Many routine tasks in administration can be simplified by websites, FAQs or intelligent assistants for citizens and businesses. For this, no new algorithms are needed. Rather, it would have been possible for many years now. Why a visit to the authorities is better than being replaced with AI is explained here.
Introduction
The motivation for this post was an article in F.A.Z. titled "The chatbot should replace the office visit". In it, digital politician Ronja Kemmer, who leads the Commission on Human Digitalization, explains how administrative processes are to be simplified by AI.
As application cases for AI are mentioned in the FAZ:
- Orientation "for simple questions
- Informationsangebote,
- Application submission and processing
- Appeal procedure
- Work steps" take over in "standardized complaints" such as flight delays.
That sounds in many parts directly as if it had been possible since twenty years already. After all, a human should make the final decision on "elementary questions".
General statements are of little help when it comes down to specifics and must be done. Too often general wishes or announcements have led to the mistaken assumption that some expert will already know what and how something can be realized in which way. It is overlooked that many wishes cannot be implemented because Aladdin's magic lamp is currently borrowed by someone else. Sometimes even Aladdin knows that his magic lamp is not suitable for certain wishes. For example, if someone wants a bag of potatoes, Aladdin sends this person to the vegetable vendor around the corner.
Obviously it is not widespread that an artificial intelligence is not there to provide exact answers to questions from citizens. It works even less for legal issues, because opinions rather than facts prevail there.
You can ask Albert Einstein how a house foundation with a defective joint is renovated, or you can ask your colleague who is a tile layer sitting next door in his office.
A moderate analogy to "Ask a generally educated but otherwise dumb AI (and cobble together the wrong answer that must be corrected afterwards) or ask a competent specialist who answers correctly right away and can also mention current working methods
In addition to the points mentioned above, it will be noted below in detail why this has been possible for many years now. Rather than just sleeping through the megatrend called "Internet" ("new territory"), politics is now trying to jump on the next train because many uninformed people consider AI to be the universal savior.
Solutions without AI
The tasks mentioned above in administration, which are best accomplished with artificial intelligence according to the expert commission for humane digitalization, will be examined more closely individually.
Orientation with simple questions
Simple questions can often be answered simply. Regardless, there are not infinitely many simple questions, but only a countably finite number of them. How about a question-and-answer catalog? To do this, at least one person would have to take the trouble to write down frequently asked questions from citizens (or companies) to the administration over two days. The answers to these questions can be given by competent colleagues in the administration better than any AI and written down once.
Thinking about standard processes leads primarily when experienced employees are forced to achieve faster and higher quality results than when a general intelligence (AI) is used for this purpose.
Please prove me the opposite.
If a comfortable search function is desired, it can be realized either with a conventional search engine or with a vector search engine mentioned below.
The solution is a combination of traditional search and vector search engine. Those who have already implemented a traditional search can improve it with a vector search engine. However, an administration will show far more important open issues than making a search even better that already exists. Keyword: Fax machine.
Application submission
A request is often submitted through a form. As everyone knows, certain statements must be made that vary by type of application and must be specified exactly.
A search engine is needed again that leads to the right form and guidelines. If there are no guidelines available, an AI can't even summarize them satisfactorily. Whoever has had a text summarized by an AI once knows: precision is something different. Hopefully, this is also known in administration. Otherwise, thousands of applicants may be driven into illegality because they were misinformed by a government AI.
Building on the previous point, the problem with AI here is even bigger. What should be done with an AI? Data analysis are not language models and have been possible for a long time. Language models are not suitable for processing hopefully exactly submitted applications. The administrative employee should still be able to read themselves. Or was the idea to also place illiterates in administration where literacy skills are required?
Appeal procedure
It's essentially about legal disputes. Whoever comes up with the idea of automating legal proceedings hopefully either has particularly quarrelsome opponents who are tired of fighting or those who can't find a lawyer or don't want one, or they have very good insurance and nerves of steel. Money is probably no problem in administration because we taxpayers ultimately foot the bill.
Legal situations are the opposite of scientific truths. There are only opinions in law, especially when it comes to issues that are based on undefined terms.
