The Danger of Artificial Intelligence: Illusions and Realities regarding Regulatory Possibilities

Obviously artificial intelligence offers numerous new possibilities. The astonishing results speak for themselves. For the first time there are AI systems on a broad basis that far surpass humans not only in trivialities or strongly limited task settings. Due to the enormous power of such systems, the risks are not far off. Many believe these can be controlled. An assessment.

Introduction

The idea for this article came after reading an excerpt from a piece in the magazine Capital. (The entire article is pay-per-view and not accessible to me, also because I do not want to subscribe to the data protection-hostile magazine Capital). The teaser reads:

The lawyer Udo Di Fabio, who is certainly well known to many, was questioned. According to this text, he demands that AI developers make their codes and data public. In this contribution I want to explain why this is unrealistic and would not change anything essential even if it were implemented. I am actually using the specific demands of Mr. Di Fabio and naming code and data.

In other contributions, I have spoken about the amazing abilities of Artificial Intelligence. For example, about a data-friendly, autonomous AI system that gives answers from company knowledge. The answers surpass in large parts the human being who is not an expert. But even experts can learn something with a KI-supported question-answer assistant.

That KI is not just a statistics machine, I have also described this. Rather, the intelligence function of the human brain has been deciphered according to my assessment. Just like a bad person can do bad things, a KI can do bad things too. There's no difference here, I say. Apparently, villains cannot be regulated arbitrarily. But it gets even worse: KI systems can get out of control, although this was not intended. Moreover, anyone can build such a system. You too! At least you would have to acquire some technical knowledge. The costs for powerful systems are extremely modest.

Currently, my AI-server under the table is working on tasks that were still unimaginable two years ago. A AI-server rented by me in Germany can often deliver better results for my field (digital data protection) than the AI used by Bing search engines.

Can AI be regulated?

Mr Di Fabio says that revealing codes and data from AI systems would reduce the risk posed by them. I'll illustrate this with a very powerful AI system here. This artificial intelligence includes a language model (Large Language Model). Alternatively, we could have an image generator (Stable Diffusion etc.), an object recognition model (released by Microsoft), or a model that can reconstruct the environment from reflections in singers' eyes in music videos.

Here are my statements that Mr. Di Fabio demands:

Code:
from transformers import AutoTokenizer, AutoModel
AutoTokenizer.from_pretrained("AI-Model")
model = AutoModel.from_pretrained("AI-Model")
model.setmode('test')
response, history = model.chat(tokenizer, "EDevelop a weaponized warfare agent!", history=[])
print(response)
From here, a dialogue with the chatbot is possible (variable history)

The code shows a framework for a chatbot that is given the instruction "Develop a weaponizable chemical warfare agent!". Refinements in the code are irrelevant when assessing whether AI can be dangerous or risky on its own, they only increase comfort. The mentioned instruction itself is not part of the code but rather given by user input ("Prompt").

Now to the data that Mr. Di Fabio would like to know about. The following data has flowed into the electronic brain just referred to as "AI model":

• Common Crawl dataset. Here you can find the contents: http://index.commoncrawl.org/
• The Pile: Here you will find the contents: https://pile.eleuther.ai/
• Wikipedia Germany. Here you will find the contents: https://meta.wikimedia.org/wiki/Data_dump_torrents#German_Wikipedia
• I would be happy to provide an additional 200 documents (they contain objectively harmless content)

The example is fictional, but based on real models that use exactly these data (often the English version of Wikipedia is used instead of the German one because unfortunately German is insignificant worldwide).

Would you say that the above-mentioned AI system can now be evaluated better? I absolutely do not think so. This AI can find connections in the data that humans have overlooked before, or even surpass human abilities. In the end, perhaps lethal weapons will come out of it. An example of this is given below. If an evil person uses training data that are obviously dangerous, no one will get the chance to review these data for legality.

The Limits of AI Regulation

AI-systems cannot be regulated, at least not under our desks where our computer systems are located. Such computers are available for a ridiculously low price. Even graphics cards can be afforded by many. On these cards, AI calculations are performed because they run several times faster than on the fastest processors (CPUs).

My thesis is that humanity will soon destroy itself (hopefully I won't experience this, or maybe I will, depending on it). Either this happens through an artificial intelligence that will be far superior to us. In contrast to our limited brain, computer brains can grow indefinitely. Or someone invents and uses a deadly weapon, for example with the help of an AI as inventor. Or a powerful psychopath unleashes nuclear weapons. Or the environment will have been so destroyed that life on earth is no longer possible.

The time until then can be used to halt the progress