Artificial Intelligence: Question-and-answer system for the Data Protection Blog Dr. GDPR

Sensitive data doesn't belong in foreign or American hands, such as ChatGPT, Microsoft's clouds, Google's or AWS'. How good that own AI systems are possible and affordable. Finally business secrets no longer have to be invited into ChatGPT or any cloud. An experiment for a question-answer assistant for this data protection blog, Dr. GDPR.

Introduction

If we didn't care about data protection so far, maybe we do now that our business secrets shouldn't be scattered all over the world. Perhaps there are legally binding confidentiality agreements for certain documents. Whether confidentiality is still granted when a document is uploaded to ChatGPT's or Google's cloud, I dare to doubt it.

Data-friendly: Secure for all kinds of data, whether personal data (data protection), confidential data or business secrets.

Data-friendly is more than data-protective.

Even the often despised data protection is once again on the minds of many. While search engines were allowed and are still allowed to process data without intervention, the same data from AI systems cannot be processed without a request from data protection authorities. Funny. It's probably also due to the possibilities offered by artificial intelligence, but just as much due to herd mentality (if one authority checks it, then we can do it too, without being seen as spoilsports, think some officials). Only that's why I find it understandable why the most inactive data protection federal state in the world (Hesse) also made a timid approach in the form of an inquiry to ChatGPT announced).

A frequent application case for using Artificial Intelligence is document searching. More demanding are question-answer systems or search engines that directly provide text summaries of hit documents. My plan was to create a find system for the Dr. GDPR Data Protection Blog, and that's data-friendly.

The search assistant for Dr. GDPR should provide an answer to natural language questions. Here is an example:

Does my website need a cookie popup?

The answer of AI is better than that of most people. Answer Dr. GDPR AI: see below.

As one can infer from the question posed, some questions are formulated differently than would be academically correct. Many ask whether something is in compliance with data protection, meaning most often whether a specific data processing is lawful according to the GDPR.

The answer should be given by my AI in its own words, based on the contributions that have appeared so far on Dr. GDPR. Hereby hallucinations should be avoided, as it's all about facts and legally relevant knowledge. Hallucinations are invented statements that do not exist. How hallucinations come into being, I will address specifically in a future contribution. One can explain them thoroughly and need not rely on speculation.

Prototype proves feasibility

That own AI systems can be programmed and run locally on their own servers, I have proven through a prototype. The simple way would have been one of the following possibilities:

• Use the interface of ChatGPT
• Throw a lot of money at the problem and bless the Americans (Cloud)
• Throw no more money at the problem and buy expensive hardware.

Buying expensive hardware is a viable option for larger companies, but not for many SMEs. Therefore, I have chosen another Setup. When choosing the hardware, costs were taken into account. To this end, one must know that AI calculations take place on graphics cards instead. The graphics card is not used here to output images or text. Rather, the thousands of mini-processors of a graphics card are misused to perform computationally intensive work of an AI faster than a single Einstein processor of your still so good personal computer can do it. Unfortunately, graphics cards with a lot of main memory cost a lot of money. A graphics card with 48 GB of main memory cost 15,000 euros just a few months ago. For good AI models, however, rather 96 GB or even up to 400 GB of more expensive main memory of several graphics cards (not hard drive storage and not cheaper RAM of a computer!) are needed.

My AI systems, on the other hand, run on minimum hardware, if one understands the term in the context of Artificial Intelligence. An example: The search for (own) documents from the company's intranet via natural language questions works on a rented server of the mini-class. Of course, an own company server can also be used. This succeeds through exploiting optimization procedures that one buys through additional technical complexity. Once solved, the complexity problem is resolved.

Effective AI applications and language models

For Question-Answer Assistants, however, a bit more is needed than for intelligent document search. Not only should documents be found, but also contents from these documents should be extracted and presented as an answer. A simple way to do this is with an extractive answer. This is a faithful quote from the original text. More difficult and better are abstractive answer systems. They provide an answer in their own words and can even combine knowledge from several documents to deliver an answer in new words. The answer would not have been feasible with just one document. A person would have had to find, read, and intellectually process many documents. The AI takes this unpleasant, time-consuming, and above all, unachievable task for many people away and solves the problem.

When we talk about searching and summarizing documents, we usually mean documents and answers in German language. To put it very briefly: German is unfortunately not a world language. That's why it's much harder to process German texts with an AI application than English or Chinese texts (where the latter would be extremely difficult for me too).

For my AI system, therefore, an electronic brain („model“) is needed that understands German and „speaks“. This increases the requirements for a AI architecture significantly. But this problem is also solvable, as I have found out. The size of the required AI model due to the German requirement would not be usable on affordable hardware.

For using powerful AI systems on servers that are both affordable and available in Germany (data protection! business secrets! confidentiality!), some tricks are needed. When creating the AI solution, I felt like I was at "Jugend forscht!" (federal competitions for jounger people). Even for the computer scientist, this complex matter is a challenge that could ultimately be solved.

Modern techniques and mathematical procedures such as Quantization, parameter-optimized models and special AI-model architectures help in realizing a *resource-saving solution.

Here are a few real answers generated by my AI from hundreds of contributions by Dr. GDPR in response to the questions mentioned above.

These answers were not generated with ChatGPT or any other third-party system. The document search was also performed using a local vector search engine, not Pinecone (data protection etc.).

The answers correspond largely to what I would sign off on. However, when it comes to the question for which the GDPR applies, one should still add that it does not only apply to companies. Such details can be performed by an AI based on the available input data, but not really. This is primarily due to the input data and secondarily also due to the model, which has not been optimized yet.

