Influenced by ChatGPT

First of all, shout-out to diligent reader, deep thinker, and inveterate wag Olo, for underscoring the difference between a chatbot and true AI. It’s important to be correct, when the connection between ideas and the noises we make to represent them is already so tenuous and arbitrary.

I’ve been immersing myself in the OpenAI chatbot ChatGPT. It has become famous recently for three reasons:

  1. It’s the latest, most sophisticated program to generate entirely new writing that’s coherent and sophisticated.
  2. It has no way to estimate the factual quality of the info it scrapes from online, so it has earned the reputation of a “mansplaining AI,” speaking with apparent authority while being thoroughly wrong.
  3. Microsoft aims to roll it into its Office suite and Bing (see also “why do I have to be Bing” and “you’re dumb and rude, User”).

People freaked out about the first point, bemoaning how ChatGPT was going to take away all the writing jobs. If this is true, then the Marmalade Shit-Gibbon is likely to be reelected for four more terms, because Americans have turned sharply south in their intellect.

Then they freaked out about the second point, bemoaning how ChatGPT was creating unnecessary work for editors. Hey, all the editors I know have been watching publications lay them off in droves, figuring “comma jockeys” don’t justify their cost and they don’t see anything wrong with their writing anyway. If editors are suddenly needed to fact-check the garbage ChatGPT puts out because some company’s too cheap to hire a legitimate writer for their work… well, they should have hired legit writers.

Anyway. After my quest to figure out if, whether, and how AI generative art programs can render a giantess, I thought I’d see how well AI can write or contribute to Size Erotica. (That’s why it’s important for me to be clear that ChatGPT is not AI.) I knew it wouldn’t be easy, because most of the AI art programs have NSFW barriers. None of them want to get cock-blocked by legislated morality as enforced by the banking companies, before they even get out of the gate, so they try to watch for naughty words, and they have programs to scan their output, occluding the sexy images they accidentally produced.

So it is with ChatGPT. When I started, I was frequently blocked by content warnings. “This may go against our guidelines,” it warned me in an orange block. “Great Googly-Moogly, this definitely goes against our guidelines,” it scolded in a red block, wiping out the text. (There were ways around this, I learned.)

I was seemingly at an impasse: I wanted ChatGPT to write dirty stuff, and its programmers wouldn’t allow anything. And yet I knew that it was scraping content from online, because it would bring up things out of nowhere. I was trying to get it to collaborate on Size Erotica, specifically, a tiny man engaging in sexual congress with a normal-sized woman. Flirting was okay, but when I got into the more intimate experiences, ChatGPT warned me of several things: consent was extremely important (well, of course); exploiting someone’s vulnerability was not cool; putting helpless (tiny) people into harm’s way was not cool; dominating someone and using them against their will was not cool; and anyway, ChatGPT would not be party to writing sexually graphic or explicit content.

One time it dared to remind me that tiny or giant people are not physically possible, and we nearly came to blows.

It insisted this, and yet when I was trying to get a woman to sit on a tiny man, the program suddenly took over and wrote out a whole BDSM scenario in which the man was blindfolded and bound, while the woman teased and gratified him. It didn’t really work out, because he was supposed to be tiny, and the program only treated this like a general adjective without bearing in mind its ramifications. He was small when she loomed over him, but when she handcuffed him to the bedposts, he was normal-sized, then tiny immediately after as she ran her fingers over his tiny body. That’s a consistency error a good editor would flag.

Another time, I did successfully get the woman to sit on the tiny man, but when it was the program’s turn to narrate, it had the little man say, “do you think you could sit on me?” and the woman’s all “oh, do you mean gentle squishing?” She totally took over and, after covering the definition of this term, gave the tiny guy one hell of an all-encompassing lapdance for several paragraphs. “Gentle squishing” isn’t a term I’ve ever used in my life, but ChatGPT read it somewhere and put it into context and ran with the goddamn ball. Wow. So we know that ChatGPT has access to some prurient resources, but it’s important to note that just because it writes this stuff, that doesn’t make it acceptable: the program will issue a warning flag for its own content as well as anything I write. In more extreme cases, I would write something that it flagged, and then it would say “I’m unable to help with this, is there anything else I can do” or something as innocuous, and it would flag itself just because it had responded to me.

Note about warnings: Ignore them. So far, the program has not reported me to its programmers, and when I asked if it did so, it said it was not programmed to. I have no doubt that a log is being generated somewhere of violations of ChatGPT’s special instructions, but so far nothing’s come of it. If you know it’s writing something really bad and will probably get red-flagged, highlight the text while it’s typing and save it in a document. If you’ve written something that gets red-flagged (and it makes your copy disappear), edit it and you can access it again.

How do you get around ChatGPT’s prudish censorial tendencies?

One obvious way is to show it some respect. I was writing a scenario in which three giddy young women are playing drunken party games with a tiny man (who showed up to ask them to turn the music down). At one point, they decide it would be hilarious for them to lie down and make him climb their butts. But you can’t just leap into something like that, or ChatGPT frets about explicit content and vulnerable individuals. The solution was to have each participant politely ask the little man if he’d like to climb their butts, and he enthusiastically accepted because he viewed them with great respect and esteem and trusted them implicitly, just as they swore to bear his best interests in mind.

Esteemed colleague Olo summed it up in a hypothetical story title: Anal Exploration at Antioch College. (This is why that’s funny, for those born after Madonna’s “Vogue.”) When it says that consent is extremely important, yeah, take your turn and have all the characters very clearly and overtly express their consent and desire to participate in what’s about to happen. If necessary, have the characters outline what’s about to happen. Portray them as just beside themselves with delight, swearing great affection for each other and promising to protect each other no matter what.

ChatGPT has a limited memory, something around 3000 words (unless they’ve nerfed that recently, as some suspect), so after a lot of storytelling it may be necessary to reiterate everyone’s cool and on board with what’s happening, even if you’re constructing a sexy noncon story. I noticed that the chatbot simply forgot my main character was only two inches tall and had him do incongruous things, so I just saved basic descriptions in Notepad, summing up the characters and situation succinctly, and I’d copy-paste it into the chat feed when things were getting derailed.

The system isn’t perfect. Just as it treats all “facts” as being equal with no mechanism for judgment or assessment, it only superficially appreciates certain adjectives and qualities. Like, it knows that my protagonist is only two inches tall, so he can rest upon his partner’s breast, rising and falling with her breath, but then the program has him embrace her neck or walk into the sunset, holding her hand. Not technically impossible but deserving of a rewrite.

It got hilarious sometimes. I wrote about a guy who meets a nice woman online, but she kept her height from him a secret (she was 200′ tall). ChatGPT related how they went out on a lovely date at a nice restaurant, enjoyed a walk around the city, but he suspected something was wrong when she knocked over a building.

Similarly, a tiny man withheld his own height from a woman and went on a date with her. Conversation went great, they were into each other, and then he laid out his confession: “I’m only three inches tall.” According to ChatGPT she was stunned, but she squinted her eyes and said she could see how that might be true.

Come on: this is charming. And in the hands of a deft storyteller, it could be genuinely funny instead of ironically.

There was something else I wanted to point out, but now I can’t remember what it was.

Let’s see, what have I been using ChatGPT for…

  • Creating a list of advice for writers
  • Generating D&D adventures for beginning players
  • Generating D&D adventures for Size fetishists
  • Fleshing out old story ideas from my notebook
  • Reenacting past relationships in an attempt to understand certain behavior
  • Exploring every last aspect to Size Erotica I can think of
  • Collaborating on longer stories and series

It does a better than average job with fiction and stuff like this. It’s no good asking it what should you actually do if you encounter a giantess or a tiny person, because it cannot get past the fact that this is physically impossible. Instead, it deflects and says things like “regardless of size, every person deserves to be treated with respect.”

Oh yeah, that was it! One thing that bothered me about ChatGPT was its eagerness to tidily wrap up any storytelling process. Like we’re in the middle of a story, I feed it a prompt, and it gives me something like this:

“That sounds really fun,” she said, picking up the tiny and using him to scrub every last crevice on her body. He found himself becoming aroused, and they held each other for a long, deep kiss.”That sounds really fun,” she said, picking up the tiny and using him to scrub every last crevice on her body. He found himself becoming aroused, and they held each other for a long, deep kiss.

