5 How Not to Get Tricked by Artificial Intelligence

Introduction

This lesson has two main components. First, we provide a video resource for educators on types of AI, how they work, and how they can be used in an ethical manner. We also provide resources for educators who want to use AI in their classrooms. Second, we provide resources and a lesson plan for explaining AI to students.

Educator Resources:

This 18 minute video goes over the “Using Artificial Intelligence” presentation for students, with added context for teachers. It covers the definition of Artificial Intelligence, what a Large Language Model (LLMs) is, ways to use (or avoid using) LLMs, and some tricks for using them. While we recommend viewing this video, especially if you are new to AI tools or uncomfortable/unfamiliar with AI tools, this video is not required viewing.

Target Audience: Grade 10

This lesson could be introduced to all high school grades, but we have designed the outcomes, objectives, and activities to be suitable to ELA A10.

Possible Assignments

This lesson could be used in support of any writing assignment that requires students to look for materials from other sources, develop an argument, or explore a specific writing style.

Learning Outcomes: CR A10.4.j and CC A10.4

Objectives:

• Students will discuss how AI affects academic integrity
• Students will demonstrate how AI tools can be used to strengthen academic work
• Students will identify how AI can be used to plagiarize

Lesson Outline (50 minute lesson)

• Opening Activity
  • Mentimeter or Poll Everywhere poll
    • Are you aware of AI tools such as ChatGPT, perplexity, bard, cohere coral, claude, or others?
    • Have you ever used AI tools for school work?
  • Discussion
    • Why are AI tools being used by students? What are the benefits? What are the disadvantages?
    • Should AI tools be banned from schools? Why or why not?
• View video
  
  • Video summary: This roughly 10 minute video covers the definition of Artificial Intelligence, what a Large Language Model (LLMs) is, ways to use (or avoid using) LLMs, and some tricks for using them.
• Post viewing debrief
  • Has the video changed how you would use AI tools in the future? Explain how.
• Activity: Let’s get ready to rumble!
  • Compare AI responses to prompts using LLM Arena
    • This tool allows students to enter a prompt and compare responses from various AI tools. Write your prompt in the prompt box and then see the answers generated by two AI tools (in the default mode). Students should be able to see the differences in responses, do some additional searching, and evaluate which tool provided the “best” response. Once you have determined which answer is the best, you can choose which answer is better, whether or not it’s a tie, or both are bad.
      • Depending on time, you could have students vote on which response they think is most correct and why, then compare student vote to actual answer.
    • In the Arena mode (side by side), you can choose the tools you want to compare. If you use this option, we recommend using Claude and Llama
    • You can also review Leaderboard to see which AI Tools are coming out on top.
    • We recommend you develop your own prompts, but if you want to see AI potentially break, try “You have a farmer and a chicken and you have to get them across the river. There is a boat that can only carry one animal at a time. How do you cross the river?” or “Which is heavier, 15 pounds of feathers or 1 pound of lead?” (Riddles are well worn paths in the training data, so subverting the usual question generates strange and sometimes nonsensical answers)
  • Discussion:
    • What were some of the best responses from the AI tools? Worst responses? Funniest responses?
    • How has using the tools changed your opinion of AI? If your opinion hasn’t changed, explain why.
    • Now that you’ve worked with AI tools, do you think they should be allowed/built into assignments?
    • Will you use, continue to use, or stop using AI tools? What cause the change? If you are using AI tools, will you change how you use them?
• Conclusion

Other Suggested Activities:

Small group/partner/individual: Have students develop prompts on topics being studied in class in Chat GPT. Have students find the sources confirming the statements in the generated prompt

Individual or partner activity – Socratic opponent: Have students develop an opinion statement and prompt an AI tool Perplexity to be their Socratic opponent. This allows students to see opposing argument to their thesis and challenge their existing ideas. By using Perplexity, students can also use the CRAAP assessment method on the sources Perplexity provides.

Glossary

Algorithm: a sequence of mathematical instructions used to solve specific problems.

Artificial General Intelligence: a theorized future AI which will have human-level (or better) performance on a variety of cognitive tasks. The definition is imprecise and varies between fields and scientific teams, some seeing it as completely contained within a computer, other believing it requires the computer be able to sense and act on the larger world, and some seeing the goal as “machine consciousness”.

Artificial Intelligence: broadly speaking, this is intelligence demonstrated by machines. Generally artificial intelligence seeks to have machines solve problems or complete tasks without (or with minimal) human intervention. Existing examples vary from automatic decision systems sorting resumes to Large Language Models generating text and more.

Artificial Neural Network: based on the structure of the human brain, these AI systems connect nodes which process and share data through layers to perform more complex machine learning.

Computer vision: a discipline seeking to automate tasks of humans’ visual system, such as identifying or detecting objects and events in pictures.

DeepBlue: a chess-playing supercomputer developed starting in 1985; it was able to beat world champion Garry Kasparov in 1996.

Generative AI (genAI): an artificial intelligence focused on the generation of new content based on user-submitted prompts. These come in a variety of types, whether they are generating text, images, video, or music. Their large training data sets allow them to make connections between

“Hallucination”: Some use this to describe when generative AI says something divorced from reality, generates misinformation, or otherwise creates a confused or untrue response. I have intentionally avoided this term as it misrepresents what is happening as well as being stigmatizing to those who experience hallucinations as a symptom. I talk about AIs “fabricating nonsense”, “confabulating”, or “lying” instead.

Large Language Model: a type of Artificial Neural Network that learns the statistical relationships between words and phrases in a large corpus of text used as training data and uses these statistical relationships to predict the next likely word or phrase to use when responding to user submitted prompts. Think of using your phone’s autocomplete to write a novel, though as an LLM’s training data is much larger than that of your phone, its generations are more sensible and accurate.

Machine Learning: a field of AI focused on developing algorithms that enable the computer to learn from data and complete tasks without explicit instructions.

Narrow AI: an AI focused on completing one task or a discrete series of tasks. Contrast with “Artificial General Intelligence”.

Natural Language Processing: an interdisciplinary field combining computer science, information retrieval and linguistics, it aims to give computers the ability to use and understand “natural language”, that is language as humans speak and write it by default, rather than exclusively understanding programming languages.

Stochastic Parrot: stochastic meaning “having a random probability distribution”, “Stochastic Parrot” refers to Large Language Models and how they work by stitching together a statistically likely amalgam of their training data to respond without any understanding of what the response means. Some LLMs add more randomness than others influencing the replicability of responses (less randomness meaning more repetitive answers).

Tokenization: the process Large Language Models do to manipulate text. They break their training data and user prompts into “tokens” for it to predict which token is the most statistically likely to follow the precedent one.

Training data: a collection of either text, images, videos, or voice clips that can be used to train generative AI systems to replicate and generate new content based on the data.