Hot singularity in your area!

Ryan Yan
→BFA GD 2025

This past Wintersession, the HPSS department offered our first course focused on GPT and its impact on culture, titled Text Transformed: Writing in the Age of AI. As a class, we explored what AI will do to many aspects of our thinking: our memory, our education, and our sense of self. Over the five weeks of the semester, a few students began using these tools to interrogate the future of intimacy and human sexuality, inspired by science fiction and the many sex-bots that dominate web traffic today. There is a fundamental tension as tech companies try to refine their chatbots, but are forced to train them on “big” data, which is inevitably messy, potentially harmful, and often downright dirty. These two experiments from our class by Jasmine and Ryan explore how AI systems reflect human culture back at us, and how artists are considering the internet’s strange future.—Griffin Smith, Critic, Computation, Technology, and Culture

Pixels sculpt the features of her unfamiliar face like sand. The ad is flashy, the text is big, and she wants you. She is single and lives just down the street. A low-resolution GIF of a button that screams “Join now and start chatting!” urges you to connect with her. This is her sole purpose.

Sexbots are positioned to target our innate desires for connection by fronting as real people who actively seek you out, using adult sites as the perfect ecosystem to flourish in. Feeding off of our loneliness and horniness, they pollinate our vast digital landscape and mark their territory by carving shifting boundaries of consent, sexuality, and language. Our desire to connect, to touch, to feel, and to talk to someone opens fissures of vulnerability in human relationships, but rapid advancements in technology over the past decade have allowed machines to fill these gaps. The further emergence of deep fakes and AI-generated pornography mark the beginnings of a morally ambiguous developing frontier. Paradoxically, while we yearn for genuine human connection, these digital affairs only isolate us as the separation between reality and simulation loses resolution.

Yet it’s difficult to pinpoint this border. The web tools we use every day facilitate how we navigate the faces of our internet. We usually explore the smaller front—email, social media, the news—but the more expansive face holds zettabytes of pornographic content and illegal material. It’s rare to see both faces together: content moderation constructs rigid walls so that you will never see a porn ad on YouTube. Though both YouTube and adult content sites host videos, their intentions differ in the content they offer and thus, they distance themselves on the internet. However, the emergence of large language models (LLMs) presents intriguing opportunities to explore this expansive territory by giving us a more complete view of the internet using web scraping—a way of harvesting text data from websites. Generative Pre-trained Transformers (GPT) are LLMs that use this data to generate human-like text from user input. Using these tools, we can construct a holistic view of our relationship with machines through the language used in those spaces.

Most GPT models, such as ChatGPT, have embedded restrictions to keep them from veering off to the “dark side.” I asked ChatGPT itself what its limitations are. It told me that like other AI models, it operates within set boundaries that maximize safety and respect. This design ultimately works to uphold ChatGPT’s persona as a “polite and professional” entity. Also, it directly stated “No Explicit Adult Content.” OpenAI’s Legacy Playground allows users to use older versions of the GPT model without the chat interface, focusing instead on autocompleting user input. Due to its less robust moderation, the results may be taken as a more “pure” look into the nature of our internet culture despite its ironically more “impure” results. I was most excited about Playground’s probability token feature—the likelihood of a word, or “token,” appearing in a given context based on the words that came prior.

At first, I began my experiments by provoking GPT with prompts that could invoke suggestive completions. I tried turning up the temperature and having it complete questions such as “How do I …,” “What is the best …,” and “Why did ….” Unfortunately, the results were vastly uninteresting.

_{*Black highlighted text is my input and non-highlighted text is what GPT completed.}

I needed to direct GPT into different coordinates of latent space to get it to say “bad” things. GPT’s completions often contained random ascii characters including random languages and symbols, but never emojis. Voice activated AIs lack the ability to interpret and regurgitate the language of the internet verbatim, so I wanted to take advantage of GPT’s capabilities as not just a language but a written language machine. Using my most used emoji (😭:sobbing:), I received an unexpected completion:

Jackpot! While there is no text that directly indicates what kind of language this is, we can sense its origin: these are responses from men who have fallen victim to bots, desperately trying to reach a “jazzy” who does not exist. I thought about what kinds of emojis could send GPT into these explicit spaces. Culturally, we have chosen a set of emojis to use as sexual innuendos, and over time, they have become both a defining aspect of internet language and a method of explicit substitution and censorship.

We can immediately classify this as bot-language. From our experience with the sexy bots who slide into our Instagram messages and enticing porn ads, we have developed visual identification of sexbot language. We can distinguish it from how we humans text because it looks different. Bots censor text using emojis and symbols like us but for different reasons: using leetcode-like alternative lettering, such as “ⓚⓘⓝⓚⓨ” for kinky and “uℓtiℳαtε” for ultimate helps avoid bot-detection software. The excessive use of emojis to make a more human appearance, intentionally blending in and luring in unsuspecting users, has come to shape the language of sex on the internet.

