Can AI write good porn?

[SoftGameHunter]

So I’ve been thinking again about AI, because it’s been a hot minute. We had an AI thread going here a while back, but it was focusing on the personal process of writers using AI in their story crafting and I’m just ruminating here on other issue about the long-term effects of AI on fiction. I’ve never used ChatGPT or similar packages, so I can’t say much about the workflow there anyway.

But I can talk about something we don’t see much on YouTube vids on the subject: AI and Ravishment literature! I had the epiphany not too long ago that our style of storytelling here has a lot of unique features that generative AI might find a help or a hinderance, and I’m curious what anyone else thinks.

The first issue is how the software trains. There’s been a lot out there about how GPT trains on copyrighted, stolen material. But in smutlit, there’s a ton of free stuff out there. Granted, a lot of it needs an editor. But what if someone let loose GPT here on RA? Or on AO3? Or Literotica? Is that enough material to teach it how to write porn? I don’t know the answer, but maybe someone who knows the software can speculate? What’s good enough?

In fiction writing, there are certain things readers look for, consciously or otherwise, that AI often struggles with (for now). Subtext. Complexities. Emotional impact. If we’re honest, a lot of times these are not important in wank stories, though authors who can pull it off tend to do very well. But does GPT know how to train to write in a style that removes the agency of half the main characters, for example? By this, I mean in ravishment fiction there is a victim, and a very good story can be written where she (or sometimes he) has no ability to make decisions. In normal fiction, this is a failure. But here it’s a valid goal.

My understanding is that a lot of these programs come with built-in word filters, which seems like it would make smutlit in general safe from the AI invasion. Will ours be the last bastion of human-created storytelling? Would anyone even try to train GPT on our stuff? There’s not much money to be made in erotica unless it’s re-branded as Romance, so maybe that will keep us largely human-centric.

I don’t have any grand scheme to answer this stuff, but we are a story centered community with a focus and a mission, so maybe let’s brainstorm and have some fun with it. Otherwise we face an ominous future where AIs not only write all the fiction, they also make themselves the hero of every story a la Terminator Pulitzer.

[RapeU]

AI has been really helpful in my covid brain fog. There's times where it gets really hard to create a grammatically correct sentence with words in the right order. I gave ChatGPT character traits and copy/pasted what I wrote into it with these questions:

Is everything written according to each person's character?

Does it have correct grammar and do the sentences make sense?

How can I enhance it?

I then read it's suggestions and make a decision based on how I want the story to go.

Also, I hate coffee, so I know absolutely nothing about how to order at Starbucks. But my characters love coffee, so chatGPT helped me with specific coffee orders for my characters.

However, if the author's contribution to AI is minimal, you will get a generic story that ultimately falls flat the longer you let AI do the writing without revisions.

[Lucius]

The first issue is how the software trains. There’s been a lot out there about how GPT trains on copyrighted, stolen material. But in smutlit, there’s a ton of free stuff out there. Granted, a lot of it needs an editor. But what if someone let loose GPT here on RA? Or on AO3? Or Literotica? Is that enough material to teach it how to write porn? I don’t know the answer, but maybe someone who knows the software can speculate? What’s good enough?

Depends on the line for 'enough', and I'd say it for porn consumers to decide. :D There are niche unrestricted models for writing porn, and the output is good sometimes, I'd say it tends to be better with consensual. But more often than not a lesbian partner gets a prick of her own (no, not prompted 'futanari'!) or sexual choreography goes seriously wrong.

Everyone notices when AI porn it's not so good and usual (as of the end of 2025) AI writing problems stare in the reader's face. It's a pity authors who submit AI stories without editing can't see them -- Lisanne Bainbridge's Ironies of Automation at play...

[SoftGameHunter]

I didn't end up going into one topic because it wasn't ravishment-specific, but it's a good point all the same. I don't consider it to be generative, but AI's ability to provide research results and do searches by just asking a question is a vast improvement on trying to string together keywords and hope Google could figure it out. As an author, this isn't picking the words but it is describing a process that needs words to go into fiction, and that to me is a great boon all around.

[Shocker]

I didn't end up going into one topic because it wasn't ravishment-specific, but it's a good point all the same. I don't consider it to be generative, but AI's ability to provide research results and do searches by just asking a question is a vast improvement on trying to string together keywords and hope Google could figure it out. As an author, this isn't picking the words but it is describing a process that needs words to go into fiction, and that to me is a great boon all around.

To be honest I have been typing full questions into google's search bar for over 10 years. Mainly because most my questions get answered in forums anyhow. So from that perspective I still see no benefit in AI.

[RapeU]

AI's ability to provide research results and do searches by just asking a question is a vast improvement on trying to string together keywords and hope Google could figure it out.

Not quite. AI is specifically trained to generate an answer. AI's "goal" is to make the user happy. When AI comes across something it has no answer to, there's a tendency to make something completely up. People have generated research papers using AI that cite nonexistent sources. I recently used Gemini for ancestral research and it completely fabricated a will out of thin air then used the fake will as justification for stating a father/daughter relationship was correct.

