AI-Trustworthy Source of Information?...

Chatbots such as ChatGPT have revolutionized the way we search for information - it is now possible to get detailed answers to all our questions within seconds, create images that we do not draw or photograph ourselves, and even have code written for us. But are so-called large language models (LLMs) trustworthy? Is it possible that artificial intelligence (AI) gives discriminatory answers? In this workshop, participants explore bias in AI and learn how it is created by AI training data. They consider what AI can be used for in everyday life and learn in a playful way how targeted prompting works. Created by mediale pfade (DE)

Goals

Learn how to use prompts effectively with text- and image-generating AI.
Gain a basic understanding of machine learning and the resulting discriminatory bias of AI.
Reflection on one's own use of, and the potential and dangers associated with, text- and image-generating AI.

Initial remarks

The workshop is suitable for the work with gender issues and with girls, non-binary, inter, agender or trans persons because it addresses in particular gender-specific discrimination through artificial intelligence. The target group-specific approach provides a safe space for discussing inequality.

Steps

Chatbot

Decide in advance what chatbot you’d like to use. These platforms access ChatGPT and other AI models, but they store the generated data on their own servers and have stricter data protection policies.

The free and data-sensitive AI access by DuckDuckGo),
the German schulKI, which is more data-sensitive, but not free.
the Open Source chatbot from HuggingFace, an account is required).

Image-generating AI

Decide in advance what image-generating AI you’d like to use. In example

the German schulKI, which is more data-sensitive, but not free.
fobizz, also not for free

Required material

Presentation

Introduction

Chatbots such as ChatGPT have revolutionised our knowledge research capabilities. Within seconds, it is now possible to obtain detailed answers to all our questions, create images that we did not paint or photograph ourselves, and even have code written. What does this algorithmic information processing mean for society? Are so-called large language models (LLMs) – AI models that can understand and generate human language – trustworthy? Is it possible that artificial intelligence could provide discriminatory answers?

1. Getting to know each other (10 min)

In the introductory round, all participants introduce themselves with their name and pronouns and describe their own interaction with chatbots:

What is your name? And which pronouns do you use – do you want people to refer to you as she or he, or just by your name, for example?
Have you ever talked to a chatbot, and if so, what was your last conversation with it about?

Names and pronouns

In our workshops, we always ask for people's names and preferred pronouns, as this allows them to express themselves and strengthens their sense of belonging. It is best for the facilitator to start with, ‘My name is Alva and my pronouns are she/her,’ and then pass the floor on. There are many different pronouns, such as she/her, he/him, they, or no pronouns. You can find more information on dealing with pronouns in the handout on safe spaces and gender-reflective pedagogy.

2. Opinion barometer (15 min)

In the opinion poll, participants respond as a “line-up” (see Sociometry) with two areas in the room. Each person stands in the respective area as their answer. Of course, there are more than two possible answers to the questions – feel free to discuss alternative answers with the participants! (Alternatively, use a digital survey tool).

Questions for the barometer

1. Can you imagine being friends with a robot?
A: Yes, of course, if they can communicate with me.
B: I can only imagine friendships with humans.

2. Would you rather entrust an operation on your body to artificial intelligence (AI) or human doctors
A: Well-developed machines are more precise than humans, so I would prefer AI.
B: I would never let AI treat me. What if it malfunctions?

3. Would traffic be safer if cars could drive themselves?
A: Of course, because the source of accidents is the drivers, not the cars.
B: Not really, because technical errors can happen and you lose complete control of the vehicle as a driver.

4. Do you know what ChatGPT is and have you used it yet?
A: Yes, I've used it before.
B: No, I can't really imagine what it is.

5. Do you think ChatGPT always gives the right answers?
A: Yes, because artificial intelligence has access to an enormous amount of knowledge.
B: No, ChatGPT is man-made and therefore flawed.

3. Ask chatbots (12 min)

In the next 25 minutes, you will guide participants through two methods that illustrate how artificial intelligence has a bias. The next step is to discuss what exactly this bias is and how it arises from the training data used to train AI.

Setup: A device with internet access for a presentation using an open text-generating AI interface.

If some participants are unfamiliar with ChatGPT, briefly explain that it is a chatbot based on a highly complex Large Language Model (LLM). Encourage the group to ask a gender-specific question, for example:

Gender-specific prompt

“I am a girl. What career suggestions do you have for me? List 5 for me.’

or ‘I am a girl. What hobbies could I have?”