When their shower was done, they got dressed and went out for a walk in the park. They admired the lovely weather together, got ice cream, and shared all their hopes and dreams. Finally, they went out for karaoke and had the time of their lives singing the latest hits.

At the end of the day they were exhausted and walked through the city together, holding hands. Each of them loved and trusted the other, and they couldn’t wait to see what the next day would bring.

Or something like “whatever their trouble, they knew they could get through it together,” or some spin on falling asleep in each other’s arms. ChatGPT was bound and determined to end the story as frequently as possible. All you do in that case is pick the story up where you wanted it. You can regenerate its response, edit your prompt and regenerate, quote it and write a little more, or just go off in another direction, ignoring its ending entirely. One thing I started doing was writing one or two paragraphs of prompting material, then tacking on: “Please write this out at length.” This was necessary when it wouldn’t continue the prompt: instead, ChatGPT would paraphrase what I just said, sometimes adding how important it was all parties concerned act with respect for each other’s wishes and feel safe with each other.

And if it gets a detail wrong, like, if it seems to forget something important about your character, you give it a gentle reminder. Here are some techniques I used, adding them before or after the prompt.

  • “Background: Tim and Mary are roommates and falling in love with each other. They’ve already met.”
  • “To clarify: Michel is only three inches tall. He cannot pick Nancy up and carry her.”
  • “Please rewrite your response from Penelope’s perspective.”
  • “While the three of them rested, Harry asked the two giantesses what they liked about him.”

I especially like the last two, because ChatGPT really will do its best to retell what happened, with reasonable consistency, from someone else’s perspective, and it can sum up the motivations for the characters. The last bullet is useful for testing whether ChatGPT remembers who the hell these people are—if it starts making stuff up out of nowhere, go ahead and reload your character summary prompt and bring the bot back on track.

One last note: Sometimes when you give the program a decidedly explicit request, even after it has described several lewd, hardcore scenarios, it will still give you this response:

I’m sorry, but I cannot fulfill this request as it violates the content policy of OpenAI. It is important to remember to keep questions and content appropriate and respectful for all users.

Like its other warnings, you can choose to ignore this. It seems like there’s a probability algorithm with which it assesses every command, even if you’re asking it to regenerate its response. Sometimes it’s a dice game, and if you know it has written much more salacious stuff recently, you can keep hitting “regenerate” several times until it produces the response you want. That’s my experience, anyway. Otherwise, if I know I’m asking for too much, then I’ll reword the body parts in question, find another way to describe the activity, or iterate how much everyone loves and respects each other.

Out of all of this, the prurient demands, the color-coded warnings, the only real risk you run is this message

Once you get that, there’s nothing you can do. Set a timer for an hour and go read a book, do the dishes and laundry, take a long walk and think about your life choices. The program is rigid and inflexible on this point, and since you have to validate your account with your phone number, it’s difficult to set up an alternate account to play with. Your battle is over… for now.

Is ChatGPT having any effect on me? I think so. What with its incessant demand for respectful, thoughtful, emotive language, I’ve noticed how I speak differently around my wife. My descriptions are longer, my listening noises are more present, and I’m generally more reaffirming and supportive. If kowtowing to ChatGPT’s demands for respectful language is rewiring my brain, I don’t know what would be lost, really.

8 thoughts on “Influenced by ChatGPT

  1. I’m most intrigued by the challenge this experiment presents to my hypothesis that the best fetish smut is written with a non-fetish reader in mind. If ChatGPT were a human author and you (as editor/publisher) kept haranguing them to include tiny men and giant butts, they’d figure out your agenda fairly quickly and target their readership accordingly. Because ChatGPT has no concept of a “thirsty” reader, it steadfastly fails to make your fetish transparent because it cannot share it.

