#sudowrite

you might have seen people talking about sudowrite and/or their tool storyengine recently

and just like… don’t. don’t do it. don’t try it out just to see what it’s about.

for two main reasons:

1) never feed anything proprietary into a large language model (LLM, eg ChatGPT, google bard, etc.).

this means don’t give it private company information when you’re at work, but also don’t give it your original writing. that’s your work.

because of the way these language models work, anything you feed into it is part of it now. and yeah, the FAQ says they “don’t claim ownership” over anything and yeah, they give you that reassuring bullshit about how unlikely it is that the exact same sentence will be reconstructed—

but that’s not the point.

do you have an unusual way of constructing sentences? a metaphor you like to use? a writing tic that sets you apart from the rest? anything that gives you a unique writing voice?

feed your writing into an LLM, and the model has your voice now. the model can generate text that sounds like it was written by you and someone else can claim it’s theirs because they gave the model a prompt.

don’t feed the model.

2) the other reason is that sudowrite scraped a bunch of omegaverse fic without consent to build their model and that’s a really shitty thing to do, because it means people weren’t given the chance to choose whether or not to feed the model.

If this makes you mad, the Federal Register is currently open to comments on AI accountability until June 12th, 2023.

It only takes a second to leave a comment to ask for legislation that works used in AI creations or training MUST secure the express consent of the original creator before they can be used. If we can get protections for artists, writers, musicians and everyone who creates that their work cannot be used in AI without their permission, we can go back to making fanworks freely available without fear of them being misused. Until then, we're stuck playing defense until the courts catch up.

(If you're a fan creator looking to do this as well, AO3 has a tool to let you do all your fics at the same time in seconds. On your dashboard, go to Edit Works and you'll be able to change the status on everything at once.)

Hi, AI ethicist + fanfiction expert here. (This is one of those times where I feel uniquely qualified to comment on something...)

I’m seeing this weird game of telephone about the Sudowrite AI that I think started out pretty accurate, but now has become “Elon Musk created an AI that is stealing your fanfiction” (which frankly gives him far too much credit). I can probably say more about this, but here are a few things that I want to clarify for folks, which can be boiled down to “Elon Musk has nothing to do with this” and “this is nothing new”: Elon Musk is not involved in any way with Sudowrite, as far as I can tell. Sudowrite does, however, use GPT-3, the widely-used large language model created by OpenAI, which Elon Musk co-founded. He resigned in 2018, citing a conflict of interest due to Tesla’s AI development. It wasn’t until after he left that OpenAI went from being a non-profit to a capped for-profit. Elon Musk doesn’t have anything to do with OpenAI currently (and in fact just cut off their access to Twitter data), though I can’t find anything that confirms whether or not he might have shares in the company. I would also be shocked if Elon actually contributed anything but money to the development of GPT-3.

Based on Sudowrite’s description on their FAQ, they are not collecting any training data themselves - they’re just using GPT-3 paired with their own proprietary narrative model.  And GPT-3 is trained on datasets like common crawl and webtext, which can simplistically be described as “scraping the whole internet.” Same as their DALL-E art generator. So it’s not surprising that AO3 would be in that dataset, along with everything else (e.g., Tumblr posts, blogs, news articles, all the words people write online) that doesn’t use technical means to prohibit scraping. 

OpenAI does make money now, including from companies like Sudowrite paying for access to GPT-3. And Sudowrite itself is a paid service. So yes, someone is profiting from its use (though OpenAI is capped at no more than 100% return on investment) and I think that the conversations about art (whether visual or text) being used to train these models without consent of the artist are important conversations to be having.

I think it’s possible that what OpenAI is doing is legal (i.e., not copyright infringement) for some of the same reasons that fanfiction is legal (or perhaps more accurately, for reasons that many for-profit remixes are found to be fair use), but I think whether it’s ETHICAL is a completely different question, and I’ve seen a huge amount of disagreement on this.

I have been seeing people panicking about sudowrite, and I have decided to write a small information post to hopefully clear up some common misconceptions I have been seeing.

Elon Musk is involved.

