that’s the point … if the accountability is unmanageable, we need a bigger, IFTF wide or community-wide approach to AI policy setting. (I know that Spring Thing is not part of IFTF like IF Comp is, but hopefully I’m not blurring the lines here too much.)

I’ve linked these before. Steam’s approach is a good one, Itch io’s is not. Not in what they allow or ban, but how they describe their reasons for acting.

2 Likes
66

50+ posts in a day, this is why I mute on sight debates on AI.

over and out,
dott. Piergiorgio.

4 Likes
67

As someone who uses—even admires—em dashes, this test has always frustrated me :laughing:

Bumping this point, since it seems to have gotten lost in the discussion.

7 Likes
68

This is a slightly different question. One thing Spring Thing has done recently is ask for games to be archivable (except for those with very large multimedia, which are usually hosted on itch.io). Live AI generation is kind of fragile, as APIs are evolving rapidly. While I’m obviously leaning towards banning AI in general (which some people have been asking for for years but I was waiting on to try to figure stuff out), if I did go the ‘labelling’ route, what are your thoughts on the archivability of such games? It’s not a hard hard line but I try to do the best I can so that future people can enjoy games of the past!

2 Likes
69

Maybe the use of “AI” could be defined more specifically, so it’s clear that LLM slop is the target. Something like “not to enter games that use text or art generated by prompting chatbot X, Y, Z, or similar services”.

1 Like
70

Not sure if this was meant in response to me since there was a comment in between …

One thing Spring Thing has done recently is ask for games to be archivable (except for those with very large multimedia, which are usually hosted on Itch io … if I did go the ‘labelling’ route, what are your thoughts on the archivability of such games?

If you went with labelling, it would be fair to have games with AI content over a certain file size exempt from archiving, or archived at your discretion, just like any other game.

But I guess this can be abused. If people want the ability to take down their work, they can deliberately submit big files knowing you won’t archive it.

Anyway, labeling is not going to satisfy people who want a ban.

Live AI generation is kind of fragile, as APIs are evolving rapidly

Are most IF games with AI content using live services? I didn’t get that impression, but yeah, that would make them un-archivable.

[Labelling] is a slightly different question [from banning].

Admittedly yes, but each is a policy for limiting content, each is exclusive to the others, and and they’re all on a spectrum of “light touch” to “nuclear option.”

I understand if you go ahead with a full ban based on the massive community pressure. But it would be reassuring to the few of us who are against restrictions in general if we knew the bigger picture (to whatever extent it can be disclosed or described).

I’m muting this thread because I end up going around in circles, but if you want input on the policy when it’s time, feel free to reach out.

71

I believe that was all a response to prevtenet’s suggestion that “live” AI queried during gameplay be allowed.

1 Like
72

If so, thoughts:

This is a great question. While there is some variation, most LLM providers now support some form of the v1 OpenAI API as the lowest common denominator. Services like OpenRouter are also popular, which automatically route OpenAI-compatible requests to whichever provider is currently cheapest. This means that, as long the game doesn’t do anything too unusual and provides some way for future curators to swap out the API URL and API key, it should be possible to keep it running even if the game’s original provider disappears. Archiving the full source code is the easiest way to accomplish this.

The bigger fragility problem, IMO, is “who pays for the API access?” Most IF contests are free-to-play (and should absolutely stay that way). Free API endpoints do exist, especially for Google’s models or via OpenRouter, but they tend to be more transient and have rate limits. LLMs are getting cheap enough that it’s becoming feasible for the author to pay for free access for the duration of a comp, but after that, what happens? In practice, experimental AI projects usually solve this by allowing the user to input their own API key, giving them the flexibility to use whatever credits/endpoints they want.

This is an ongoing problem, but IMO not something Spring Thing needs to be the final word on. You can just say “it needs to be archivable and free to play” and let authors figure out how to actually do that.

TL;DR Games using live LLMs create problems for the author, but (IMO) not for the comp, where they are already covered by existing rules. This is the exact opposite of static AI-generated text, which is easy for the author but creates problems for the comp that are not covered by existing rules.

