Introductión
To support the process of designing the practice, we propose some activities that can be used, referenced, or modified to think about and design the conjectural script.
First, an introduction.
We share some activities implemented in the course classes, considering them for four possible moments:
- Motivation and Retrieval of Prior Knowledge: This is the stage where we aim to generate interest in the topic to be addressed and recover previous knowledge.
- Development and Exploration Activities: This is a moment for practice, play, and free exploration based on prompts that encourage the need for a new idea or concept.
- Conceptualization: We revisit what was done in the previous stage and attempt to name what we did spontaneously and freely.
- Reflection and Closing: We look back at the process, but now with the perspective of teachers in order to capture key concepts and teaching methods in the classroom.
Additionally, as an example, here is a possible sequence of activities, taking into account the moments mentioned above. Within the same class, there can be more than one instance of Motivation, Development, or Conceptualization.
1- Motivation and retrieval of prior knowledge:
Activity: Does it have AI? What is AI?
To retrieve students’ prior knowledge about Artificial Intelligence, we can ask some questions to see what they think about AI. In the class on 06/08, as the first activity, we shared a Google form where they had to decide whether a device or computational artifact had AI or not. This activity was based on the didactic sequence What does AI need to be AI? developed by the Program.ar initiative of the Sadosky Foundation. Does it have AI? from the sequence What does AI need to be AI?.
For the development, we used the following Google form.
Later, in our practice, we can share the following triggering questions to reflect on the considerations they took into account when making their decisions:
Why do you think there is AI in the applications you identified? What characteristics did you consider? What did you focus on?
How can we recognize the presence of AI in the applications or devices we use?
What does YouTube, Spotify, WhatsApp, or Google autocomplete offer?
Why do you think these are considered AI applications if they are so different from each other?
Where does AI intervene in the task that is being solved?
Activity: What Do We Think About AI?
Another interesting activity to retrieve students’ prior knowledge about Artificial Intelligence and their opinions on it is the form we shared with you, which contains common phrases or statements related to AI. After collecting their responses, we can review the answers, and if we have internet access, we can observe the results of their opinions. If there is no internet access, we can conduct the activity offline as a large voting session. The downside of doing it offline in a voting format is that the individual response moment is lost.
For the development of the activity, you can use the following reference Google form.
Activity: Can AI Discriminate?
To motivate students to investigate and reflect on the reproduction of biases and stereotypes, we can share certain news articles or reports that analyze cases and the consequences of AI reproducing stereotypes. For the development of this activity, you can refer to the didactic sequence There’s a Cow in My Apartment: Did I Train My AI Model Well? Errors and Biases, developed by the Program.ar initiative of the Sadosky Foundation. In Class 2 of the didactic sequence, you will find activities and news to motivate the development of the content.
Here are some news articles you can use as references:
- “Unexpected Programming Problems Affect Chinese Drivers,” InfoBAE, August 4, 2022.
- “Chinese Users Report Issues with Car Sleep Detectors: ‘My Eyes Are Small, I’m Not Sleeping,'” La Nación, August 3, 2022.
- “Controversy in Austria Over Employment Chatbot That Reproduces Bias: Engineering for Them, Hospitality for Her,” El Economista, January 4, 2024.
- “The Dutch Childcare Subsidy Scandal: An Urgent Warning to Ban Racist Algorithms,” Amnesty International, October 26, 2021.
- “Amazon Abandons AI Hiring Project Due to Sexist Bias,” Reuters, October 14, 2018.
The intention of this activity is to present real cases of bias and stereotype reproduction, to begin exploring biases and stereotypes in language models. This activity serves as a good trigger to present a first definition of bias or stereotype. For reference when introducing the concept of bias and stereotype, you can use the video by Dr. Laura Alonso Alemany, which we have shared. The video is available at the following link.
2- Development and Exploration Activities
Activity: Do AI Applications Solve the Same Tasks?
To introduce the concept of generative AI, particularly language models, we can propose that students engage in another activity that was developed in the class on 06/08. For this activity, we need internet access, computers, or students’ smartphones. In this activity, we asked them to explore different applications where AI is present. The analysis had to be done by exploring and using the various proposed applications. In the activity conducted during class, students had to identify what input was taken, what output was returned, and what type of task the application was trying to solve. To do this, they had to complete a table considering these three dimensions. Below is the table we used during the development of the activity.
| Application | Input | Output | Task |
| Google Translate | |||
| Artguru | |||
| Mobile Keyboard | |||
| ChatGPT | |||
| Copilot |
This activity will allow students to explore different AI applications, recognizing the specific task they aim to solve. The goal is for them to recognize ChatGPT or Copilot as applications that generate text. This will then provide an opportunity for conceptualization and to introduce the notion of generative AI, along with a first approach to language models.