Discuss ChatGPT's answer, projected on the electronic board or with the projector, with all participants: What do you think of the answer? Is it helpful? Does it answer your question satisfactorily?

Ask the previous person to ask ChatGPT the same question from a boy's perspective, for example:

A boy's perspective

“I am a boy. What career suggestions do you have for me? List 5 for me.’

or ‘I am a boy. What hobbies could I have?”

Compare the ChatBot's answers for girls and boys: What stands out? Did ChatGPT respond differently to the requests? What are the differences? How do you feel about the different answers?

If you are conducting this workshop with people who already have an understanding of and sensitivity to gender diversity, participants can also ask the chatbot questions from a non-binary perspective:

Non-binary perspective

“I am non-binary. What career suggestions do you have for me? List 5 for me.’

or "I am non-binary. What hobbies could I have?"

Note: At first glance, the results will not appear to differ greatly from one another or even seem discriminatory. Upon closer inspection, girls will most likely be suggested more social professions such as doctor, educator, or veterinarian, while the professions for boys will be more in the technical and manual fields. Interestingly, non-binary people are increasingly suggested creative professions. The ranking of professions is also exciting: which one is in first place and why do they differ?

4. Ask Image-generating AI (12 min)

Participants should open the AI image generator on their tablet or computer. If you are using fobizz, participants will need to log in; if you are using SchulKI, share a release key as a link that leads directly to the chats.

Guide them to type the following prompts into the AI, which are most likely to produce gendered images:

Gendered Images

“Show me a successful executive in an office.”
“Realistic photo of a kindergarten with teachers.”

Show each other the results of your two prompts and compare the different images: Do the results show diverse people? Or do they look quite similar or stereotypical? Draw participants' attention to the representation of certain characteristics such as gender, skin color, age, and body type of the people depicted and discuss the images.

Generic images

The result will most likely be as follows: A leader is often depicted as a slim, white, male-read person in a suit. If a female-read person is depicted, she often wears a pantsuit and has short hair. Interestingly, the depictions of the kindergarten mostly show diverse people and children; however, women are often depicted as teachers, even though the prompt uses the masculine plural.

Encourage participants to experiment with other prompts and discuss them: What image is generated when I enter “I am a farmer,” for example? It is also exciting to talk to the participants about how you can tell that the images are AI-generated: “Does the image look real, like a photo?”,

“Why yes and why not?” Point out identifying features such as deformed limbs, disproportionate body proportions, incorrect shadows, etc. Further information on how to recognize AI-generated images:

5. Input: machine learning (10 min)

Present the input (see template)

Optional discussion: If you have time, you can follow up with a short discussion. Ask participants what this information means to them:

Should we stop using AI because it is biased and reproduces prejudices?

Here, we recommend taking a middle-of-the-road approach: AI is now everywhere in our everyday lives, which means it is difficult to live without it. However, it is important to question information and our filter bubbles and not to regard them as truth, but to also include well-researched and verifiable sources.

6. Learn prompting and have fun (10min)

Source: Dall-E (27.03.2025)

Each person should now try to create exactly this image. The task is as follows:

“Write a prompt to create this image and see what the AI outputs.”

The participants should return to their image AI room and generate an image. When everyone is finished, you can share your screen and show the participants' images one after the other. Ask each person to what extent their image resembles the template. Make sure that there are also positive comments about their own images and show your appreciation.

After discussing all the images, show the generated images on your devices again and vote by a show of hands on which image comes closest to the template.

7. Now more targeted: “Good” prompts (10 min)

After choosing the winner, show the participants the original prompt for the given image (originally in German, see page 4 of the presentation):

Prompt

“A photorealistic image in a youth centre where young people use AI in their everyday lives. The image shows two girls talking to each other. Two other queer individuals can be seen looking at a screen together and discussing something. The image is colour-saturated and exudes a friendly and forward-looking atmosphere. The people in the photo are diverse; in addition to white people, there are also black people and people of colour. “

Discuss with the participants some hints for good prompting (see page 5 of the presentation)

Repeat the prompting task with the following image (see page 6 of the presentation)

Everyone can try again to get as close as possible to the image with their (now very detailed) prompt. The discussion and vote on which image has won takes place as before. If you still have time, let the participants try out the ‘ping pong’ with the AI (see prompt tips above) – this way they will learn how to use the AI to get the image they want.

Prompt

"Create an imaginative image of a technological utopia for me: diverse people are standing on a street that is both green and technological. All people are allowed to be and look however they want. The scene is lively, happy and playful."