73

Not really germane to the current drift of the conversation, but one point I’d like to bring up is, no one event has to be a venue for everything, and I feel that with solo-organizer events like this especially, the organizer is entitled to make any rules they want to make running the event more manageable and pleasant for themselves, as long as they can draw clear lines about what is or isn’t allowed/required and communicate them clearly. As long as that’s the case, the rules don’t need to be maximally inclusive or logically watertight or to reflect the norms of the community.

Like, the Short Games Showcase requires photosensitivity warnings. Required warnings of any kind are not a community norm, and it’s not strictly “fair” or consistent to require that one and no others. But it’s my event and I’m photosensitive and so I get to make that totally arbitrary rule because it makes my life easier. If the event really catches on and five years from now I hand it over to someone else, then it will be their event and they can get rid of that rule if they want (and make whatever other arbitrary rules they feel like making!), but this is what works for me right now.

All this to say, if AI art and text specifically are causing you as the Spring Thing organizer headaches, and other possible uses of AI (such as for coding) are not, I think it’s fine to ban the specific things that are causing problems for you, and it’s still fine even if it’s not strictly logically consistent and even if you don’t get majority agreement that it’s the best AI rule that there could be, and it doesn’t have to be Spring Thing-as-an-event’s final word on the topic forever, either. It can just be about what makes this an event that you feel OK about running for now.

36 Likes
74

They are perfectly archivable. If the author volunteers about $10 in his own API account, it is enough for hundreds of people to play the game for as long as they want, since every AI query costs a small fraction of a cent.

75

Playing interactive fiction requires a much higher mental investment and a longer time commitment than other types of games.
Right from the start, I begin to visualise the background and characters, and of course drawing a map (physical or in my mind). This comes before any type of enjoyment the game could give me.

This activity (which requires effort from the player) does not happen in — say — graphical adventure games, where everything is laid out in front of you.

When the upfront commitment is substantial and LLM slop is allowed, severing the good from the bad would require more and more energy, and in the end quickly drain my willingness to try new games.

The policy in OP is reasonable.

5 Likes
76

For as long as the API keeps working. Companies retire APIs eventually. And OpenAI and other AI companies are hemorrhaging cash. Eventually the investment money will run out. No doubt there will be some form of GenAI which is economically sustainable in the long term, but there’s no guarantee the current companies will last. They might not even be operating next year!

12 Likes
77

In the game I am working on (for which I am using some AI input), I am doing all those things; creating characters and their backgrounds, drawing an extensive map with useful object locations and creating a plot (this may become plots) with a number of subplots. I don’t think I have done any less work with this current project than I have done with previous ones, for which I did not use any AI assistance.

1 Like
78

Andy, my point was different (apologies if I was not clear).

In brief: trying IF takes time and effort, much more so that other types of games. I do not pass moral judgment on LLM usage, but I would be way less likely to play competition games if LLMs were allowed.

4 Likes
79

(post deleted by author)

80

I’ve not participated in any competitions (I just like playing games and tinkering with the z-machine) so my opinion means pretty much squat, but here it is anyway.

There seem to be two issues of concern: Size of AI games with art, and overall text quality.

The first seems relatively straightforward - put a hard cap on submission size because of archiving constraints, no need to make it AI specific.

The second is a lot more difficult and will be seen as wrong by someone no matter what choice is made.

AI-slop has become a very real and growing problem across many fields, including security research and bug bounties. It’s become enough of a problem that AI assisted or generated bug reports are being banned from some open source projects due to overwhelming the capability of teams to investigate and handle the volume. That seems an apt analog to the situation here. Personally I’d ban AI generated text while making clear the intent:
We are not going to do deep inspections looking for AI, this is not an ideological standpoint, but a practical one to maintain our capability to thoroughly evaluate all games. If you manage to slip an AI based game through then good for you, but be warned that usage of AI generated text can result in a game being pulled from competition and possible ban from future competitions (three strikes rule?).

If someone is super-committed to AI generation, then maybe start a separate competition.

One further thing: I strongly disagree that live-service usage isn’t a problem for archival. It most definitely is. The surest way to make sure something dies an early death is to make it depend on something somewhere else on the internet.

16 Likes
81

I prefer that my posts not be used to generate LLM content.

21 Likes