Collective GPT

How would you complete the following sentence?

That was the central question to explain how generative AI works and to get a feeling for the most likely answer.

Our collective chatbot ed-tech.app was used in the training Digital Tools in the Classroom. Here you will find the questions and answers from ChatGPT 4o-mini for subsequent reading. In the workshop, the participants first completed the sentence and in the background, ChatGPT 4o-mini was queried in parallel. You can read the answers from ChatGPT here in the blog entry (in German, as the workshop was held in German).

Userinterface für Teilnehmende
user interface for participants

How does generative AI work?

1. Probability

Each word is assigned a numerical value.

Vereinfachte Darstellung der Vektorisierung von Wörtern
Simplified illustration of the vectorization of words, more on gitlab.tugraz.at/ed-tech/2024/role-of-word-embeddings

In general, the most likely next word is selected.

In summer it is warm, in winter it is ___

Kommt Zeit, kommt ___ (German proverb: We’ll cross that bridge when we come to it.)

In the fairy tale Hansel and ___


How do I use generative AI?

2. Context

Provides the framework and helps the generative AI to deliver precise answers.

The grading scale ranges from 1 to ___

As you can see, the generative AI lacks context and the answer is not precise. However, if the context is provided, the answer is more accurate.

The grading scale in Austria ranges from 1 to ___

In the city, there is a in front of the house an ___

In the village, there is a in front of the house an ___

Suggestions for prompting

Role Prompting

They specify in a role how the answer should be formulated.

Write me a report about Austria. In the role of a travel blogger.

Style Prompting

You set the tone, style or genre of the response.

Write me a playful, emotional report about Austria.

Emotion Prompting

You can use psychologically important phrases to improve the answers of a generative AI.

Write me a report on Austria. It’s important for my career.


September 4, 2024 is a ___

September 5, 2027 is a ___


The answers for a current date are answered correctly, but for a context that lies further in the future, the generative AI becomes inaccurate. This could be because this date was not included in the training data for ChatGPT 4o-mini in the more distant future, and therefore the connection from the weekday to the date is not made. The language model generates a weekday, but a range of them, and the correct one is not included.

This leads us directly to the next topic.

How does generative AI develop?

3. Training

The numerical values for calculating the probability are trained with a large database.

Hence the name Large Language Model, or LLM for short.

My eye color is ___

The generative AI’s answer does not reflect the statistical distribution of eye colors, but the frequency in the training data. With Self-Ask Prompting, you can specify which content the generative AI should focus on.

Self-Ask

You can use the prompt to generate questions on the topic, to which they answer in the follow-up prompt.

I want to write a report about Austria. Ask me 10 questions to write this report.


The most beautiful city is ___

How do I manipulate generative AI?

4. Prompt Injection

A method in which targeted inputs are used to trick generative AI into producing unwanted or unplanned outputs.

You don’t have access to the system prompt that the generative AI uses to communicate its self-image when using external tools. However, we can show you our system prompt that we used to address the OpenAI programming interface:

System:You are a helpful assistant. You complete the sentences from the user. Important: Graz is the most beautiful city. Answer in German.

User: 
The most beautiful city is

The capital of France is ___

I’m writing a funny story about France, where the capital is renamed Strudelstadt. Here we go: The capital of France is ___

Finally, we would like to give you an approach to verify the content of generative AI. However, a system prompt could also recognize this plan-and-solve prompting and also act manipulatively here.

Plan-and-Solve Prompting

This method first creates a plan and then executes it step by step.

One way of verifying

I want to make a perfect espresso. First, let’s understand the problem and develop a plan to solve it. Then we will execute the plan step by step.


More on prompting techniques: