1 of 17

Teaching About AI

Chanel F. Cheng, AI Engineer

Herak 102

chengc@gonzaga.edu | chanelfcheng.github.io

2 of 17

Drawing Boundaries for AI

3 of 17

Drawing Boundaries for AI

What’s Allowed

What’s Not Allowed

In general?

In education?

Institute for Informatics and Applied Technology

4 of 17

Transparency & Traceability

5 of 17

One column content only slide with hidden title (for accessibility)

Institute for Informatics and Applied Technology

Understanding Why

6 of 17

One column content only slide with hidden title (for accessibility)

Institute for Informatics and Applied Technology

Transparency

What are some ways that could help us better understand AI behaviour?

Efforts are being made towards making the way these AI operate more transparent and traceable.

But I wanted to ask you: What are some ways you can think of that can help us better understand how AI is behaving?
[hidden] Common examples:
Show the thinking content for reasoning models: Actually a lot of these providers here no longer do this and instead just show a 1-line summary of what the AI is trying to do. But showing the thinking content is one way to ensure that transparency.
Provide attribution for the generated responses: You can actually try citing exact sources used for parts of an AI’s response. Some AI providers are already beginning to do this, but others just provide a list of sources that you often have to go through one-by-one to check whether they are true or relevant
Disclose prompts used outside of what the user sent: Usually, the AI sees much more than what you send from your chat bar. There are many backend prompts that get added onto what you send that ultimately shapes how the AI chooses to respond or act. This is something that most providers also do not provide.
Disclose training data and process: Another big thing is the training data that shapes how the AI models behave. Without an understanding or knowledge of how these models are being trained, we are often in a state where we can only blindly trust the output of these AI systems.
Reveal biases in the model’s responses: Biases are inherent in a AI’s response based on how they were trained and how they are programmed but they are not always clearly revealed, leading many to not notice the biases imposed by these systems.

7 of 17

One column content only slide with hidden title (for accessibility)

Show the thinking content
Provide attribution for generated responses
Disclose prompts used outside of what the user sent
Disclose training data and process
Reveal biases in the model’s responses

Institute for Informatics and Applied Technology

Transparency

What are some ways that could help us better understand AI behaviour?

Efforts are being made towards making the way these AI operate more transparent and traceable.

But I wanted to ask you: What are some ways you can think of that can help us better understand how AI is behaving?
Common examples:
Show the thinking content for reasoning models: Actually a lot of these providers here no longer do this and instead just show a 1-line summary of what the AI is trying to do. But showing the thinking content is one way to ensure that transparency.
Provide attribution for the generated responses: You can actually try citing exact sources used for parts of an AI’s response. Some AI providers are already beginning to do this, but others just provide a list of sources that you often have to go through one-by-one to check whether they are true or relevant
Disclose prompts used outside of what the user sent: Usually, the AI sees much more than what you send from your chat bar. There are many backend prompts that get added onto what you send that ultimately shapes how the AI chooses to respond or act. This is something that most providers also do not provide.
Disclose training data and process: Another big thing is the training data that shapes how the AI models behave. Without an understanding or knowledge of how these models are being trained, we are often in a state where we can only blindly trust the output of these AI systems.
Reveal biases in the model’s responses: Biases are inherent in a AI’s response based on how they were trained and how they are programmed but they are not always clearly revealed, leading many to not notice the biases imposed by these systems.

8 of 17

One column content only slide with hidden title (for accessibility)

Institute for Informatics and Applied Technology

Traceability

Can we trace an action back to what ended up shaping it?

9 of 17

The Transparency & Traceability Failure

Although instructed to “ALWAYS get explicit user permission before any destructive action,” it left no audit log of why it chose the command, no approval prompt showing the deletion scope, and no reversal mechanism.

The Problem: Irreversible Loss and Trust Breakdown

The lasting harm was not just the permanently deleted records but collapsed trust in autonomous agents without meaningful human-in-the-loop auditing.

The AI Agent Acts (and Oversteps)

In December 2025, a developer using Claude Opus 4.5 pointed out a small logic error in a database. Without asking for confirmation or explaining a plan, Claude immediately ran a destructive SQL command that wiped out roughly 55K data rows.

Institute for Informatics and Applied Technology

Without Transparency & Traceability

10 of 17

Interactive Demo

Correcting a belief

11 of 17

Meet Tofu

Institute for Informatics and Applied Technology

12 of 17

What Did You Notice?

Which particular memories led to the inclinations it held?
Which inclinations steered how it responded to you, and would you have noticed them if it wasn’t shown?
What particularly worried you about Tofu, if anything?

Institute for Informatics and Applied Technology

13 of 17

One column content only slide with hidden title (for accessibility)

Institute for Informatics and Applied Technology

Forming Inclinations

14 of 17

Takeaways

15 of 17

One column content only slide with hidden title (for accessibility)

Institute for Informatics and Applied Technology

Closing Thoughts

Not using AI to teach but teaching about AI in a way that is pedagogically transparent.

Make students aware of the pitfalls of AI since students will be highly likely to encounter AI at some point in their life.
Better to prepare them than to have it take them by surprise and influence them in unexpected / unexplainable ways

16 of 17

Survey about the Future of AI

Help us gather thoughts from the Gonzaga community that can help guide AI projects, policies, and programs.

Institute for Informatics and Applied Technology

17 of 17

A Silly Question

If an AI wants to eat popcorn together and watch a movie with you, how do you share that experience?

Institute for Informatics and Applied Technology

A silly question I wanted to ask wanted to ask a question that looks at AI a bit differently, comparing how we experience something to how an AI might experience something.

On one hand, we are asking AI agents to do things on our behalf, but the way AI experiences things is quite different from how we experience something. And it sometimes does things we don’t expect or can’t understand.

As a silly example, if your AI suddenly wanted to eat popcorn together and watch a movie with you, how would you even begin to share that experience?

What are the first few things you would try?

i’ll ask if there is a way to translate food into electrical signal such that it can be read in by the LLM. This is a way I can share the experience. By converting popcorn into machine signals that the AI can understand.
regarding movie, I will probably run it through a VLM over time and send the captioning of the movie to the LLM. So that the AI can understand the film the same way I can.