AI fan wikis riddled with errors are muddying the waters of games culture (www.eurogamer.net)

72 points by rdmuser •

Fandom (ex-Wikia) has been a dumpster fire since day one.

As witness the choice to squat on the name "Fandom" when they were notoriously averse to treating fandoms reasonably.

fan-Dom

Wait until you have to audit your average codebase post-AI.

AI is rotting pretty much everything that was carefully curated over the last couple of decades. If it can be expressed in bits it will be degraded. This is one of the most annoying side effects of AI, the general decrease in quality. Quality of discourse, quality in literature, quality of information, quality of software. It's everywhere.

I’m not sure I’d call most enterprise code repositories, carefully curated. I think lava flow is way more standard https://en.wikipedia.org/wiki/Lava_flow_(programming)

There are at least different kinds of messes, and the new kind may offer different challenges which may eventually require different (easier? harder? invented-yet?) techniques than the ones we're used-to.

The problem I'm already having with LLMs (in miniature) is code/comments/docs where I don't know whether it's someone's intent or whether its un-reviewed LLM babble. That creates a continual stumbling block for figuring out whether the thing should be the way it is, or whether I'm looking at an unusually-coherent error.

Take the exact same messy pile of bytes but promise: "Everything in this made sense to a human at some point in the process", and that changes how one approaches the problem.

Before code would be intentionally a certain way, done as a known compromise that's less than ideal, or the developer flubbed in some way. Using commit comments and other surrounding details you might be able to identify which in many cases which could guide your next steps. I don't think you can do that with an LLM's output.

Post-AI?

So this is why we are seeing all these CVE's and vulnerability disclosures? Because of the careful curation? Feels to me it is more like ignorance is bliss.

Those bugs (assuming validity) were there in spite of the careful curation. People tend to do what they can to the best of their abilities. Before compilers did much in terms of static checking a lot more errors made it into production than today. Better tools -> better code.

Whether AI written code is going to be better in the longer term remains to be seen. With what I've looked at so far I'm not super optimistic about this.

LLMs reliably avoid whole categories of bugs, some of them the most serious offenders in terms of impact.

To me it has become clear that LLMs will, on average, improve the code quality by a lot.

It was just 5 minutes ago that I prevented a Claude PR from introducing an incredibly obvious RCE bug into a codebase I maintain. The LLMs are no better than the people driving them. But those people spew out more code than ever before.

Better than what? 99% of stuff produced by humans is a garbage. Arguing about "abilities" or compilers does not change that.

Even here we see the AI bit rot - we're getting CVE's and vulnerability disclosures for imaginary issues.

I know the steel made before this particular year is look after because was not affected or build during periods of radiation, and that gives it qualities sought after.

When AI started I got a copy of 6 months before of wikipedia and open street map.

At some point you cannot trust anything because everything has been touched by everything as a regurgitation of everything.

In case anyone isn't familiar with that type of steel: https://en.wikipedia.org/wiki/Low-background_steel

Similarly, radioactive signatures are used to identify modern forgeries of classic oil paintings!

Are you willing to share that? I didn't have the space at the time to download the entirety of Wikipedia (though I think I have an old copy of Wikipedia sans images floating around somewhere).

I believe someone created a website specifically for that purpose with regards to LLMs: https://lowbackgroundsteel.ai/ (not mine, nor can I remember who shared it on HN)

And of course they had to use high-background steel to build this site.

You know it’s not gonna be updated, because it was created on a whim and the author has already moved onto 79 other projects since.

So what's the solution? I imagine the right one isn't specific to fan wikis. We're going to need sites where 100% of contributed content is human-written. Who is building this? How will it work?

If the solution merely makes it high friction to post AI content that is probably good enough. In other words if someone has to copy/paste the output of an LLM into a text field and even then the tech detects that and throws an error that would be a big step in the right direction.

Reddit, Wikipedia, fan wikis, reviews, Hacker News comments - I'm sure there are a dozen more categories I'm missing.

Every high friction system that doesn't involve manual approval line-by-line is vulnerable to LLM circumvention and most of the ones that do are still also vulnerable to it, as long as the users of the end product are financially viable to prey upon. Also, if a game feels like work, most players will go find a different game, leaving only a narrow playerbase possessing certain personality traits. That kind of narrow playerbase has no reason to ever worry about a shortage of human wiki curation.

We can put these together to conclude that every game and community that lacks that kind of dedication will die out and be replaced by psychologically personalized short-form dopamine-agonist videos, as exemplified in an infamous comic:

https://x.com/Merryweatherey/status/1395152040389713922

FYI, that tweet is not visible without an X account. You may want to consider an alternative link or at least a description of what's about.

replace x. with tw1tter.

That comic is considered sensitive? By whom? The AI overlords?

god that comic hits so hard. what a horrible world we are building.

You can still hitch a ride on a pearling lugger to the horizontal falls and shoot them with a homemade sea kayak in real life .. and as yet there's no VR / AR plugin for that in the digital world - at least not one with real whirlpool stoppers, salt water crocodiles, and warm equatorial water sharks.

That link prompts for a signup.

Scientific articles: https://retractionwatch.com/2025/02/10/vegetative-electron-m... (although this is an old one)

> So what's the solution?

There is no solution (though there will be plenty of grifters trying to sell one).

The open internet we currently know will be lost to mountains of slop just like usenet and email were even prior to the AI era.

None? This is a fundamental change in the ecology, with the introduction of a new invasive species. Like with code, there is no going back, and no way to review everything that is written.

The unspoken assumptions that underpinned the web included the idea that it was made by humans and that content generation couldn’t be mechanized.

There are already YouTube ads for services that Humanize AI output.