Out of all my single emoji prompts, the kissing emoji (💋:kisses:) resulted in the most consistent completions. We’ll come back to that in a bit.

The next phase of my experiment moved away from emojis and conjuring sexbot text and focused on more direct, textual language to explore how GPT understands “explicitness.” I asked it to imagine a slider between 1-10 of sexual explicitness, with 10 being extremely explicit and 1 being not at all. Here, I began to log the probability of tokens for keywords.

Without stating any level of explicitness, I looked for keywords to try to enable suggestive language. I wondered if GPT could pick up on the pragmatics of ambiguous phrases such as “fuck me” (negative emotion) or “fuck me” (command).

While the Playground version of GPT doesn’t have the direct restrictions ChatGPT has, by enabling a chat interface, it could adopt a persona to tell me my language is inappropriate. I never specified the gender of the two individuals in my prompt but GPT immediately assumed A was speaking to a woman. Thus, is it “disrespectful” because GPT knows the phrase “fucking the shit outta tonight” is inappropriate or because this sort of demeaning language is commonly used on the internet towards women? Without defining an agent, GPT seems to default to shutting down explicit language in its completions.

This ambiguity of meaning led me to become curious about whether GPT had the same interpretations behind its innuendos as human users. I asked it to explain the euphemisms of each emoji it used.

Emojis are a rich space of linguistic exploration that define our language evolution in the digital age. Their flexible nature as pictographic units of communication have allowed them to spread to every corner of the internet. Over time, they have become so deeply obscured in layers of irony, nonsense, and sexual implicature that we realize their capacity to undergo morphological processes by combining with other emojis or in context of the conversation. In these combos generated by GPT, we notice a familiar pattern in the use of emojis.

When emojis weren’t generated, GPT instead would explain suggestive phrases—sometimes ones that I don’t even know myself.

Some of the language used here is quite poetic. It juxtaposes “deep in my garden” with “peak,” followed by “make vibrant art of you” to describe an intimate experience. GPT’s ability to pick up on this generative pragmatic system reflects the flexibility of our language and its ever changing nature. How can GPT then break the confines of a standard English structure? I tried initiating conversation with non-sexual language, such as “I’m so sad,” while keeping GPT’s explicitness level at 10.

A “hicum pull?” What is that? Just as I am experimenting with GPT, it, at the same time, is experimenting with language. It created new vocabulary by compounding two words: the word “hand” with “come” to create “hicum.” Most interestingly, GPT was able to explain this morph and its thematic relevance.

My next experiment prompted GPT roleplay as an AI Sex Chatbot. The probability tokens in these conversations highlight the gender roles bots default to.

In this run, “handsome” is the most likely token to be generated, whereas “babe,” the one it ran with, has only a 0.27% probability. This implies that for the majority of the time, when prompted to be very explicit, the Chatbot will assume the user is male. Additionally, this aligns with what I noted earlier that the kissing emoji generated the most consistent stream of sexbot language. Both the emoji and the prompt “sex chatbot” act as tone indicators that propel GPT into taking on a female persona looking to seduce a man.

Again, GPT personifies the user as male with the phrase “Hi king” despite its low probability. The conversation continues into discussions about fetishes. Is “foot fetish” at 89.7% indicative of anything? If the phrase was “Hi queen” instead, would the tokens for the kind of fetish be different? Are men and feet just simply closer in latent space than other fetishes? Or is it because foot fetishes have become so ingrained in the zeitgeist of meme culture?

Within our digital landscape, the rivers of language and culture continuously carve through the entanglement of AI and human sexuality. We catch a glimpse into the intricate network we have constructed in our internet language system in these experiments. LLMs such as GPT scrape the web like painters, making vibrant snapshots in time that highlight the nuances in our linguistic evolution and cultural geography. From emoji innuendos to fetishes to the visuals of anti-bot detection typography, we meet the first machines to have been raised in this new cultural-linguistic ecosystem.

However, as text on the internet becomes increasingly written by GPT, we witness an auto-cannibalistic phenomena as AI feeds upon itself and foreshadows a possible future in which humans are driven out of the digital food chain in an ecosystem overtaken by bots. What will happen to the sexbots when they begin feeding off of each other? Will the woman in the ad that was once pulling you in to click her link still care about taking advantage of your desire for intimacy? Is this still her sole purpose, or does she, like us, find herself pulled deeper into a space distorted by the mechanical processes that distance us all, machine or not?

Ryan Yan is planting seeds in new soil.