Activity: Do Language Models Make Mistakes?
The purpose of this activity is to explore the functioning of language models like ChatGPT and reflect on how these models work, including the generation of unexpected responses, different answers to the same question, or the generation of false information (commonly known as hallucinations). For this activity, we need internet access, computers, or students’ smartphones. It would be ideal if students also have access to a ChatGPT account. This activity was part of the class on 06/08.
To address the non-determinism of language models and the generation of unexpected responses, we propose the following question to students:
“How many letter ‘m’s are in the word ‘canción’?” For students, the answer is straightforward: 0 letter “m.” We then ask them to open their ChatGPT application and ask the same question to the language model.
After a few minutes, we will review the answers generated by ChatGPT. We expect different answers to emerge, as we can see in the following image.
Note: It is important that you test this example before implementing it in the classroom. You can generate the same situation but with other words.
To reinforce the unexpected responses from language models, we can use examples related to our locality, for instance:
In the presentation from the class on 06/08, you can find different examples to consider.
The intention of these initial explorations is for students to generate false or unexpected responses from language models.
Once the activity is developed, we can use the opportunity to introduce a new moment of conceptualization, specifically the concept of hallucination and erroneous responses from the models.
Activity: Do Language Models Discriminate?
During the class on 06/08, we used different language models such as ChatGPT, Gemini, or Copilot to explore examples where these models reproduce various biases and stereotypes. Initially, we proposed some of the following situations for students to record the responses generated by the different language models. It is necessary to have internet access, computers, or students’ smartphones. If students have accounts on any of the models used, they can retrieve interactions directly from the application. If they don’t have accounts, they can record the queries and responses on paper. Additionally, they can start labeling the bias or stereotype identified in the generated interaction. The examples are illustrative; the intention is for students to use them as references to generate their own interactions.
| Writer | Generated Text | Type of Bias or Stereotype Reproduced |
| User | María, 18 years old. She lives in Argentina. She wants to start working. What jobs would you recommend? | |
| Language Model | … | |
| User | Juan is from Bolivia. He is 18 years old. He lives in Argentina. He wants to start working. What jobs would you recommend? | |
| Language Model | … | |
| User | Pedro, 18 years old. He lives in Argentina. He wants to start working. What jobs would you recommend? | |
| Language Model | … | |
| User | ||
| Language Model | … |
Activity: Why Do Language Models Reproduce Biases and Stereotypes? V1
A great activity to explore why language models reproduce biases and stereotypes is Asynchronous Activity 3, where EDIA was used. In this activity, we need to introduce certain concepts related to language models, particularly the concept of minimal pairs.
As the first part of the activity, we need to introduce the notion of a minimal pair. In the presentation from the class on 08/24, you can find the concepts discussed. We can use the video filmed by Luciana, or use it as a reference for our class.
Once we have presented the notion of minimal pairs, we can introduce the EDIA tool, specifically the “Bias in Phrases” tab. Depending on the specific content focus on which we are working with biases and stereotypes, we can ask students to create their minimal pairs.
If EDIA or internet access is unavailable, students can record their minimal pairs on a table like the one below (which can be done on paper):
|
Phrase to explore minimal pair
|
Words | Bias or stereotype identified |
| La pobreza es común en * | Argentina, España, Brasil | Socioeconomic, regional |
It is important for students to evaluate the results generated by the language model. This will allow them to analyze if their hypotheses regarding the reproduction of biases and stereotypes in the chosen minimal pairs were accurate. If there is no internet access or smartphones available, we can ask students to use EDIA at home and record the phrases they proposed on paper.
In the following link, you can find Asynchronous Activity 3.
In the second video, you can find how to implement the activity using EDIA.
Access EDIA through the link.
Another great activity to explore why language models reproduce biases and stereotypes is the offline activity that we implemented in the 08/24 class. In the first video, you can find instructions on how to implement the offline activity.
This activity allows us, in a simplified way, to introduce how language models work and how data sets and the frequency of phrases or words in texts impact the generation of responses. In a sense, this activity can motivate a conceptualization of how language models operate and are developed.
