The current system seems fine? The games are clearly labeled by whether and how their authors made use of generative AI and the large majority of games this year made no use of generative AI at all.
Reviewers who object to generative AI on ideological, environmental, or ethical grounds can abstain from playing or reviewing those games.
Reviewers who believe in good faith that a game that used generative AI is of low quality can rate it poorly.
I would not want a competition where most entries are LLM-written slop; or where outsiders manipulate the vote for the sole sake of boosting public perception of generative AI; etc. but as far as I can tell these are hypothetical concerns?
I’ve admittedly got my fool’s cap on as I write this, but since IFComp already bans games that infringe copyright, and since no LLM can be guaranteed not to have infringed copyright in its selection of training data, AI games are already banned.
I gather the actual legal situation is complicated, but I do find it odd that I’ll get emails from IFComp confirming that I’ve got the rights to a stock photo used in a cover image, but “Actually, I generated that from a bunch of obviously stolen training data” would apparently be an acceptable response.
A quick look at KL3M suggests that it’s probably not going to be a whole lot of good for IF specifically. It’s good to see someone out there is making the effort not to simply grab stuff off the internet willy-nilly, but unless you can get fantastic results by adding your own writing to the mix I don’t know if this really supports the idea that IFComp entrants currently have an ethical option.
I’m in the unusual position of having 250,000 words of my own work all packaged up in a TXT file ready to go (I set up a Markov chain text generator a while ago), and I wouldn’t know where to begin trying to make meaningful use of something like this in an IFComp submission. That’s partly because this really isn’t an area I want to sink my time into, but also KL3M seems a long way from the straightforward API option offered by the “screw the rules, I have money” crowd.
Versu not being available any more for quite some time means that others here may not be readily aware of this but it was not an LLM platform or even to my knowledge used ML or generative networks at all; it was a traditional rules-based agent behavior system, ‘video game AI’ in the conventional sense (I feel bad for all my colleagues who now have to append “not that kind of AI” to every discussion of their work).
My point is that when Versu and other types of automation were being developed, not one person found it unethical or flat out wrong.
So when we talk about bans, we can’t just say AI because that does have a wider context.
I was reluctant to bring up Versu because we don’t (I think) know what tools were used, but we know there was a long standing goal to develop automated conversations based on context. In today’s environment, this community seems like it would react differently to these kinds of efforts.
When peeple say “AI” in this conversation they mean generative AI and LLMs. Versu did not use these, nor do most video games, which use “video game AI”, a longstanding concept since the 80s. There’s a difference between creators writing procedural generation or rules-based behavior for NPCs and using LLM black boxes.
This is the lamest rhetorical trick I can think of. C’mon.
Versu is from 2013 and definitely isn’t using generative AI. It’s nothing like chatGPT and its cousins so it’s not really relevant to this discussion. And we do know what tools it uses, btw:
But even without that, one very big difference is that it requires a human to actively create content! It never generated text or images for you–it only habdled how characters delivered your content to readers:
This is exactly the problem people have been referring to upthrrad about the collapsing of the term “AI” to refer exclusively to large language models. It’s really confused decades of earlier uses of the term.
If you genuinely do not understand why people are opposed to LLMs but do not mind these older things, you might check out wikipedia as it has a good summary of the ethical issues.
That is absolutely not what everyone is saying. It’s not a rhetorical trick. There is some level of automation we’re comfortable with and I’m curious why a potential ethically trained model would still break “the rules”.
I mean, manually implementing a bunch of different rules that can interact in potentially-surprising ways is just programming; that feels obviously and intuitively different from describing a result and having something else implement the rules in potentially-surprising ways. I suppose there could be edge cases where the difference is blurred but I think specific examples pointing to the concern would be needed to understand why this is something we should be worried about (even granting the premise of a hypothetical ethical LLM).
This feels backwards to me. I can see the posts above where people have said things along the lines of “There’s no objection to translation or grammar-checking even when they’re advertised as ‘AI’ - we’re talking specifically about LLMs” but haven’t seen people explicitly discussing “automation” as a concern.
Admittedly I think there is a concern over authorship, but that doesn’t strike me as being quite the same. People happily using electric mixers in a baking competition might reasonably take issue with entrants submitting shoplifted cakes even if both things save some effort. Insisting that pre-made cakes be bought rather than stolen wouldn’t necessarily resolve the problem.
It’s more than that. It also matters who is presenting the game. We instinctively trust some people because we believe they did due diligence and are presenting something fair and meaningful.
This debate (and yes I’m poking the edges) is much more challenging than simple thoughts about GenAI.
Now what if we describe the Versu system to ChatGPT and build a simple logical system with NLP? We’re still delivering a pure non GenAI system but we augmented our capabilities with it.
But I think, like Bruno says, there’s a pretty significant difference between using some automation to assist your creation process and generating content from whole cloth with limited input.
I see a lot of people elsewhere point out things like cameras or photoshop as “advances in automation that artists didn’t like but use now,” but to me those arguments miss a lot of points.
This article by Cory Doctorow makes great points that I think are relevant here:
ETA: my least favorite internet word (debate) has now entered the chat, so I’m going to mark this thread as muted. I’m interested in providing resources and educating people to make informed decisions–not scoring points in a debate. I encourage people to fill out the official survey form if they want to see the change that this thread’s title suggests. (And would reiterate what @zarf said above about some of the reasons why monitoring these threads is NOT a mechanism for IFComp changes!)
The following is addressed at no one in particular and is just an observation:
Obtaining AI responses through prompts is no more creative than entering a seed for a Minecraft world. You didn’t create the result, you just seeded the algorithm that did.
Note: This isn’t meant to compare LLM’s to clever procedural generation either.
/s