Until not many years ago, artificial intelligence seemed like something out of a utopian story. Commonly described as “systems and tools” to “help” and “improve” the quality of life, we have missed something so essential to their creation: their education. Yes, as if it were a child, the way you train an AI will say a lot about its futureand this intelligence trained in 4chan shows us how not to educate them.
The youtuber Yannic Kilcher decided to train an artificial intelligence using 4chan as a model. Specifically, he used the board politically incorrect (Politically incorrect) to generate a creation unthinkable even by the devil himself. Of course, this movement is not without controversy, as a lot of people have already come out against this idea.
The result was terrible, but not because it went wrong. In fact, the AI was so successful in learning from 4chan that ended up creating about 15,000 publications of racist content up to date on the board.
How Kilcher’s AI model works
To get started, we’ll give you some background on what the Politically Incorrect (/pol/) board is in 4chan. In this space the users are dedicated to posting racist, misogynistic, anti-Semitic, LGBTphobic content and on many other unfortunate topics. Thus, Yannic Kilcher decided to create an AI that he ended up naming GPT-4chan (inspired by the OpenAI GPT-3 model).
This artificial intelligence model not only learned to detect the words used within the Politically Incorrect board. In addition to this, he managed to capture the tone used by its users. “It perfectly encapsulated the mix of offense, nihilism, trolling, and deep distrust of any information that pervades most messages on /pol/,” Kilcher said, “proud” of his creation.
Once Kilcher finished training his AI, he decided to release it onto 4chan in the form of multiple bots. With them, the model managed to reach up to 15,000 posts in a single day. Yannic took special care not to be detected by the bots on the platform, even going so far as to use VPN services to spoof the region from which they were posting any of these computer messages.
Although the creation of youtuber has been quite criticized by internet users and AI specialistsKilcher has taken the subject as something of a joke. After all, it is content that users of 4chan they’re used to, so people didn’t really notice anything weird on the board.
The harsh criticism against GPT-4chan
Although the idea may seem quite striking, has not escaped controversy. In fact, Lauren Oakden-Rayner, director of medical imaging research at the Royal Adelaide Hospital and senior researcher at the Australian Institute of Machine Learning; has commented on it.
“This experiment would never pass a human research ethics board,” commented the director in a publication of Twitter; further adding that “Open science and software are wonderful principles, but they need to be balanced against potential harm.”
“Medical research has a strong ethical culture because we have a terrible history of causing harm to people, generally from unprotected groups. It has carried out experiments on humans without informing users, without consent or supervision. This violates all the principles of research ethics. research with humans.
Dr. Lauren Oakden-Rayner
Oakden-Rayner expands on her idea on one of the talk pages for GPT-4chan, adding that “There is nothing wrong with making a model based on 4chan and test how it behaves. The main concern I have is that this model is freely available for use.”
The harsh moral dilemma proposed by this AI
The model developed by Yannic Kilcher is undeniably a very interesting one. After all, seeing how an AI is able to learn, and showing that the sources used in its training matter, is something that can be both useful and fun. See the different scenarios in which she is “maid”to then see it in action in the real world.
Here, there is no doubt that Kilcher did a good and interesting job. However, certainly Leaving this type of technology available to anyone can be potentially dangerous. In a world that lives more and more on the internet, the spread of hate messages becomes a really dangerous factor for people belonging to communities that have historically been discriminated against.
Just imagine that someone is dedicated to using these trained bots on 4chan to sabotage any campaign for the human rights of certain communities. Seeing as it has the ability to generate at least 15,000 posts of pretty dark content in a one day period, could end up being used to the detriment of these social movements. Thus, these movements could be totally invisible, and even lead to a part of the population rejecting them completely.
All against GPT-4chan
Of course, despite these dangers, these artificial intelligence researchers they might be getting a little carried away by internet activism. After all, it’s not the first hate-message bot to exist across the web, and it certainly won’t be the last.
For its part, after warnings from researchers, Hugging Face has decided to stop external downloads of the model. However, while these AI experts hoped that the platform would remove the bot entirely, Hugging Face has refused to do so.
“After a lot of internal debate at HF, we have decided not to remove the model that the author uploaded here under the conditions that:
1. The model file and the video clearly warned of the limitations and problems posed by the model and the POL section of 4Chan in general.
2. The inference widget was disabled so as not to facilitate the use of the model.
Clement Delangue, CEO of Hugging Face
In addition, Delangue has publicly stated that the model developed by Kilcher has been “useful for test what a trained model could do with that data and how it performed compared to others (in particular, the GPT-3), and that it would help to draw attention to both the limitations and risks of those models.”
Kilcher is right. After all, it is necessary to know the dark side of things to be able to manage them. Better to know how a relatively “primitive” AI trained in this field might behave than to wait for it to happen a hundred years from now and take us by surprise in the best Skynet style.