AI/LLMs Murder Hobos are out, Paladins and Bards are in, according to the data

[GregH_AAI]

I keep saying if someone had an LLM trained on public domain works from Gutenberg plus a bunch of context from Creative Commons/OGL/copyleft game supplements, I'd be all for it.

disclaimer I’m one of the co-founders I love this idea. I think a lot of people would be more supportive of LLM usage if it could be trained in purely public domain material. The challenge is that this would be possible through fine-tuning, but you’d still require a base model that was trained on all the stuff that people take issue with. There just isn’t enough of the data you’ve described to pre-train a sufficiently-capable foundation model I hope we can find a solution like this though. It would be a massive step toward truly ethical models!

 

[SableWyvern]

LOL. To be fair though—I've been using em dashes for a long time. I'm kind of miffed that it's now considered a tell of AI generated text. Maybe AI has spent too much time scraping my blog.

I use the British method -- like this -- and I've never once had an AI accusation, even when they're automatically converted to en dashed by a word processor. Something to consider, perhaps?

 

[GregH_AAI]

Both are very powerful in the 2014 phb.

However, this "data" tells us nothing and required LLM workflow to gather that...nothing, which means it is actively bad data.

Just to clarify one point, an LLM wasn't asked to look at a bunch of campaigns and tell us what it thought was happening. These are aggregate database statistics from actual recorded campaigns.

Things like session length come directly from timestamps, system comes from the campaign's system setting, class/alignment are fields recorded on character records, and talk share is calculated from the words attributed to each speaker. We also published the methodology and limitations alongside the numbers.

Archivist does use AI as part of the product that creates some of those campaign records, but just wanted to be clear that “an LLM generated these conclusions” isn't how this analysis was done.

 

[GregH_AAI]

Meaningless word salad clickbait is annoying.

I say that paladins and bards are welcome in any civilised murder hobo club, while random arson is not and has never been normal for any alignment.

Edit: aka, I'm implying what @doctorbadwolf said directly.

I can understand your take on the headline. It's intentionally playful.

That said, the article is just us publishing aggregate data from over 25k actual TTRPG sessions because there isn't much large-scale observational data about how people are actually playing.

Also agree that playing a Paladin or Bard doesn't somehow make someone incapable of being a murder hobo. Those are separate observations in the article. I don't think the language implies a causal relationship but if it was interpreted that way that certainly wasn't the intent.

If the headline got in the way of that, that's fair criticism. Hopefully the actual data is still interesting. Since the raw numbers are there we also wouldn't mind hearing other interpretations

 

[GregH_AAI]

1) Murderhoboism has nothing to do with the alignment on the character sheet.

2) Alignment as an indication of behavior is very poorly corelated. I find in general that people justify all kinds of terrible behavior anyway.

UPDATE: Reading more of the original post more closely, it says all kinds of things that the data doesn't support. What in the world does "found family storytelling" have to do with anything, and why would a 2.5% uptick in paladins and bards registered in DNDBeyond lead to that conclusion? Not impressed. Made a strange conclusion, and as evidence of it, showed data that had nothing whatsoever to do with the conclusion. A much more compelling conclusion would be that Chris Pine and Sam whatshisface's character from Vox Machina have led to the popularity of the bard, and the mechanical changes to the paladin led to its uptick. How in the world you've concluded that murderhoboism is extinct based on that is... opaque, at best.

That's still just a hypothesis, not proven by data, but it's a much better null hypothesis then the one you've proposed, which is completely unrelated to the data that you've shown. Yours isn't a hypothesis supported by the data, as you claim, it's a non sequitur just so story, and you've thrown unrelated data at it and pretended like it somehow supported the just so story.

I think we're actually more aligned (pun absolutely intended) than the headline makes it seem. You're right that alignment on a character sheet isn't a direct measurement of player behavior. The article deliberately calls this section “Evil Alignment Is Rare On The Sheet” exactly because of that. What we can say is that among D&D PCs with alignment recorded, 58.7% are Good, 1.6% are Evil, and 0.3% are Chaotic Evil. That's interesting observational data, but it doesn't mean we watched every player and classified whether they behaved like a murder hobo. Same thing with Paladins. The full analysis actually says the 5e/2024 class distributions are statistically indistinguishable (p=.81), and explicitly says the Paladin movement is a signal worth continuing to measure, “not yet a finding.”

The “found family,” actual-play influence, mechanical changes, etc. are possible explanations, not conclusions drawn from the data. We don't collect demographic data that could answer the “why,” which is why we didn't pick one.

So I think some of what you're objecting to is fair criticism of how the headline summarized it, but it's considerably more qualified in the actual analysis.

 

[GregH_AAI]

@Wold Newton Welcome to the boards. Jumping in with AI. You should get a lot of.... responses.

I would go with the idea that paladins and bards get more goodies in 5.5e and that people are taking the things that give them power.

Haha yeah... the spirit behind it was that we just had a meaningful amount of data that we could anonymize/generalize and share it with people. Actual stats on large (not saying ours is massive, but it's meaningful) groups of players is pretty rare and we just thought fellow nerds would think it was interesting

 

[One Man [RPG] <studio>]

That conclusion depends much more on sampling and labeling than on the raw class counts. Before reading it as a cultural trend, I’d want to know which tables entered the dataset, how behavior was classified, and how many campaigns do not treat alignment as an important descriptor at all.

Real Campaigns Are Overwhelmingly Good also does not by itself establish that the murder-hobo stereotype is false. Character-sheet alignment and play behavior are different variables: a non-evil character can still solve every problem through opportunistic violence. The stronger test would be to code session behaviors such as violence against non-hostile NPCs, purposeless looting, deliberate destruction of social ties, or a consistent preference for violent solutions, then compare those with class and alignment. That would actually measure the myth the headline claims to test.