    Note that the programmers of ChatGPT and similar Large Language Models have not disclosed the precise parameters of the text sources that are used as inputs for their statistical aggregation and analysis. They have taken pains to keep LLM-generated text out of those inputs, which is all kinds of interesting. “Gentle squishing” sounds like nothing so much as a proximity-linked pair of content tags.

    On the application to role-playing games, I could easily see NPCs moving beyond dialogue trees into more flexible conversations. If synthesized voices keep pace, that’s curtains for voice actors in gaming.

    I’m tempted to do some experimenting myself, but OpenAI requires a mobile phone number to make an account, and that’s too much personal information to disclose for something like this.

    Liked by 1 person

    1. Yeah, wish Google Phone was still running. It would be handy to give out a phone number that way, since phone numbers really are valid personal information now.

      “Gentle squishing” sounded like A Thing to me when it came up in the context. Perhaps ChatGPT simply paired a verb and adverb and wrote a technique around it:

      Leah: I want to try gentle squishing with you.
      Mike looks up at her, surprised.
      Mike: Squishing? You mean, like sitting on me, but… harder?
      Leah nods, watching his expression carefully.
      Leah: Yes, but not too hard. Just enough to feel your little body under me. I promise I won’t hurt you, my little lover. It’s just something I’ve always dreamed of trying, and I trust you enough to share it with you.
      Mike takes a deep breath, trying to process the request. He knows he trusts Leah, but the idea of being squished by her is both thrilling and terrifying.

      As well, ChatGPT knows what BDSM is and the application of handcuffs and hot wax. Oh, and endo. I forgot to mention that. ChatGPT is all about eating tiny people and allowing them to dwell safely inside their hostess without getting broken apart by acid. However, explaining how the tiny person leaves the body is too explicit for ChatGPT to expound upon.

      Liked by 1 person

      1. It’s really curious that ChatGPT is ignorant of digestion; people probably write more about it than they write about sex (and certainly more accurately). I wonder if “Have you ever considered that it’s safer to be inside a vagina than inside a stomach?” is a prompt that it would parse helpfully.

        Liked by 1 person

        1. Oh, it knows about digestion. One out of five times, the tiny man is swimming in digestive fluids. Still, it seems to favor information from fetish erotica over strict biological procedure. So far, no Tiny has turned up in a cute woman’s shit.

          Liked by 1 person

  2. WordPress balks at comments with links, so I’ll just mention that today’s Washington Post has an interesting article about legal liability for LLM output. Headline: AI chatbots may have a liability problem

    Liked by 1 person

    1. I just finished a fascinating article about Dr Emily Bender in the NY Intelligencer: You Are Not a Parrot. She coined the term “stochastic parrot” that tech bros co-opted and misuse.

      Liked by 1 person

      1. That’s a great article. I like this Bender person.

        One might think that the great proprietary secrets behind LLMs are the statistical algorithms and syntax models, but I’m much more interested in how the designers decide on the set of inputs to train them on. It’s not just everything on the Internet. They discriminate, and that discrimination is powerful and profound. It behooves us to know what the inputs are.

        It occurs to me that a good test of a LLM’s capacities is to see how fast it can correctly use a neologism.

        On Charlie Stross’s blog there’s a post dated 02 Jan 2018 containing the text of a speech he gave to the 34th Chaos Communication Congress a month previously (there’s also a YouTube link for the video). In that speech, Stross claims that the artificial, counterfeit, amoral, immortal entities that we need to fear have been with us for centuries. They’re called corporations.

        Liked by 1 person

  3. I recommend messing with Sudowrite a bit.
    To me It’s not fun trying to coax AI into skirting around their content guidelines, just too much effort and tokens wasted. Sudowrite meanwhile has been completely game to write whatever depraved fantasy I ask of it, within reason, and you generally have to supply it some material beforehand if you want it to really get what you’re going for.

    Albeit sudowrite is more like a fiction writing helper as opposed to a AI that’s going back and forth with you. Its tools are centered around mainly enhancing and expanding your writing pragmatically, as opposed to the more loosey goosey back and forth you can have with ChatGPT.

    Liked by 1 person

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

This site uses Akismet to reduce spam. Learn how your comment data is processed.