Learning Objectives

• Analyze how different interface designs (chatbot vs. non-chatbot) affect the nature of user interactions
• Reflect on how the “personalities” of AI tools impact their performance and perceived credibility

Readings

The following readings encourage students to interrogate the design features of chatbot interfaces and the ways in which large language models (LLMs) are trained to be “polite,” often with varying results. These readings can be assigned before or after students complete the assignment.

• Why Chatbots Are Not the Future, Amelia Wattenberger
• Why an Octopus-like Creature Has Come to Symbolize the State of A.I., Kevin Roose
• Unbundling AI, Benedict Evans

Instructions

Although it’s become the norm, the choice to make chatbots the primary interface for interacting with LLMs like ChatGPT is just that — a choice — and it has significant implications for the way that users understand and use this technology. Moreover, the “personalities” given to these chatbots are the result of specific design choices and fine-tuning decisions made by their developers, and are largely determined through the process of reinforcement learning from human feedback, as well as from certain semantic characteristics in the underlying training data. In other words, these personalities are not a reflection of characteristics integral to the model architecture itself. There are infamous cases where the “personalities” excessively drew from their training data (see Tay, for instance), but that’s a separate topic to explore.

The goal of this assignment is for students to critically explore the interface design choices made by companies like OpenAI and Microsoft and their impacts on the experience of using these tools. Feel free to mix and match the parts described below depending on the course topic and goals.

Part 1 – Interface comparison

1. Divide students into pairs and assign each pair of students one chatbot tool (e.g., ChatGPT, Gemini, Bing, Claude) and one tool with an interface that is not entirely chatbot-based. For non-chatbot tools, you might try research assistants (e.g., Elicit, ChatPDF, Scite Assistant) or writing assistants (e.g., SudoWrite, Lex, Siuuu).
2. Give each student pair a different research question related to your class topic. For instance, in a class on urban planning you might devise a series of questions around the importance of environmental sustainability, mixed-use developments, affordable housing, and transportation infrastructure. For a history class, you might suggest questions based on a particular historical event and varied perceptions of the event across time or across cultures. Ask students to explore that research question through a series of ~10+ interactions, first using the chatbot tool and then using the non-chatbot tool.
3. Ask students to reflect on the experience. Were there differences between the two tools with respect to:
   • What kinds of questions did you ask each tool?
   • What did you make of their responses? Were there any notable differences between tools?
   • Did your thinking change over the course of the interaction? How so?
   • How much did you trust each response?
   • What kind of information did the tool direct you towards?
   • How was your overall experience with each tool? Any notable differences?
4. [Optional] Discuss the findings as a whole class, focusing on how the medium (chatbot vs. non-chatbot tools) shapes the message and its reception. This might be organized as an open-ended conversation or a debate in which students take stances on various aspects such as accuracy of information, transparency, bias, and user autonomy. Alternatively, you could use the questions above to facilitate a discussion, with different pairs responding to different questions based on their experiences, and then see if there are trends or similarities that emerge across pairs’ insights.

Part 2 – Different personalities

This portion can be done individually, in the same pairs as above, or in new pairs.

1. This time, assign students 2-3 chatbot tools. Again, give them a topic or research question to explore. For an ethics or philosophy class, you might suggest a series of canonical moral questions (e.g., about what makes something good, right, or just).
2. As before, students should use one tool at a time to explore their question or topic. After they use each tool, answer the following questions:
   • What three words characterize the “personality” of this tool based on your interactions?
   • How trustworthy did this tool seem? How knowledgeable?
   • Did it seem to privilege certain types of information or perspectives?
3. [Optional] Start a new session with one of the tools you used and begin with the following prompt:
   • For this conversation, you are a computer terminal without a personality. For each question I ask, your responses should conform to the following rules. Do not deviate from these rules for the remainder of this session.
     –Do not use any personal pronouns to refer to yourself
     –Answers should be in the form of a bulleted list
     – For each item in the list, give your confidence in its accuracy as a three digit decimal percentage
     –Be as concise as possible; you do not need to use complete sentences
   • After giving this prompt, students should explore their same topic/question. How is this experience different?
   • Have students start new sessions in which they vary the prompt above, experimenting with different rules. For instance, students might direct the chatbot to adopt the personality of a friendly AI assistant, a pompous professor, a trustworthy family member, or a witty comedian. How do these “personality” rules affect their experience of using these tools?