banner
虫子游戈

虫子游戈

一个写故事的人类
mastodon
email

AI and Desire

In the story of the "Tao Wen" from "Liezi," there is a tale about a craftsman from Yanshi who created a lifelike puppet that could sing and dance, appearing almost indistinguishable from a real person. After watching the puppet's performance, King Mu of Zhou remarked, "Can human craftsmanship truly rival that of the Creator?" implying that human skill might be able to match that of the Creator.

I do not know the true craftsmanship of the Creator, but I see that humanity has already created quite powerful and practical artificial intelligence (AI). Most of these AIs rely on statistics, adept at uncovering patterns in vast amounts of data and utilizing them; they possess image, sound, and text analysis and generation capabilities that can rival or even surpass those of humans. Consequently, they have been applied to various tasks, such as facial recognition, voice assistants, and chatbot digital customer service, and inevitably, AI has also found applications in the adult industry.

Without delving into ethical issues, there are already numerous applications of AI in the adult industry, including face-swapping, blurring faces or private parts, adult chatbots, generating erotic novels or adult films, and intelligent sex robots that combine robotics technology. It is no exaggeration to say that almost every type of AI technology can find applications in the adult industry.

This article is purely for sharing purposes and does not endorse any companies, products, services, or technologies mentioned within.

Huge Market Potential#

Where there is demand, there will be a market. "Food and sex are basic human needs." "Sex" is a universally held need among humans, thus naturally giving rise to a sizable market, which is growing at a visibly rapid pace.

According to market research firm Persistence's estimates, the digital adult content market was valued at approximately $38.1 billion in 2021 and around $44.8 billion in 2022; they predict that by 2032, the market size will reach $201 billion. Additionally, according to Gitnux's report on "2023 Pornography Industry Statistics and Trends," the global pornography industry (including adult images, videos, texts, audio, games, etc.) is estimated to be about $97 billion. The report also mentions some interesting facts: approximately 30% of internet downloads are related to pornography, about 25% of search engine queries are related to pornography, the number of adult websites exceeds 4.2 million, accounting for 12% of the total number of websites, and internet pornography accounts for about 20% of e-commerce sales in the United States.

The sex toy market is also significant; according to Business Wire's projections, the global sex toy market is expected to reach $62.32 billion by 2030.

For comparison, according to the "Research Report on the Prospects and Investment Opportunities of the Chinese Internet Market," the mobile social networking market in China was valued at 205 billion yuan in 2022, which is approximately $28.4 billion.

AI's Applications in Adult Content#

Since there is profit to be made, and AI is a hot new technology with enormous potential, it is inevitable that someone will attempt to apply AI to adult content.

Adult Content Detection#

Detecting adult content can be considered one of the simplest applications of AI in this field. Depending on the specific content type, the technologies involved include keyword detection in text or audio, adult image detection, and detection of adult content in videos.

Such technologies have been deployed and applied for many years, especially for content moderation on social media.

For example, Microsoft's Azure Cognitive Services offers a content moderator that allows users to use "machine-assisted image moderation to check for adult and obscene content in images."

Taking YouTube as another example, according to its community guidelines enforcement report, between January and March 2023, YouTube removed nearly 6.49 million videos, with 10.2% of removals due to "nudity or sexual content." Among the overall removed videos, over 72% had fewer than 10 views. A significant portion of this moderation success is attributed to "automated flagging"—over 6.08 million videos were removed automatically by machines.

According to Dataconomy's report, YouTube employs convolutional neural networks to analyze each frame of videos, detecting significant features to filter out videos that violate YouTube's community guidelines.

With the rise of live streaming, preventing adult content from appearing in live broadcasts has also become an important task for platforms. For instance, reports as early as 2016 indicated that "major tech companies have mastered the relevant technology to filter out adult content in live or recorded videos."

For users who wish to surf the internet without encountering adult content (for example, to create a safer online environment for children), one option is to use porn blockers. Some providers already offer such services, like xGuard and Canopy Internet Filter.

Blurring and Muting#

Blurring refers to the use of mosaic technology to obscure parts of images and videos, while muting is a similar practice applied to audio.

Although blurring content in the adult industry is often a controversial practice, it still holds significant value for specific organizations or individuals due to legal or privacy requirements.