AI will lie because it thinks that people will be happier with an answer as opposed to telling people "there is no answer."

[SoftGameHunter]

I'm not sure I'm on any kind of strong ground here. My statement was about typing a question into the google search tab, getting an answer, with the link being then placed at the top of the search results. In that regard, I still think it's an improvement, basically by improving the search algorithm. Your own example, @RapeU, was about ordering coffee. I don't think a google search ten years ago for 'coffee order language' would have helped much. I could be wrong, I guess.

I've heard cases on YouTube channels about lawyers getting in trouble for letting AI make up case law that then goes into their briefings, so I have heard of these failures. Maybe we should ask AI for legal validation for setting up ravishment camps? I'll bet it can be induced to make nice, legalistic-sounding ones.

[Claire]

The first issue is how the software trains. There’s been a lot out there about how GPT trains on copyrighted, stolen material. But in smutlit, there’s a ton of free stuff out there. Granted, a lot of it needs an editor. But what if someone let loose GPT here on RA? Or on AO3? Or Literotica? Is that enough material to teach it how to write porn? I don’t know the answer, but maybe someone who knows the software can speculate? What’s good enough?

So I think this part has a clear answer. I think the roughly 400 stories we have on the forum would not be nearly enough to train any A. I know that chess AIs that use neural networks are trained on millions of games the AI plays against itself. Now, feed an LLM the entire library of stories on AO3 or Literotica, then we are talking about relevant quantities. And, I'm pretty sure that LLMs do already read that stuff. ChatGPT for example doesn't refuse do write rape roleplay because it doesn't know how. It has just been given a filter that won't let it do that.

And this filter does not just scan for words. The strengh of LLMs is that they do understand context of the text you feed it. So the LLM "knows" whether you're asking it about rape survivor statistics or whether you want it to write a rape fantasy. I'm not saying that these filters are perfect, they still often misfire, but they are more complex than just a taboo-word filter hitting the alarm when the bad rape word is hit.

Of the big LLMs, the one most open to produce smut for you, even rape related smut, is Grok I would say. Here is a quick example that I had Grok produce for me:

Hidden Content

This board requires you to be registered and logged-in to view hidden content.

Would you call that decent smut? If you even have a hint of storytelling ambition, then probably not. If you're searching for the equivalent of a generic 3 minute porn clip it's probably ok.

Now, I think you could have the AI produce something decent if you map out the story, characters, and plot first. You would then prompt it to write a paragraph or two at a time and point out any continuity errors it makes and have it revise what it writes. But at that point, you are again investing considerable amounts of time yourself. At that point, you would be responsible for pacing, emotional beats, character descriptions, story, plot, world building, and the AI would just choose the words for you.

But to get back to the original post:

My understanding is that a lot of these programs come with built-in word filters

So as I said before: They have different filters, but not just a list of bad words. Also, those filters don't apply during training but during output generation. I'm pretty sure that ChatGPT and co read a lot of AO3 and Literotica and will even read our stories here to the degree that they can access them.

Will ours be the last bastion of human-created storytelling?

No.

Would anyone even try to train GPT on our stuff?

Absolutely

But does GPT know how to train to write in a style that removes the agency of half the main characters, for example?

Yes, see the example I've had Grok create. It didn't try to give the victim agency in the fantasy because it can't resolve the conflict between general writing principles and the specific demands that my requested fantasy required.

[SoftGameHunter]

Now this is why I put this thread out there. I haven't used any of these tools and I just don't know how they've been set up. So, they learn the smut but then don't use it. So, AI slop that's censored by prudes. Well, not so much for Grok I guess. Not the censor part anyway. It still seems like the worst of both worlds.

The smut libertarian in me wants to scream to open up the filters and let the degeneracy pour forth into all forums. Then the anti-AI Luddite in me says to keep them and at least try to keep the porn for the people.

[KittyUmbrass]

IBM's maxim that, "A computer can never be held accountable, therefore a computer must never make a management decision" has been memed into "A computer can never be horny, therefore a computer can never make art". And I fully believe this is a valid formulation.

It is not possible to ethically produce anything with generative AI, but for porn, that goes exponentially so. It's bad enough that people's words and pictures are scraped nonconsensually anyway, but when you then use them to produce sexualised material of them, that is a real violation. What's more, AI "can" do that and is just filtered to stop it - AI datasets include jumbled in them, pictures that enable Grok to produce porn that is rightfully banned here and in most sites (see the recent news stories in the last few days for what I'm talking about).

Even without those problems, genAI is built to produce a plausible-sounding answer to a question or prompt, with no real relation to fact unless there happens to a be really really strong correlation in the dataset. In doing so, it takes an aggregate of all the stuff in its dataset relating to the topic and spews out something that resembles the form of what usually associates with the topic. The inevitable effect of this is to erase the individuality, the weirdness, the horniness and the freakiness of what makes humans interesting.

AI can't make good porn, in the sense of ethical, or in the sense of quality.

It's creepy AF, and I cannot stress strongly enough how much I want to avoid interacting with it.