The best analogy is that we are at the start of the Dark Forest era of the net.

I assume people will adapt, based on the things that don’t change. If you cannot make out authenticity of participants, then you index on correctness of argument. If the output is useful to you, then you got what you needed and can move on. A more transactional and less community oriented web.

I don't think there is a solution. The internet as we knew it is over, I give it 5 more years before it becomes borderline impossible to find new human-created content on it.

Its really just the web that has this problem. The internet is OK, and we can build a new web for it.

>a new web for it.

Incentives will make sure that this new web will become the old one given enough time.

Instant information was always a mistake. I hope we will go back to libraries. Time has come to preserve the paper books.

I can make something and produce a digital signature for it. AI can't fake the signature. We meet in real life and exchange keys. You pass my key to your browser so when it encounters content with my signature it marks it as having come from somebody whose humanity you have verified. We build a network of such relationships. Maybe it's enough to just see an annotation, a sort of glow around the content that has verified human signatures. Or maybe we go full dark-forest and configure the browsers to only render content that has a trusted signature.

The problem is that the majority of the time, the one ultimately trying to dupe you is not the AI model, but the person who handles it.

In other words, signatures don't address the fact that a human can ask an AI model for content and then sign it to pass it as their own.

[bell ring] What is PGP?

I think we're back to IRL interaction. Things I could ask my friends about are unfindable on the internet. Meanwhile travel is much more accessible than before the internet, so "but the person I want to talk to lives in another country" isn't as much of a deal as it used to be.

Case in point: the proliferation of furcons.

Make the use of AI expensive. This is the same problem we have with email

Unfortunately it's worse than the problem with email, and it's also adding to the problems of email, but making it very cheap to send infinite unique and persuasive messages.

I've seen a couple of these in search results while looking for info about more niche games and at first I thought they were fan made but like the article mentions they're not actually wikis, nor do they contain accurate information.

The ones I came across just contained nebulous prose with variously incorrect info about the game on a few barebones pages, with a custom domain name. They weren't using any wiki software.

I have uBlock Origin installed so not sure if the purpose was ads but I was baffled as to what the goal would be otherwise since they're useless for proper information, let alone a wiki.

I think there are people who just assume there's money to be made if you're pushing the sloppified tomorrow hard, and no time to stop and think about why.

Maybe there are other people selling the tools, who try to convince the former people that what they're doing is smart and profit will roll in any day now.

Anyway it would be a mistake to think something is profitable just because someone bets on it.

On the other hand, frontier LLMs are extremely good at getting super accurate info from games. They’ll happily disassemble the actual game binary to get to the bottom of certain game mechanics, including ones not obvious in the game itself, without even an explicit prompt to do so; I can merely ask questions about certain mechanics. That’s way better than often guessed (and guessed wrong) mechanics commonly found in game wikis. I’ve observed this over and over again when vibe coding some helper tools for some games I’m playing; I used to build these for wider communities and got some recognition for them, now I can do what I used to finish in 100 hours in like half an hour of prompting. I no longer share them though.

It’s the shit tier free or super low cost AI hallucination engines used by the masses that are the problem.

I fully agree! I have done a few game decomps and some deep RE passes on other games, and that uncovers a lot of details that game wikis/discords/forums get wrong or just don't mention, going as deep as undocumented mechanics that change the online gameplay. It's only a matter of having the LLMs properly reverse engineer the games instead of relying on other players' info, lossy wikis, or game manuals.

I’m building a shed at the moment, and it’s become basically impossible to find reliable advice on tools or techniques by searching the web. It’s all just slop feeding back on itself now, unless you are lucky enough to find a forum post from before, say, 2023.

For example, try searching about what sort of nails you should be using to attach cedar lap siding.

On Google, you might find a few good results, but you’ll really need to scrutinize them to make sure they’re reliable. Most are slop on the ‘blogs’ of siding and roofing contractors.

On DDG, it’s even worse. Usually nailerguy.com ranks highly - it seems to be completely AI generated, complete with contradictions and impossible pictures. And, often, nailicy.com, a (seriously) nail salon in New Hampshire that I’m sure is confused about why all their web traffic isn’t translating into clients.

YouTube is about the only ‘reliable’ source for recent advice.

Thank you for mentioning DDG - I don’t know what it is about their search algos but whenever I tried switching I found I got far less helpful results compared to Google. Like at least half the first results were total slop fest unusable.

DDG has become so unusable because of AI word vomit sites that I'm finally seriously considering the paid search engine that HN loves... Kagi, I think?

> ... I put the question into Google and clicked the first result: the Nighthawks Wiki.

The problem here is Google. Bad websites have always been around.

LLMs have created a new problem. Previously it was a lot of effort to create say a bad wiki with bad information that looked plausible.

In this new era it is almost zero effort to stand up a convincing website on any topic, which is mostly correct, but riddled with small confabulations and errors. It can take typing one sentence (even with grammatical errors) to do so, the bar is now zero for creating plausible 'content' and flooding the internet with it. As the article notes bad actors, but also people who have good intentions, are flooding the zone with slop.

If it was all clearly completely incorrect, it wouldn't be such a huge problem, but it is plausible and hard for humans to detect in a short time unless it is long form text.

Even search providers are giving us this new kind of generated result at the top of their results, which most people will be using blindly to verify answers. So it's turtles all the way down.

If you're not a domain expert you have no idea which things are right and which are wrong so it's very hard to determine the truth or trust results.

To top all that, LLMs are being trained on this new untrustworthy data without differentiating it from older data, and trained by other LLM instances in a sort of recursive loop to improve performance on very narrow goals. What could possibly go wrong!