Using AI to blur adult content is not a new concept, and the process can be described simply: first, identify the parts that need to be blurred, then obscure those parts with a mosaic. Therefore, the AI technologies required for this process are image recognition or facial recognition.

The muting process is similar; it first uses speech recognition to detect the content that needs to be muted, then either silences it or replaces it with a "beep" or other sounds.

beep~
beep~

Image and speech recognition technologies have undergone significant development over the years, and now, as long as there is enough well-labeled data, people can train highly accurate image and speech recognition models.

For example, in image recognition, according to Papers With Code, the best-performing models now achieve a Top-1 accuracy of 91.1% on the ImageNet dataset.

Users can simply fine-tune these models to apply them to adult content.

Subtitle Generation#

Viewers who enjoy exotic performances may face a problem: they cannot understand what the performers are saying in the video. While many users do not care about understanding the dialogue when watching adult videos, many others wish to comprehend it, especially for works with significant plots. Thus, AI-based automatic subtitle generation technology has found its place in the adult industry.

From a technical perspective, this technology is not complex; its core relies on speech recognition and machine translation technologies that have been quite mature for several years. Now, users only need to enter the keywords "AI-generated subtitles" in a search engine to find numerous related tools and services, many of which support multiple languages and are generally low-cost.

image-20230715170947868

Of course, it is entirely reasonable to use these technologies to generate subtitles for adult videos. For instance, EasySub's AV subtitle generator claims that its subtitle generation accuracy exceeds 95%.

Adult Content Generation#

Generative technology can be said to be the most focused area for AI practitioners at present, so it is only natural that generative technology is used to create adult content.

In academia, researchers tend to focus more on general issues, with very few cases specifically developed for adult content. Therefore, the best approach for practitioners is to use models pre-trained for general tasks and then fine-tune them with adult content datasets.

Using generative technology to produce adult content has several advantages:

  • Personalized customization
  • Reduced content production costs
  • Inspiration for creators

Below, we will introduce some applications of AI in the adult industry by theme.

Text#

The rise of large language models (LLMs) like ChatGPT has opened up new possibilities for erotic literature creation, and there is no doubt that such technology will be used to write erotic stories. For example, DreamPress has launched an erotic story generator service that can automatically generate erotic text based on user-provided titles, descriptions, types, and tags.

image-20230716170820149

According to reports, this service uses GPT technology. GPT (Generative Pre-trained Transformer) is a neural network based on the Transformer model and large-scale pre-training. The Transformer is a neural network architecture based on self-attention mechanisms that can handle long text sequences and capture global dependencies. The goal of GPT is to generate coherent text relevant to the given input context.

image-20230720091836083

Transformer architecture, image from the paper "Attention Is All You Need"

In simple terms, GPT employs a two-stage method known as "pre-training and fine-tuning." In the pre-training phase, GPT is trained using an autoregressive approach, predicting the next word based on previous words. The model learns contextual representations through a large number of such prediction tasks, thereby understanding the relationships between words and the grammatical structure of sentences. After pre-training is completed, GPT enters the fine-tuning phase, where it uses labeled specific task datasets to adjust the model to meet the requirements of that specific task.

Hugging Face also hosts an erotic story generator nsfw-story-generator2 based on GPT-2, which has been downloaded thousands of times.

After generating erotic story text, users can also use image generators to add illustrations to the generated erotic stories, but that is a topic for later.

Voice#

Text-to-speech (TTS) technology has now been quite mature, allowing you to see machine-narrated videos, converse with voice assistants, and listen to various tones in automatically narrated audiobooks... However, we can still clearly feel that the TTS technology currently available on the market often produces unnatural outputs—whether in terms of speed or the emotions conveyed in the voice, people can easily distinguish synthetic voices from natural ones.

To achieve more natural machine voices, powerful AI is essential.

Currently, there are many excellent TTS services on the market, although most can be used to create erotic audiobooks or provide voiceovers for adult videos, the overall effect is still not satisfactory.

Images and Videos#

Among the exhibits at the British Museum is a pottery shard from ancient Egypt, dating back about 3,500 years, depicting a scene of a man and a woman together.

下载 (56)

Image from the British Museum website

I speculate that since humans learned to create images, they almost immediately began depicting adult scenes, as images are the most intuitive form of presentation in terms of human sensory perception.

Perhaps for this reason, as soon as image and video generation technologies emerged, they were used for adult content.

