How does artificial intelligence (AI) work, and why is AI revolutionary and world-changing now? What are the legal conditions for processing large data sets to train an electronic brain? And how about storing user inputs and outputting images and texts? Excerpt from one of my lectures.
Introduction
People who don't understand much about AI often use terms like ChatGPT as a placeholder to sound interesting. Many think that ChatGPT is a search engine. Spoiler: That's complete nonsense. ChatGPT has an ancient dataset by today's standards. That's intentional and deliberate. Because ChatGPT serves as a response machine, not for finding current knowledge.
Those who understood more about AI and had an eye on the stock market bought Nvidia stocks some time ago and have since been able to observe a gigantic price increase. For Nvidia is the manufacturer of graphics cards that are absolute champions when it comes to AI applications.
I predict the downfall of the stock market in its current form because soon anyone can make predictions about stock prices with a probability of over 50%.
My theory along with the assumption that this will soon be accomplished by me.
What's that? Quite simple: A Graphics card like the Nvidia GeForce RTX 3070 has 5888 cores in its GPU. The GPU is the processor of the graphics card. In contrast, the CPU, the classic processor of a computer, stands for it. Good today's Intel processors have 10 or a few more cores.
An Intel core is mathematically something like a Albert Einstein (who could do maths very well as a physicist). A Nvidia GPU core is a moderately gifted mathematician. AI algorithms are based randomly on calculation operations that can be executed particularly well by graphics card processors (GPUs). While the Albert Einstein core performs multiplication with ease and gets bored for half the time, the GPU mathematician is heavily loaded but finishes this trivial calculation almost as quickly.
Unfortunately, 5888 average mathematicians working in parallel would take less time for, say, 100,000 simple multiplications than 10 Einsteins working simultaneously. While the PC equipped with a graphics card has long since finished the K-computation, one thinks that the Intel-only driven PC would have hung up. One can count on a performance increase of a factor of 50 or more with the graphics card. The graphics card is not used for displaying pictures or videos or games, but only for calculating. This is also heard from the loud fan of the card, which can outdo any PC fan.
While in Villariba the CPU is still glowing and only 20 percent of the goods have been unloaded, everything is already shining in Villabaj.
Please excuse the silly comparison with these two fictional villages, which are likely known from advertising, and about which more is known than about current AI algorithms.
The graphics card already makes a quite significant difference for algorithms that are gladly trained 10 days straight for more demanding tasks, or which need 10 seconds on a GPU to generate an image, but take 8 minutes on a – gähn- CPU. You've probably heard of DALL-E or Midjourney and know you don't have to wait 10 minutes for an image.
Functionality of current AI
Systems of artificial intelligence like ChatGPT are based on artificial neural networks. A neural network is also found in the human head or brain. It works roughly as follows:
Shown is the way people process information and how intelligence arises. We understand just as much from looking at the picture why there's even such a thing as intelligence in the first place. I claim that we know nothing about it, but are only amazed that neurons with their connections are able to give rise to something like intelligence. Spoiler: It has nothing to do with God, which I will show soon.
In the above image, on the left we see a series of environmental influences, that is signals. This can be sounds, tones, still images, moving images, smells, air movements etc. Bats are also very familiar with ultrasound. In the middle comes our brain, which receives and processes all these signals. On the right we see the neural network, in which the signals are processed and stored.
A Cell is comparable to a simple processor core. There are connections between neurons, and there are many, very many of them. Billions exist. Whether a neuron fires, that is, is active, is determined by the action potential created by other connected neurons towards a target neuron.
Now we come to the technical realization of today's AI algorithms.
You see the same links in the picture as above with humans.
In the middle you see the electronic brain, above it was the human one.
On the right side of the picture you see the neural network in digital form, which is present biologically and thus more analogically in humans.
So far, so good. But it's going to get even better. The electronic brains transform all signals into number sequences, called vectors, thanks to the Transformer approach (known since 2017). Exactly the same thing does the human brain. At least qualitatively it is the same. That there are fine differences in the general implementation of biology and electronics is nearly irrelevant and only ensures a slight performance boost for biology over electronics. You probably know the Moore's Law: Every 12 to 24 months, the computing power of a processor doubles, often with a simultaneous decrease in price. So the performance winner is the machine, and that's now (around 2023).
These vectors, therefore number sequences that represent images, texts or videos or anything else can now be compared with each other. Now texts can be compared with texts, images with images, videos with videos, texts with images, images with videos, texts and audio signals with images or videos etc. Now you know how image generators like Dall-E or Midjourney work. Now you know that with schnellstens and simplest previously exclusive applications such as SoundHound can be programmed to identify music pieces in just a few seconds.
