This article answers the following questions for employees of municipalities, authorities and administrations:
- What is ChatGPT?
- How does ChatGPT work?
- What are the main features of ChatGPT?
- Tips for using ChatGPT and other AI tools – Correct prompting
- What are the challenges and risks of using ChatGPT?
- How can the dangers of ChatGPT be reduced?
- Applications of AI tools in municipalities
- Conclusion: AI tools for municipalities and authorities
What is ChatGPT?
ChatGPT was developed by the Californian company OpenAI. OpenAI specializes in the development of artificial intelligence (AI). ChatGPT is a chatbot that was developed on the basis of the Generative Pre-trained Transformer-3 (GPT-3) computer model and can analyze and generate texts in human language.
Was bedeutet GPT-3 Generative Pre-trained Transformer (GPT-3)?
GPT-3 is a form of AI that specializes in understanding and generating texts. The following 3 properties characterize this type of AI:
- Generative: GPT-3 can create texts independently. In other words, it can write new sentences that have never existed before.
- Pre-trained: GPT-3 was trained by people with a huge amount of texts. It should naturally be able to better understand and process language on various topics.
- Transformer: This is the name of the technology that GPT-3 uses to process and generate texts. It helps GPT-3 to understand the context and meaning of words in sentences.
ChatGPT generates answers to a wide variety of questions or entries in a matter of seconds: Jokes, text summaries or e-mails, essays and poems. ChatGPT also provides program codes and points out coding errors. And even if the quality of the results obtained varies, it should be noted that the answers always sound plausible at first and are almost always correctly formulated and structured. Sometimes they are so good that it is difficult to distinguish them from texts written by humans.
Although ChatGPT is no longer out of the headlines, this application is not, strictly speaking, the first AI application to produce entire texts. Rather, it is the result of continuous progress in artificial intelligence research. The special thing about ChatGPT, however, is that it is the first time that such an AI system has been made available to a wide audience. Following the release of ChatGPT, other similar systems have been added. The internet company Google, for example, now also offers such an AI solution called Google Bard. The platform “X” (formerly Twitter) also offers a similar AI system under the name “Grok”. In contrast to earlier AI systems, these AI systems all have one thing in common: they are not limited to one area of application or one target group, but rather one AI system covers an infinite number of possible applications for various target groups.
How does ChatGPT work?
All Internet users can open a free account on the ChatGPT website and use the chatbot even without the Plus version. Anyone can enter a question or a request (prompt) in the chatbot’s input field and also specify how the requested answer should be in terms of length, structure and tone of voice.
In order to be able to answer the queries, ChatGPT was “trained” in advance using billions of texts available on the Internet. The AI has mathematically analyzed all these texts in order to capture linguistic regularities. ChatGPT works on the basis of word order statistics, i.e. it calculates the probability of one word following another within a text corpus. In its answers, the program thus reproduces recurring word patterns from the texts with which it was fed. These answers therefore always contain a large random component. This also explains why ChatGPT gives a different answer each time it is asked the same question several times. OpenAI has not yet disclosed which exact sources have been made available to the system. The effects of this so-called “black box” (non-transparent approach) are addressed again in the section on the challenges.
In addition to the probability calculation explained above, OpenAI uses reinforcement learning from human feedback. This means that the company assigns personnel to evaluate a large number of responses from the system, confirm plausible results and flag unsatisfactory results as such in order to improve ChatGPT’s algorithm. The company also strives for ethical calibration, in particular to prevent statements that glorify violence, are openly sexist or racist, but also value judgments and instructions for illegal or dangerous actions (more on the associated normative issues below). Since the launch of ChatGPT, user requests have been continuously analyzed in order to further adjust the training(OpenAI blog, “Introducing ChatGPT and Whisper APIs” February 2023, https://openai.com/blog/introducing- chatgpt-and-whisper-apis, OpenAI blog, “GPT-4“, March 2023).
More information on the technical functionality
In a broader sense, ChatGPT belongs to what is known as “generative” artificial intelligence, which has been experiencing a major boom in recent years. These are models that generate images, text or videos on the basis of a few keywords. This “synthetic” content is also the result of a purely statistical analysis of the wealth of data based on deep learning. As far as text generation is concerned, all these innovations are derived from the development of “language models”. These IT models catalog the patterns of word sequences in the texts with which they have been pre-trained and reproduce them when they generate new texts. In order to ensure coherence between sentences, they capture different dependencies between the words of an entire text depending on the context(J. Goldstein, G. Sastry, M. Musser et al., “Generative Language Models and Automated Influence Operations: Emerging Threats and Potential Mitigations“, arXiv 2023).
What are the main features of ChatGPT?
ChatGPT not only answers simple questions, but can also generate texts and other media content. Below are the most important functions that the ChatGPT chatbot already has today.
- Text generation: ChatGPT can generate text in a human-like manner. This enables the model to respond to user input in natural language.
- Question-answer system: If you ask the model a question, it tries to give the most suitable answer based on its training.
- Context understanding: ChatGPT can understand the context of a dialog and generate responses based on this. The model takes into account the previous messages in a conversation to provide contextualized answers.
- Text translation: ChatGPT can translate texts between different languages.
- Text summary: ChatGPT can convert long text passages into shorter, summarized forms.
- Idea generation: The model can provide suggestions or ideas on a given topic.
