The AI Dungeon Saga

Heard about an ai driven text adventure game recently, decided to try it out, learned it was a glorified chatbot/story writing tool, then heard about the controversy.

play.aidungeon.io...

The creators of Ai dungeon are a company called Latitude.

In early April, shortly after receiving a large amount of investor money, it was revealed that there was a massive data leak and users personal stories were publicly available.

github.com...

In this data leak it was found people were writing stories with #ed up content. Latitude secretly released a regex filter to censor content out then a week or so later released a statement about it essentially blaming their users.

latitude.io...

Fast forward a month later, the company's made no response about the data leak and has doubled down on their filter.

Some enterprising folks scoured the training data used on the ai for the game.

Turns out according to the report available here:

gitgud.io...

The ai used in the game was intentionally trained using pedophilic, racist, homophobic and otherwise depraved and #ed up material causing the ai to frequently generate material that would trigger the filter even when users were trying to play the game normally.

At some point the company thought it would be a great idea to hire moderators for people's personal stories on a site called taskup.ai. Shortly after that a user on 4Chan began leaking users personal stories on a thread on 4Chan.

Stories available at a link here(nsfw like real #ed up #)

and

Here(same warning as above)

The company proceeded to double down and remove their job posting from taskup.ai and ban users for ai generated content.

Not much to say other than, heard about this and thought it should be shared around.

The behaviour by Latitude is astounding to me from the things i've read about the situation. There's more to the story, if you're interested and can handle reddit, the aidungeon subreddit has the whole saga there and is full of pissed off people. Anyway...sadly, it's not actually an ai generated Zork or some such a thing, but all the controversy is pretty funny.

So a company made a pedo-weeb ai for degenerates to beat off too and got bad that people noticed....sounds par for the course for reddit.


a reply to: dug88

a reply to: Athetos

Yup, the interesting part is how they managed to make millions of dollars doing this, rip off customers, break laws, do some pretty questionable things and are still continuing to do so.

Who would've guessed people would use text generating AI to create messed up sex stories lol... seriously though it's quite unfair to blame this company for people using their product like this because their AI is based on GPT-3 which was already trained using millions of web pages and comments on the internet, and I'm sure a great deal of questionable content made it into the GPT-3 training data. They fine-tuned the GPT-3 model using data which allowed it to produce interactive stories, and some of those stories are obviously going to include dark themes like rape and murder. I remember playing AI Dungeon a few months ago and it worked very well for me, I was extremely impressed.

There will come a time within the next 10 years when anyone can have access to an AI smarter than GPT-3, I'm sure some of those models will even be specifically trained using adult content. Even the openly available GPT-2 is capable of generating adult stories without any extra fine-tuning (I may or may not have tried this). We are only seeing the start of these social issues arise now, but there's very little anyone can do about it, and I see little reason to take a moral high ground on the issue like that gitgud page author. He/she is unfairly attacking the company behind the game as if they consciously tried to make the AI sexist and racist, etc, when that is highly unlikely to be the case.

a reply to: ChaoticOrder

Latitude specifically trained their ai using their own data separate from openai. They also get preferential treatment from openai. This is not an unfair attack against the company. I suggest you read the report on their training data.

a reply to: dug88

They did not train their AI from scratch, they fine-tuned the existing GPT-3 model using a separate dataset so that it specialized in producing interactive stories. I skimmed the report on the gitgud page, and as I said it's more of an emotionally charged rant than anything else. A lot of the examples provided are typical scenes of violence and crime which can be found in many stories and other forms of media, when you have thousands or millions of books in your dataset you can't check it all, and I would argue that these dark themes actually help the AI gain a deeper understanding about human society so it's not necessarily something we want to filter out of the data if we want a more empathetic and intelligent AI. Just as overprotecting a child from all the dangers in the world can cause them to be naive of those dangers and more prone to fall victim to them.

a reply to: ChaoticOrder

I would agree with you, but I actually tried to use this to create a 'real' story.. Because I actually write and publish books was having some mental blocks creativity wise.

No matter what you says, it kinda forces sex on you. I tried to used it to write a standard adventure that started out in the woods with a girl and her dog, bam gang raped, beaten and murdered in the woods by 4 men.(You can 'fast forward scenes by just pressing enter - then its like a new chapter in a book). But it jumped to waking up in a new room (as a new character) with police knocking on the door, and parents questioning what happened to the 'sister'.

Fairly certain, if I was into those types of books over fantasy adventure. I could of done something with it..(That is how detailed these things get) But I prefer working with drow/fae/dwarf/elf/human; Just easier to write, and cop drama's leave me bored.

While I am sure a lot of people lead that to their 'sexual desires'. Not everyone does, and it is HEAVILY predispositioned towards sex, rape and violence.

Either way I stopped using it and went back to generic scenario generators, it showed promise, except its NOT AI. It literally looks for keywords and openings/similarities then pulls forward what other people write.. its actually super disturbing.

Tried 3 times to start a native girl in the woods with her dog; hunting. Every time there was some form of 'stalker'. It is in fact pretty messed up.

Its not uncommon for artist to be inspired or use generated scenarios/tarot/tools to help with writing, but this is literally something else.

a reply to: ChaoticOrder

You are spot on, I think the narrative on the attempt I tried was literally taken out of that movie/book (Last house on the left? or something cabin). Where one of the kids is dad is an escaped convict.. and they wind up going to the girls parents cabin, rape her shoot her, etc.. I am 100% certain your correct that it pulls other writings from books. This is kinda obvious from anyone who tries it and pays attention.

The downside is you quickly learn that 99.9% of its copywrite. So it becomes completely useless as a generated 'ai' for creativity blocks. Maybe in another 10-20years it will be an awesome tool in the aritst arsenal. But right now its pretty much garbage.

a reply to: BlackArrow

Tried 3 times to start a native girl in the woods with her dog; hunting. Every time there was some form of 'stalker'. It is in fact pretty messed up.

This scenario probably has something to do with why you got those results. These AI rely heavily on context, and I guess it expects most girls who are alone in the woods to get into some sort of trouble. I used it to produce several stories and none of them were sexual.

The downside is you quickly learn that 99.9% of its copywrite. So it becomes completely useless as a generated 'ai' for creativity blocks.

This isn't exactly true, these AI have some randomness when choosing the next word in a sentence and the degree of randomness can easily be changed. I've used AI to generate a fairly original short story as well as two unique essays. I used GPT-2 (the predecessor to GPT-3) to generate the text and I'd estimate it's about 90% original.

Censorship and Free Speech on the Internet [Written by AI]
Does AI pose an existential threat to humanity? [Written by AI]
The Weakest Great Elder [Written by AI]