No. This is blatantly false. Elon Musk was involved with openai - NOT SUDOWRITE - when it was founded. He co-founded the company with a 1 billion dollar investment in 2016, 2 years later in 2018 he left the company. According to this Bloomberg article, he left because "Musk has to focus on “a painfully large number of engineering & manufacturing problems” at Tesla & SpaceX, he said, adding that he hasn’t “been involved closely” with OpenAI for more than a year. Musk left the company’s board in February [2018].

Why is when he left important?

The first gpt paper was published on June 11th 2018. This is 4 months after Musk left the company, and over a year since he said he stopped being involved. He has not been involved with gpt whatsoever.

What is Sudowrite

Sudowrite is, to my understanding, in the most basic terms a word processor with a generate text button.

Sudowrite operates with a gpt-3 model. This is the third gpt model. Gpt is a text-generation program created by the aforementioned openai. Not sudowrite. Basically, gpt is a transformer model. There is a complicated way of putting this, but in short it generates human-like text.

Why can it generate fanfic?

From 2011 to October 2019, an internet crawl was done to train an ai on the entire internet. This includes most websites. This includes ao3. Gpt is trained on this data. From my understanding, no new data is being added to gpt from websites like ao3. Privating your fanfic will do nothing against what is already collected and in use, as the collection of data is already finished, and not current. Please stop panicking.

What is sudowrite trained on?

Thankfully, for this we have a simple awnser.

In this faq (which you can find at the bottom of this page) it is stated that the sudowrite copy of gpt-3 has the models data, (the common crawl, web text, wikipedia, etc) and 10s of thousands of books. As it states here, again, there is no new data being collected. It is from 2019. Privating things will do nothing.

What else can I do to be informed?

Has that been confirmed? Did the bots drop other names?

Because someone out there is really helping a LOT of users learn about how Sudowrite scraped all AO3 up until 2019 to make its plagiarized database.

I wouldn’t be surprised if FFN’s data got scraped by sudowrite. I’m not gonna go and delete everything there, which means there is no way for me to protect the writing that’s already there. I will try to protect my future writing, which means that you shouldn’t be surprised if in the future I stop crossposting to FFN. Likewise, I have a few fics that are multichap WIP’s. They will likely become archive-locked when I update them, assuming I don’t update the FFN version.

I’m sorry, guest commenters. I love you, but I’m protecting myself.

I wanted to experiment with Sudowrite’s Story Engine, so I used it to generate a 12-chapter, roughly 22,500-word novella over the course of a weekend. The generation was based on around 6,500 words of story beat prompting, with the skeleton generated by AI but rewritten mainly by myself. Here is the entirely human-written (except for some names I got the AI to suggest) “brain dump” synopsis in its final version, following several rewrites as the story progressed:

A slipstream novel set in an alternate future about a hacker named Jack who is arrested by corporate security and taken to a domed city in the middle of the ocean, managed by an AI named the Curator, who produces the city's elaborate theme park-style culture, which has been corrupted and broken down by years of corporate control and subcultural resistance. A rogue executive asks Jack to come to the city to perform a mysterious job. Jack agrees, and he is assigned a partner: a former underground musician who now uses the AI to program the city's pop music. But the morning after he arrives, the executive is murdered, and Jack and his partner are framed for the crime. They escape to the floating city's underbelly, where they learn that the AI killed the executive and has secretly been running the city for reasons incomprehensible to human minds. Jack and his partner must decide whether to stop the AI or allow it to run the city.

I gave it a slew of genre and prose parameters that may or may not have made much difference, honestly:

Genre: Cyberpunk, literary, surreal, slipstream, hardboiled, neo-noir

Style: Terse, hip, literary, avoid cliches, include futuristic cultural references, focus on emotions, be weird, reduce metaphors, include surreal elements, hardboiled, noir

And here's the rest below. Each chapter includes the text of the story, which is almost entirely generated by the Sudowrite AI, sometimes with one or two do-overs. It also includes the chapter summaries, which were generated as a long list based on this "brain dump" and then almost entirely rewritten by me as the story progressed, as well as the individual story beats that the AI was working with.

Honestly, this thing is more coherent than I expected. Or as Sudowrite might put it: it will lead you down labyrinthine passages that send a shiver up your spine. The choice is yours and yours alone -- but at what cost? We're playing with fire.