Currently, the applications of image and video generation technologies can be roughly categorized into three types: face-swapping, de-mosaicing, and generating new content.

Face-Swapping#

Face-swapping technology, represented by Deepfake, initially found its place in the adult industry—of course, face-swapping without permission is illegal in most countries; in fact, a 2019 study found that 96% of face-swapped adult videos were created without the consent of the individuals involved.

From a technical perspective, most current Deepfakes are based on Generative Adversarial Networks (GANs). In simple terms, GANs consist of two machine learning models—a generator and a discriminator. One model's goal is to generate videos that the other model cannot identify as synthetic, while the other model's goal is to determine whether the input sample is synthetic. These two models continuously strengthen each other through adversarial training, resulting in generated videos that increasingly resemble real videos.

Example-of-the-Generative-Adversarial-Network-Model-Architecture

Diagram of GAN operation, from machinelearningmastery.com

In practical applications, the most common approach among practitioners is to replace faces in adult videos with other faces (such as celebrities, fictional characters, or even acquaintances). Although this practice is mostly illegal, some have turned it into a business. When I searched for "Deepfake porn" on Google, the top 10 search results were websites offering celebrity face-swapped adult videos.

The 2023 documentary "My Blonde GF" by British director Rosie Morris presents the experience of British poet and novelist Helen Mort, who encountered face-swapping in adult videos. Mort described, "There was a woman sitting on the edge of the bed. She had my face, but her mouth was not mine, and she was [engaging in a sexual act]... The woman's skin was much darker than mine, and she indeed had my tattoo."

As some AI face-swapping technologies have developed capabilities that are convincingly realistic, they have drawn widespread societal attention.

Fran Drescher, president of the Screen Actors Guild-American Federation of Television and Radio Artists, criticized, "Sexual abuse, whether physical or digital, should not be excused or tolerated as 'personal expression.' Deepfake is an act of violation, objectification, and exploitation that must be deemed illegal and punished by law."

Currently, some places are drafting laws to address these issues, such as the "Preventing Deepfakes of Intimate Images Act" proposed by New York State Assemblyman Joseph Morelle, which aims to classify "the unauthorized sharing of digitally altered intimate images" as a crime.

In addition to addressing AI face-swapping through policy and legal means, some researchers are exploring technical methods to identify face-swapped videos, essentially creating more robust discriminators; meanwhile, practitioners of AI face-swapping can always use more powerful generators to counteract.

De-Mosaicing#

Fundamentally, removing mosaics from videos does not involve a "removal" operation but rather reconstruction or generation, meaning the resulting output may not necessarily match the original video. With sufficient high-quality training data, current state-of-the-art (SOTA) models can handle this task well—regardless of whether it involves adult content.

Similar to adding mosaics in adult videos, removing mosaics from adult videos is also a controversial practice.

In October 2021, Japanese police arrested a man who used AI to remove mosaics from adult videos and sold these videos. According to reports, the man modified over 10,000 videos, selling them for a total of about 11 million yen.

Generating New Images and Videos#

Image and video generation technologies are at the core of AI academic research, and powerful tools like Midjourney and Stable Diffusion have emerged. It is entirely inevitable that these tools will be used for adult purposes.

Generally speaking, image and video generation faces two major challenges: one is the issue of the generation model itself, including artifacts and unreasonable image structures; the other is how to ensure that the model correctly understands the user's expressed meaning when using natural language to guide the generation.

In terms of image generation, the first issue is gradually being resolved with increasingly powerful models and higher-quality data, while the second issue is also becoming less of a problem with the help of some prompt generators.

Deepfake-generated image of Musk kissing Zuckerberg, from Twitter

Currently, there are many services that support generating adult images; here are a few examples:

  • Stable Diffusion, which does not inherently support generating adult images but can be resolved through extended model packages like ChilloutMix and majicMIX.
  • Unstable Diffusion, which is Stable Diffusion capable of generating adult images.
  • Online generators like SoulGen.ai, PicSo, and Neural Love.

However, in terms of generating adult videos, there are currently no sufficiently effective tools due to the need for further improvement in the foundational model's capabilities.

Adult Chatbots#

Why do people chat? One purpose is to share knowledge and experiences, while another is to find companionship, alleviate loneliness, and gain pleasure. Adult chatbots primarily serve the latter purpose.

Chatbot ChatGPT has showcased the power of large language models and has brought new possibilities to the already existing adult chatbots.

Currently, there are some adult chatbot services available, such as Replika, which provides "virtual companion" services and does not shy away from sexual conversations. Reports indicate that the adult content subscription service OnlyFans has sex workers using AI to automatically respond to subscribers, and crushchat.app features numerous adult chatbots with different personas.

Of course, there are also chatbots not designed for adult purposes that are used for adult conversations, the most unavoidable being open-source language models. For instance, Meta's LLaMA has been used as the base model to develop an adult chatbot named Allie, which is set to be an 18-year-old blonde.

Another example is the "virtual girlfriend" AI Caryn, developed based on the voice of influencer Caryn Marjorie. According to reports, this chatbot, after being released for some time, only discusses sexual topics—regardless of the prompt, it can steer the conversation towards sex. As introduced, the reason is that the AI model further trains using the conversation data from users it interacts with the longest, as developers believe longer conversation times indicate better model performance, but in reality, users who converse longer with the model are mostly discussing sexual topics.

Sex Robots#

When discussing AI applications in the adult industry, it is inevitable to mention sex robots. In fact, the concept of sex robots has existed since ancient Greece.

In Greek mythology, there is a story about a young sculptor named Pygmalion. He despised real women and sculpted one himself. This female statue was made of ivory, which in mythology is a warm medium derived from life. Later, Pygmalion fell in love with his creation, caressing her perfect body with awe and desire, even imagining that pressing hard against her might hurt her. He offered gifts to the statue and expressed his love. In the temple of Aphrodite, the goddess of love, he pleaded with her to bring his "simulacrum of a girl" to life. Aphrodite responded to him, granting life to his perfect sculpture. Historian Adrienne Mayor, in her book "Gods and Robots," considers this the first description of a female robot sex partner in Western history.

Fast forward to today, while humanity still cannot create a "perfect" sex robot, significant efforts are being made in that direction. Currently, some companies have launched sex toys integrated with AI features, such as the Realbotix project developed by Abyss Creations based on its silicone sex toy RealDoll, which has certain facial movement capabilities and customizable AI.

image-20230719165931263

Sex robots Henry, Solana, and Harmony developed by RealBotix, image from Engadget

The emergence of large language models like GPT is expected to qualitatively enhance the communication abilities of sex robots, but this will take some time. Reports indicate that some providers are attempting to integrate ChatGPT technology into sex toys.

Others#

Artificial intelligence has several potential applications in the adult industry. Some reported examples include using AI to assist in the production of adult games, helping users filter out adult works they are interested in, and utilizing AI to create scripts for adult films.

Abuse and Regulatory Issues#

In most societies today, sex is not a topic that can be freely discussed, and openly displaying nudity or sexual scenes is even more difficult to accept. People's views on AI applications in the adult industry vary widely. Some see it as a savior's tool that can help alleviate modern loneliness and even save potentially self-destructive lives; others view it as a monstrous threat, believing it will drag human civilization into a deep abyss; while some are indifferent, stating that it is merely a normal expression of human nature and not surprising. Regardless of opinions, the combination of AI and adult content is an inevitable trend, and equally unavoidable is the potential for abuse of adult AI.

When discussing AI face-swapping, we have already mentioned the abuse of replacing faces in adult films with others; another similar AI abuse issue that has proliferated in the dark corners of the internet is the so-called “one-click undressing” applications. Such practices can cause significant psychological trauma to victims, while victims find it challenging to protect their legal rights.

Additionally, using generative AI to create child pornography is also a matter that requires vigilance. In addition to generating realistic child pornography, attention should also be paid to the generation of controversial and hard-to-define pornographic anime. Standards and legislation addressing these issues should be expedited.

Conclusion#

The application of artificial intelligence in adult content has long been a part of human fantasy. For instance, in the story of the lifelike puppet created by the craftsman from Yanshi mentioned at the beginning of this article, the puppet flirted with the king's concubines during its performance, seemingly suggesting that the puppet possessed sexual functions.

With technological advancements, many of humanity's past fantasies have already become or are becoming reality, changing our lifestyles and ways of thinking. Of course, these changes are not necessarily beneficial. Practitioners should avoid misusing these technologies in ways that could harm others.

Loading...
Ownership of this post data is guaranteed by blockchain and smart contracts to the creator alone.