That it can be done quickly and most easily to program great applications, I have proven myself:
- Audio transcription of my podcasts: Audio converted to text in unexpectedly good quality. For 30 minutes of speech, an automatically generated transcript by AI comes out, where I might only need to manually correct five words. Sentence parts and unknown words like "alles tutti", "Hömmele" (yes, that's what it means!) or "Megafail from Microsoft, which had a Twitter AI tool" are easily recognized.
- Video generation from an image and audio input: My two-dimensional photo in extremely poor quality plus a voice of an American president as only input results in my three-dimensionally animated head and my mouth moving synchronously with the voice as video animation.
- Image generator: Known from other applications. But it makes a difference, if the technician knows whether something is calculated on a Microsoft cloud, Dall-E or somewhere else, or on a local system. The differences are: Local I pay nothing. In the cloud I can accidentally destroy 100,000 euros in a month with wrong programming (example: unintended infinite recursive call). Local I have full control over all data. At Microsoft and Google, praying won't help, but maybe a sedative or pure alcohol to forget about the worries with the data monsters for a short time or even free one's head completely from reason.
- Object Recognition: What objects are on an image? What are the outlines of each object? What is the name of that object? How about "Mark all teapots on the image" or "Find me all images where two or more people discuss in an office and sit on chairs" or "Replace the face on the image with Norman Reedus's, my doppelganger, many say…
- Semantic Search: Instead of searching by keywords or cryptic SQL commands, one now searches over natural language sentences or compares whole documents with each other.
I once had a few thousand pictures calculated on my computer. Here is the result in the form of a mosaic (each mosaic tile is actually an image with a resolution of 512 x 512 pixels):
The individual images of the mosaic were created from a calculation by an AI. The pictures of two artists I know were mixed together. Variety of variations is not yet optimal here, because it was the first attempts. It goes much better, as I noticed a day later. What is half eternity on the AI market is for some data protection authority a time unit they don't know (What is a day? Some authorities only know the time concepts "year", "decade" and "never")?
Local Systems as a Solution
As shown above, many demanding calculations can be performed on one's own computers. Please, don't run again to Microsoft, AWS or Google just because someone wants to play Bullshit Bingo and cover up ignorance with anglicisms and brand names.
Whoever still books a Cloud Service for every little problem is a poor sausage and knows less about AI than about data protection and internet applications. That wouldn't be bad, you don't have to know everything. The wrong advisors are bad though.
An AI is not suitable for exact statements. It's just as reliable as an exceptionally intelligent human.
Some problems are however so complex that they cannot run on standard hardware. For example, ChatGPT Version 4. This system does not even exist as open-source, so no one can be tempted to get it right.
Some are then however reasonable enough and don't dump their business secrets into a Microsoft or Google chatbot. I personally wouldn't want anything to do with someone who does that.
As a particularly plastic example of computational effort is called A Bloomberg GPT model is a type of artificial intelligence designed to generate human-like responses to questions and prompts, similar to how a Bloomberg terminal provides financial data and news. This is a Large Language Model (LLM) of the financial broadcaster Bloomberg. It is so powerful that it took 1.3 million computing hours until the model was finished being calculated. A model is an electronic brain.
The Bloomberg brain was however already finished after 148 hours of computing time because 512 high-performance graphics cards with each 40 GB graphics memory (NOT: computer main memory) were used. Each of these 512 graphics cards costs around €14,000. Whoever wants to upgrade their PC from 16 to 32 GB RAM pays for it out of pocket. Whoever wants to increase a graphics card from 8 to 16 GB storage pays a small fortune (slightly exaggerated).
Legal considerations
I limit myself essentially to some key points that I have taken from the statements of lawyer Jonas Breyer. It is no coincidence that his surname may be familiar to you (keyword: "IP addresses are personal data"). ([1])
Copyright law
Bad and good at the same time for us all when it comes to AI. There is a risk that Europe will remain world champion in regulation. Then hardly any company in Europe will be successful with AI in some fields (image processing?). Instead, we'll buy from our friends over there who do what they want but aren't held accountable (can't be).
Most fundamental premise: What I as a human am allowed or not allowed to do, is equally applicable to an AI.
Drawing a picture from memory is just as allowed for humans as it is for AI. If the result has too much resemblance with a copyrighted work, it's not allowed. All works that have a minimum level of creativity are protected by copyright. That includes almost all pictures or photos showing more than just a square or circle.
Incidentally, it is according to § 44b UrhG allowed to temporarily store works of others in order to analyze them according to patterns. Exactly that does AI usually.
The LAION Case
Here we come to the case LAION e.V., a German association, which was sued by a German photographer. LAION has compiled over 5 billion images in a training dataset for AI image applications and made them publicly available. The dataset only contains links to the publicly freely accessible original images.