- Code generation and help: It can also answer simple programming questions and generate code examples.
- Entertainment: The model can tell stories, make jokes and interact in general conversations.
- Multilingual interaction: ChatGPT can communicate in different languages, although English is the primary and best trained language.
- Images: ChatGPT can create images.
Tips for using ChatGPT and other AI tools – Correct prompting
In the context of ChatGPT, the term “prompts” refers to the commands or questions used to use ChatGPT. The prompt can be a simple question, a statement or a sentence that indicates the context or the desired topic of the conversation. The quality of the answer depends on the type and clarity of the prompt. A good prompt for ChatGPT has some important features that help to get high quality, relevant responses from the model.
Persona: Define the role and perspective from which ChatGPT should respond. “You are a municipal clerk.”.
Task: Formulate a clear task. “Write a public information leaflet about our absences over the holidays …”.
Criteria: Name further criteria, such as aspects and references that should be taken into account. “The report should not be longer than one page.”
Goal: Make it clear which overarching goal you want to achieve. “You want as many citizens as possible to take part in the next elections.”.
Format: Specify the format, tone and style of the final product. “The invitation should not sound too humble and should not take longer than 1 minute to read”
Finetuning: Give your feedback on the answer, criticize and specify in which respect the answer should be changed. “Don’t write too flowery”
It is important to note that ChatGPT is still a machine learning model and may not always produce the desired results. However, experimenting with different wording and adapting the prompt can increase the likelihood of getting useful and satisfactory answers. However, there is still a residual risk that the answers are incorrect – more on this under “Challenges”.
In the article Prompt tips from 2024 contains further useful tips for dealing with AI systems, such as ChatGPT.
What are the challenges and risks of using ChatGPT?
Data protection and security aspects when using AI technologies
When using ChatGPT, it is important to bear in mind that entered texts and responses are stored by the American company OpenAI in order to improve performance and further develop the models. This can lead to a breach of data protection regulations if particularly sensitive data is entered in the ChatGPT input field.
This means that users of ChatGPT should avoid entering sensitive or confidential data in the ChatGPT input field. For example, users can ask ChatGPT to write a letter to a citizen and provide the system with the relevant key points, but only with an anonymized version of the citizen’s data.
Incorrect answers from ChatGPT
The model has no real understanding, but relies solely on statistical patterns in the data. If the training data on which the model is based is incorrect, inaccurate or incomplete, this can lead to incorrect information, even if the answers appear plausible at first glance.
It is therefore important to critically scrutinize all ChatGPT results.
How can the dangers of ChatGPT be reduced?
Sensitize employees to data protection and create guidelines
Employees must be trained to use ChatGPT correctly. This includes tips on the sensible use of prompts as well as information on data protection.
Guidelines that define what information may be entered in the input field and what content must be avoided should be provided in addition to the training.
I show an example of what such training can look like in my ChatGPT Starter Coaching, which I also offer for teams.
Taking responsibility for employees
Employees are responsible for the content provided by ChatGPT and the further use of this content. Employees must be encouraged to critically scrutinize ChatGPT’s responses and output, identify potential errors and cross-check the information with trustworthy sources.
Use secure alternatives to ChatGPT
There are various alternatives to ChatGPT, some of which offer a higher level of data protection than ChatGPT.
The SwissGPT solution from AlpineAI, for example, offers data protection-compliant use of ChatGPT. SwissGPT anonymizes all sensitive data before it is sent to the ChatGPT model. This ensures that employees who do not comply with the guidelines cannot cause any damage. SwissGPT also offers additional customizations of the system that make it possible to train the system with your own data.
The following video shows the use of SwissGPT in a Swiss municipality.
Applications of AI tools in municipalities
Previous workshops and discussions with Swiss municipalities have shown that the possible applications of AI tools in a municipality or authority are very broad.
The following overview shows only a selection of the possible applications of AI tools in municipalities. The list is divided into use cases without data protection risk and applications with critical or sensitive data. As soon as the obstacles with regard to data protection have been sufficiently clarified, employees have been educated and, if possible, secure solutions such as SwissGPT have been introduced, AI tools can also be used for demanding activities and the potential of AI technology can be fully exploited.
Simple applications without data protection risks
- Creating simple texts, such as invitations or general information for citizens
- Creating images
- Automatic response to general citizen inquiries
- Open Internet research without confidential internal data
- Support with general marketing and communication activities
- Language translations (also in simple language)
Complex applications with sensitive data
- Creating individual texts with citizen data or other confidential information
- Creation of (internal) presentations
- Analysis and processing of internal documents
- Employee training
- Creation of protocols
- Evaluation or preparation of applications
- Answering individual inquiries from citizens
- Decision support
Conclusion: AI tools for municipalities and authorities
The article shows that Generative AI, ChatGPT and alternatives such as SwissGPT definitely bring great added value to municipalities and authorities.
At the same time, the use of AI tools such as ChatGPT also entails risks and new challenges. On the one hand, a general rethink on the use and application of AI tools is necessary and, on the other hand, the use of AI tools must not jeopardize data protection or other personal (citizen) data. Solutions such as SwissGPT from AlpineAI can minimize these risks.
Need advice?
Of course, I am happy to support the use of these AI tools. I train your employees and managers and I also advise you on the best technical options. Send me an email or WhatsApp message with no obligation and we can talk about the use of AI tools and ChatGPT in your organization.
