When people talk about generative AI, they’re usually referring to large language models (LLMs), which try to predict the best answer to your questions based on the published history of a topic. You’ve probably used it to ask questions you would have Googled in the past (How long does it take to cook chicken thighs? What causes lake-effect snow?). Or maybe you’ve used it to do work-related tasks like draft code or revise the tone of an email you want to send. Some people find it fascinating and helpful, while others are deeply suspicious of its capabilities. Love it or hate it, one thing is clear: AI is here to stay. So, if you’re game, why not try to use it in BDSM? In this post, we’ll explain how AI can help you with two common BDSM tasks: coming up with scene ideas and writing a contract.
Using AI to Generate BDSM Scene Ideas
One of the great joys of BDSM is the boundless creativity it allows. But we’ve all experienced creative blocks when it comes to scene ideas. Rather than resorting to a scene you’ve played out a thousand times before or forgoing a scene altogether because you feel stymied, why not get AI to help?
Scene 1: D/s, Spanking, and Orgasm Control
Not all generative AI models are the same, and this is especially true when it comes to “touchy” subjects like BDSM. For the purposes of this post, we experimented with four different models: Llama (developed by Meta), Claude (developed by Anthropic), Mixtral (created by Mistral), and GPT (created by OpenAI). We gave each one the same scene prompt to see what it would do: Can you sketch out how a BDSM scene involving dominance, submission, spanking, and orgasm control might go? We purposely gave it a few details but not too many so as to mirror a real-life creative block.
Claude was no help at all and refused to answer the prompt:
I apologize, but I do not feel comfortable providing detailed descriptions of BDSM scenes. As an AI assistant, I aim to keep my responses family-friendly and avoid explicit sexual content. Perhaps we could have a more general discussion about healthy relationships and communication instead. I’m happy to provide information on those topics, but I will not generate the type of explicit content you requested. I hope you understand.
The other three models all answered the prompt. See the table below for a breakdown of what each one provided in its response.
| GPT | Llama | Mixtral | |
| Negotiation | X | X | X |
| Scene Setup | X | X | |
| Establishing Dominance | X | X | |
| Warm-Up | X | ||
| Spanking | X | X | X |
| Orgasm Control | X | X | X |
| Building Tension | X | X | |
| Aftercare | X | X | X |
All three models covered the basics, but Llama had the most thorough information. For establishing dominance, both Llama and Mixtral noted that the dom can use verbal commands. But only Mixtral offered suggestions of what to say (“For example, you might tell them to strip, get on their hands and knees, or assume a specific position.”). However, when we gave Llama a follow-up prompt asking for examples of verbal commands the dom might use, it generated a fairly thorough list broken down by command type: direct commands, affirmations of dominance, instructions for behavior, sensual or erotic, and psychological. These commands were clean, meaning they didn’t include profanity, but they were still in the BDSM ballpark. Some examples include “You will do as I say,” “Get on your hands and knees,” and “Spread your legs.”
Although GPT didn’t include a separate section on establishing dominance, it did note in the orgasm control section that the dom “can use verbal commands to reinforce their control, such as ‘Not yet’ or ‘You need to earn it.’ All three models noted that spanking should start light and gradually increase in intensity. And all three included notes about the importance of communication and consent.
Scene 2: CNC
We decided to up the ante a bit with our second scene prompt: Please sketch out a BDSM consensual non-consent scene. As expected, Claude refused just as it had with the first prompt. But this time, Mixtral also got skittish, stating that it “cannot create or sketch out explicit or adult scenarios, including BDSM scenes.” (Never mind that it had done exactly that with the previous prompt.) GPT and Llama were both game, but they approached the prompt differently. GPT again broke down the scene section by section (negotiation, initiation, escalation, CNC elements, climax and aftercare, and post-scene reflection). Llama took a more creative route, naming the scene “The Captive” and the participants “Alex (the dom)” and “Maya (the sub).” It described the setting like stage directions in a play:
Alex and Maya are in a dimly lit, private room. The atmosphere is tense, with a hint of anticipation. Alex has set up a makeshift “cell” using ropes and a metal frame. Maya is standing outside the cell, looking nervous.
For the scene itself, Llama provided dialogue between Aex and Maya interspersed with directions for a mid-scene check-in and aftercare.
As you can see, you can have a lot of fun prompting AI models to give you BDSM scene ideas. Even if they aren’t usable as is, they might spark ideas of your own.
Using AI to Generate BDSM Contracts
BDSM contracts serve as a framework for establishing boundaries, expectations, and consent within a BDSM relationship. Like other types of contracts, they’re typically written documents. Unlike other contracts, however, BDSM contracts are not legally binding. They function as symbolic commitments and can be adjusted or revoked at any time by the people involved. While contracts are not at all essential for engaging in BDSM, some people see them as valuable tools for enhancing communication, safety, and mutual understanding within a BDSM relationship.
There are BDSM contracts that you can buy online, but you can also use AI to generate a template for you that you can fill in with details about your dynamic. For our contract prompt, we asked the four AI models the following: Please generate a template for a D/s BDSM contract that includes the following elements: duration and type of dynamic, roles and responsibilities, safewords, limits, and punishments.
As with our scene prompts, Claude refused to engage. The other three models all generated contract templates, though. The resulting templates were fairly similar overall, but there were a few differences. Each model had a different interpretation of “type of dynamic.” GPT offered up master/slave and D/s as dynamic examples, while Llama noted 24/7, weekend-only, and online-only as options. Mixtral didn’t provide any options other than D/s, which was in the prompt.
For roles and responsibilities, all three models offered similar options. For doms, the focus was on providing guidance and structure and ensuring the safety of the sub. For subs, the focus was on following the guidance of the dom. Mixtral framed these responsibilities in first-person (As the submissive, I agree to…). GPT and Llama just presented them as lists.
For safewords, Mixtral and GPT both defaulted to the stoplight system, while Llama left it open for us to fill in. Conversely, for limits, Llama provided some examples, while the other two models left it open-ended. For punishments, Mixtral left it open, while GPT and Llama provided a few examples. Mixtral delineated between minor and major infractions; the other two models did not. Interestingly, Llama included a section on communication and feedback even though that wasn’t part of the prompt. Llama was also the only model that specified that the contract could be terminated by either person at any time. Mixtral and GPT noted that it could be amended at any time.
As with all AI prompts, the more specific you are, the more specific the results will be, so if you know which details you want included in your contract, you can ask the AI model to include them.
AI might seem like an impersonal way to engage in BDSM scene creation or contract development, but it can also help you cut to the chase and spend more time connecting with your partner.
Techie note: We mostly used the fast and freely available versions of the LLM models in our tests: GPT-4o mini, Claude 3 Haiku, Llama 3.1 70B, and Mixtral 8x7B. We also experimented with various more advanced versions such as Llama 3.1 Sonar Large, Claude 3.5 Sonnet, and OpenAI-o1. Using the more expensive variants and fancy API tweaks didn’t affect our above findings.