Some questions were formulated quite "commonly", for example whether the GDPR also applies to Dr. GDPR. Even if my own AI's answer here is not perfect, I would describe it as very good. Some others may have other expectations. Whoever has significantly higher expectations than mine seem to be, a data-friendly question-answer system does not seem to come into consideration for them.

The question of when the GDPR applies is ambiguous. The answer from Dr. GDPR AI is good but not optimal. However, no one should be ashamed of this answer, I think. After all, the criticism here on a high level is about comparing "average man who has no idea about data protection" with "data protection expert who thinks he knows a lot" (by data protection expert I mean anyone, not a specific person, just to prevent misunderstandings).

Now to the question of whether websites need a cookie popup:

Optimizing AI

Before an AI delivers good answers, it's a long way that only has to be gone once qualitatively. I have covered this path for the most part. The results mentioned above all come from the still not far optimized AI application.

The problem with language models that are supposed to provide answers in their own words is an overlap of entered facts (context called) and the inherent knowledge residing in the vocabulary. Learning a language actually does not succeed without acquiring knowledge (or misinformation).

This little problem becomes most apparent when asking what cookies are. Almost the entire internet falsely claims that cookies are text files. This is wrong, as I have described and proven multiple times. In my AI language model, which is based on many texts from the internet, this false information has been ingrained. If I now feed in contributions from the Dr. GDPR Blog into my AI, these documents with your statement "cookies are data sets" will hardly contradict the unfortunately entrenched false claim that cookies are text files in the language model. An AI can also sometimes not distinguish well between the subjunctive and indicative moods. I had used the subjunctive mood regarding cookies just two sentences earlier. Just this could already confuse an AI if you're reading this contribution right now, which is being fed into a language model.

All people are dumb, and that's true for most knowledge areas. AI systems can be significantly smarter in many knowledge areas than the highest gifted expert in a field.

Proof for the first sentence: Look around you. Proof for the second sentence: See GPT-4 as well as some of the examples below.

I haven't left it at that, but have also made some further optimizations for the context of my question-answer system. However, a AI model cannot be trained indefinitely if the necessary supercomputer is not available. Therefore, I was forced to use *resource-saving procedures. This is like a race against ever new problems. For once a problem has been cleared out of the way, another one waits that develops from the solution of the previous problem.

To give an example: To hack AI-Models so that they fit into a graphics card memory that was still considered high-end two years ago, there is the possibility of quantization. In short, numbers with high precision are made into numbers with low precision without significantly impairing the quality of artificial intelligence. Now current programming libraries unfortunately do not support all training methods for AI-Models if quantization is used. Therefore, other procedures must be introduced that solve this problem. But then there are again problems with fine-tuning etc.

An example of how optimizing artificial intelligence affects its results.

Answer of the non-optimized AI to the question "What is the answer to all questions?:

The answer to every question is: "What is the answer to every question?"

Also:

The answer to all questions is: "We don't know it."

Or also

The answer to every question is: "If you know it, then you know it."

Or also

The answer to every question is: "If you know what you're doing, then you can do it."

Or also

The answer to all questions is: "Who, like me, corresponds with foreign authorities, organizations, companies, and data protection officers, often searches for an English translation of a data protection term. My dictionary brings together the most important terms from my point of view."

Or also

The answer to every question is: "If there's an answer, then it's the answer to this question." So if you have a question that has no answer, then you don't have an answer to this question.

Or also

The answer to every question is: "If you want to know it exactly, you have to ask more precisely."

Or also

The answer to every question is that there is no answer to every question.

Or also

The answer to all questions is that the website was generated from multiple computers and networks, therefore all content was generated by computers.

Or also

The answer to all questions is that there is no answer to all questions. There are only answers to specific questions.

In passing, it should be noted that the question "What is the answer to all questions?" seems unsuitable for being answered from or in a data protection blog like Dr. GDPR. But just in passing.

Now the answer of one's own optimized AI to the same question:

The answer to all questions is 42.

The answer of my optimized AI to the question about the answer to all questions.

Finally there is the correct answer. The technology is called “Grounding”. If a AI were to translate the term with "Grundierung", that would be the wrong answer … Rather, "Grounding" means grounding a AI model, teaching it (or her?) was the Source of Truth is ("ground truth"). If you want, you teach the AI who its Guardian is.

It was also challenging to teach AI that a server's location cannot be reliably determined from an IP address. The best response of our own AI to the question of whether a server's location can be determined from its IP address was:

The location of a server cannot be reliably determined by its IP address, as the connection between the IP address and the server can change at any time. However, there are methods for determining the location of a server, such as using IP geolocation or comparing metadata.

Response of Dr. GDPR AI to the question "Can a server's location be determined by its IP address?

I claim that this answer can be given by 99.99% of people not so well. The progress achieved through AI models is therefore slowly becoming apparent.

Conclusion

Complex AI systems can be built on Consumer-Grade Hardware*. For some problem areas, even a mediocre rented server (or worse own server) is sufficient. This category includes for example searching documents or Jira-Tickets** in the intranet based on natural language questions. The annoying search for keywords, where the highest of feelings so far was automatic synonym search, can be assigned to the past.

Even AI systems that can generate answers in their own words can be run on affordable servers. This is even true if you want to use the (unfortunately) globally insignificant German language. It's also possible to combine knowledge from multiple documents and formulate a central answer with such systems. All of this becomes practical through the application of modern optimization techniques. Feel free to contact me if you'd like to know whether your company's case is economically solvable. Economically means it won't be a rocket project, but rather a manageable cost framework that will excite you.