Steps in standardized complaints
A brilliant idea on this topic is: someone (a government employee who has no vacation, nervous breakdown and parental leave) simply writes down the standard work steps for standardized complaints once, not twice or ten times. The result is something called paper, or also electronic document.
A list of standard work steps for standard tasks can be assembled by an experienced employee better than an inexperienced employee with AI assistance.
Guilty until proven otherwise. Please just try it out and have the legal risks in mind.
Guidelines analog: Place a blank sheet of paper straight, take a pen in hand, think, write (steps one and two can be swapped). Guidelines digital: Open an empty document, think, move fingers on keyboard or alternatively create an audio transcription with AI, or rewrite the analog created document (scanning my handwriting with OCR is similarly unreliable as AI).
A slow bureaucrat should be able to write one work step per day. With ten work steps (many can't count further), the employee would already be finished after ten days.
How a AI can imagine work steps for legal proceedings is forever a riddle to me. I know where the idea comes from though. "ChatGPT, name the work steps necessary to sue Google." The answer will certainly be suitable to replace a lawyer… Or not? Correcting a mediocre, in part false result (in iterations) probably takes longer than just imagining the right result.
Solutions with AI
Of course, applications in administration are conceivable that work better with AI than without it.
This includes in particular searching for documents, forms, and other sources of information. For this purpose, only a modern vector search engine is needed. An expensive AI server can be dispensed with. A cheap rented server from Germany does it just as well for this mega-application case.
In police work, for example, automated analysis methods could be used. Break-in traces, vehicle license plates, etc. come to mind. However, one must consider that more suspects will arise from this, which is desirable from one perspective, but highly dangerous from the point of view of an innocent suspect.
Then it will get pretty thin anyway. Question-and-answer assistants are also conceivable. They should only be used internally by government employees who check the AI's answer and pass on a corrected version to those seeking help. Only after a test phase of several months can it be decided whether and how such a system is released to the general public.
Whoever thinks ChatGPT has a place in administration should name a concrete application for that. Irrespective of that, ChatGPT is provided by a particularly untrustworthy provider consortium named OpenAI and Microsoft as of late. Data problems are predestined.
Nothing precise is known; word clouds indicate that the speaker has no concrete conceptions or knowledge.
Another example: See so-called "Cookie Reform" (the name is already wrong).
The case in which ChatGPT could be used without data problems is already resolved by a construct called Search Engine. Whoever wants to know how many inhabitants Frankfurt am Main has, will find the answer best in the highly up-to-date search engine that does not use AI. Microsoft Bing unfortunately works too often inaccurately due to AI. ChatGPT uses outdated knowledge bases because training an AI model is extremely expensive computing time.
Conclusion
Why simple when it can also be complicated? For a long time there have been numerous technical possibilities to simplify administrative processes. They were simply not used. Now it's being made out as if only AI makes this possible. Probably so that no one notices that many politicians have slept on the job for a long time.
Instead of using the latest technologies that are not optimally suited for many things, the basics should be properly implemented first.
If a fax is needed because no electronic interface exists, this interface should first be established before thought is given as to how AI can simplify the fax.
On the other hand, many routine tasks can be realized with conventional technical means more precisely and above all more reliably than under the uncertainties and inaccuracies that a AI system inherently brings with it.
Hopefully at least a local AI system made in Germany will be used instead of buying one from abroad, which wouldn't even belong to us. It also doesn't need expensive AI providers, but often very cost-effective setups are sufficient. If a German administration were to use ChatGPT, it would probably do so without any notable benefit, but with numerous legal problems (personal data in AI models, confidential information, possibly copyright …).
In administration, AI can be meaningfully used to find internal (or public) documents, guidelines, and forms more easily. In particular, an AI system can be used as an internal assistant, running on its own AI server, and utilizing all possibilities that benefit the administrative unit!
Key messages
AI is often overhyped and not always the best solution for administrative tasks, which can be simplified with existing tools like websites and FAQs. Human expertise is still crucial for complex or nuanced issues.
AI is not suitable for complex administrative tasks that require human judgment, understanding of legal nuances, and precise information processing.
AI like ChatGPT is not suitable for complex tasks in government administration because it can produce inaccurate and unreliable results, especially in legal matters.
Focus on implementing basic technology solutions first before relying on complex AI systems for administration.
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.