The LAION 5B dataset is globally popular and the basis for every second image generator that's worth something.
My own observation after researching numerous Stable Diffusion algorithms.
LAION had downloaded these images once and used them to compute an electronic brain. After that, LAION reportedly deleted the images again. With the brain, images like those above or the one below can be computed.
A photographer has now found some of his pictures in the LAIOn dataset. As I said, only the links and not the originals. The photographer demanded from LAION the deletion of his material from the training data. What exactly is meant by this, I have not yet been able to find out. Maybe it's about the photographer wanting the insights from his images deleted from the electronic brain.
Here is the LAION image database as schema and with original attributes (excerpt):
The question now is whether LAION must delete the findings from short-stored and analyzed images afterwards or not. LAION means "no" and has sent a lawyer's bill with a damages claim of around 850 euros as an answer to the photographer's deletion request.
Data protection law
Personal data is subject to protection under the GDPR. AI applications always involve automated processing. Therefore, they are always subject to the GDPR if personal data is involved in the game.
Personal data can be not only texts but also images. I think that public information can have little or no significant protection of privacy. False statements generated by an AI have nothing to do with privacy protection in the first place. Also, copyright is something other than privacy protection.
So I see data protection as a subordinate problem when it comes to AI applications. It has its very good justification here and prevents, but protects. Otherwise, I see it more so in copyright or trademark law.
Other legal provisions
Trademark law protects explicitly protected trademarks (§ 3 Trademark Act) and implicitly protected trademarks (§ 4 No. 2 Trademark Act etc.).
There are also word and image marks, scents, sounds, shapes, and probably more.
Patents, utility models, and some designs are equally protected. Even a bottle shape can be protected.
The Data Act is currently only in draft form by the Commission. It obliges larger companies to release their data if someone asks for it. This sounds curious and is also.
There is also a draft of a AI-Regulation. In question is whether there should be a reversal of the burden of proof if someone uses a work generated by AI. Was this work lawfully created? To prove this, for example, one would have to save the state of the AI system as a backup when using a picture generator in order to be able to secure the evidence later on. This prevents AI and makes it impossible.
Regulation protects large companies that can afford legal counsel and personnel for extensive proceedings, thereby suppressing small businesses.
Better than regulation, which I consider practically impossible, I find Deterrence and Sanctioning. Two words that are completely unknown to data protection authorities like those in Hesse (and some courts from there). Where do fewer cars park incorrectly? There, where even street parking is fined 800 euros (but never gets a ticket) or there, where one out of five people who park incorrectly get written up?
Recommendations
Use local AI systems. The world can be that simple. No more problems with Google and Microsoft. What was that Privacy Shield again? Is it still there? No, it's not anymore. US is out.
Don't confuse ChatBots with search engines. However, through a semantic search, without ChatGPT at all, every document or image can be found better than ever before. Without OpenAI or Microsoft altogether.
Check the exercise data: Only own or handpicked data is safe from legal problems. Difficult because AI relies on large datasets. Luckily there are ways out. Ideally, only publicly available data should be used if it's not your own.
User inputs should not be stored without reason. They might contain personal data. The reason should, if stored, be carefully selected and justified. After all, ChatGPT was banned in Italy for a short time because of this.
The expenditures generated by an AI vary in form and are handled differently. Text outputs are not scientific reports. See the case of a New York lawyer who was misled into celebrating 12 verdicts from ChatGPT that never existed. Foolish to then submit these results in court and pretend they're true.
I gave a lecture on this topic at a data protection congress hosted by German Air Traffic Control in late May 2023.
Key messages
Nvidia's graphics cards are essential for AI because their powerful processors excel at the massive calculations required for training AI models.
AI systems mimic the human brain by using artificial neural networks to process information and learn from data.
These networks, made up of interconnected "neurons," allow AI to understand and generate text, images, and other types of content.
Powerful AI tools can be used effectively on personal computers, offering advantages like cost savings, full data control, and privacy.
Don't blindly trust AI and cloud services for every problem. Understand their limitations and potential risks, especially regarding data privacy and copyright.
The main point is whether using copyrighted images to train AI, even if they are publicly available and deleted afterwards, violates copyright law.
AI regulation is complex and may unintentionally harm smaller businesses while favoring larger corporations.
Focusing on deterrence and sanctions, rather than extensive regulations, could be a more effective approach to responsible AI development and use.
My name is Klaus Meffert. I have a doctorate in computer science and have been working professionally and practically with information technology for over 30 years. I also work as an expert in IT & data protection. I achieve my results by looking at technology and law. This seems absolutely essential to me when it comes to digital data protection. My company, IT Logic GmbH, also offers consulting and development of optimized and secure AI solutions.
