AI 101

What you need to know

What we’ll cover:

What AI is
Examples of AI in your life
How AI works (conceptually)
Why AI is gaining attention (again)
Practical concerns when using AI
Examples of how to put this information into action

Goals for the Training

Understand at a basic level what AI is
Understand that there are different kinds of AI models
Understand different AI models are appropriate for different tasks
Be able to apply this information to evaluate if an AI model is probably appropriate
Understand why the term AI has become seemingly ubiquitous
Have a basis for further learning (come back for AI 102! I’ll be a nerd and tell you why AI is cool!)

How would you define AI?

“AI is whatever hasn’t been done yet.” — Tesler’s Theorem¹

Defining AI

How do you define “Intelligence”?
Academic definitions differ and are not always helpful for practical use

AI Definition Overview²
	Human-based	Ideal Rationality
Reasoning-based	Systems that think like humans	Systems that think rationally
Behavior-based	Systems that act like humans	Systems that act rationally

Defining AI

AI is “an engineered or machine-based system that can, for a given set of objectives, generate outputs such as predictions, recommendations, or decisions influencing real or virtual environments. AI systems are designed to operate with varying levels of autonomy”³

Defining AI

Questions to help determine if it’s AI:

Does the technology use data to provide predictions, recommendations, insights, or decisions?
Does the technology augment or replace human decision-making?
Does the company use words such as “personalized”, “tailored”, and “adaptive” in its marketing?”
Would it be difficult or impossible to write out how a program does what it does, either because the process is opaque or because it is mathematically complex? ⁴

Examples of AI

What are some examples of AI you interact with on a regular basis?

Optical Character Recognition

Video Games⁵

Spam Filters⁶

How AI Works Conceptually

Understanding the different kinds can help you know when and how to use AI

Lets say you need an AI to help organize some recipes:

Logicist AI

Probabilistic AI

Neurocomputational AI

Think of logicist artificial intelligence as a series of rules. If you know all of your cake recipes are made with flour and butter and are cooked at 350 degrees, and none of your non cake are you can give that rule to an AI who can sort them based on the logic you gave it.

Think of probabilistic AI as showing a model a bunch of recipes labeled ‘cake’ and ‘not cake’ with notes on each ingredient; once it is trained you can show it a recipe and it’ll say how likely it is a recipe for cake or not

Think of “neurocomputational” as having a bunch of robots working in groups. You can show them a bunch of recipes, and each group will work to figure out one part of what makes a recipe. You can then ask them to label something as cake or not, or even potentially ask them to give you a new cake recipe. They could be super accurate (maybe they figured out some cakes are made with almond flour), or could be really off (they could think a recipe for wet cement is cake), because they don’t have the context to know what cake is, and aren’t good at telling you what they do / look at.

Real Life Examples of Each Kind of AI

Logicist

Probabilistic

Neurocomputational

Optical Character Recognition

Video Games⁷

Spam Filters⁸

Lots of video games use AI to control NPCs (non-playable characters). For example, in Skyrim NPCs have to know where to walk without someone controlling them, how do they do that? How do NPCs know how much they should charge for a book, or spend on a sword you want to sell them? Newer video games often use a combination of several kinds of AI, so perhaps in Skyrim there would be a logicist AI to determine where NPCs can walk, and then a neurocomputational AI to determine where they actually walk.

Unfortunately, spam emails don’t usually tell you they’re spam, so how can spam filters work? It would take hours and hours of manual work to create rules that would work for the spam I get on my work email and my personal email, much less that would be mostly accurate for most people. This is somewhere ML excels- select an appropriate model, give it labelled data, and it can create rules to catch most spam.

Optical Character Recognition is used to e.g. add words to scanned PDFs. It’s usually based on Neural nets, which are especially good at computer vision

AI Model Type Overview
	Logicist	Probabilistic	Neurocomputational
good for:	clearly defined logic	clear factors	No clear process
bad at:	handling uncertainty	operating with limited /bad data or assumptions	Truth; Explainability
example:	email inbox rules	(some) population forecasting	chatGPT

To give another example of each: In my personal life, I’ve signed up for a few store rewards cards. I like looking at the flyers before I go, but I don’t want them cluttering up my inbox, so I set inbox rules to move those to a folder. Some of the companies use different emails, so I can’t create rules that always filter correctly because there is some uncertainty about what emails should / should not be moved.

At my last job, I had to predict how many cars we’d sell of certain types so we could buy appropriate stock. There aren’t clear, human definable rules for predicting that, but we had previous sales data. The data sucked, so the forecast had an upper limit on how good it could be. As a company we artificially inflated demand in December by running sales then, so we had to account for that outside of the Machine Learning model because the data was not accurate.

When I was a lab manager, one of the projects I worked on was creating an algorithm to sort hand-drawn paperwork. There were no clearly defined features I could use, and no clear rules about which was which I could communicate to a computer. I cross-trained a computer vision neural network to sort the documents into different folders but it couldn’t have told me e.g. why it sorted them a certain way.

Why AI Is Gaining Attention (again)

Why do you think there is a lot of hype around AI right now? Has this happened before?

Why not before now:

The idea of ‘artificial intelligence’ can be seen even in writings by Ancient Greek philosophers, so why haven’t robots taken over yet?
Turing’s Chess
early neural nets (1950’s)
Impracticality of probability
not user-friendly
AI Winters

AI as a field of study came from formal logic theory (which came from philosophy), with some important additions from probability theory and neuroscience.

Alan Turing created an AI which could play chess by designing an algorithm which would take the current state of a chess board, and then assign points to all possible next moves and choose the move with the highest points. Because of technical limitations, he never ran this on a computer.

Early neural nets had to be written by hand on punchcards- if you messed up somewhere in the middle you had to completely start over. There’s a hard upper limit on how much you can do in a system like that based on how many people you can have work on the project.

Until recently, probability-based AI was impractical for most cases. In a lot of situations, you’d need to run a few hundred calculations for a basic model, and gathering the data and running those calculations was difficult.

Typically, if you wanted cool AI tools you’d have to find/hire a nerd to make them for you. Most of the closer to off-the-shelf options required technical expertise to correctly use.

AI history is full of booms and busts- we figure out something cool and then over promise and under deliver. So people lose interest in AI because they think all of it is overhyped instead of just some, and then things that objectively are AI stop being called AI. This probably includes things you use every day, but don’t think of as AI.

Why now is different (technical):

Graphical Processing Units (GPUs)⁹
Incredible access to data
- Raised some new needs
Bayesian probability is easy(ish) now
Attention Is All You Need¹⁰

GPUS are designed to run multiple processes at the same time, so models that would be too complex to run with CPUs because of the time became practical with GPUs

The increase in access to data is due to many factors, including the normalization of data collection e.g. by grocery stores and smart phones, but the easiest to point out is the rise of the internet. It’s a lot easier to scrape a few dozen websites and get usable data compared to trying to convert printed books into a format usable by ML. These changes have also created new (or amplified existing) use cases- like Netflix and Amazon wanting to show you things you might want to watch/buy.

Bayesian probability can handle uncertainty better, and is often more useful for ML than frequentist methods (come to AI 102 for a crash course), but requires a lot of computation which was impractical before modern computing.

A paper put out by computer scientists at Google in 2017 introduced a new way of adding context to things as they were processed by neural nets. Come to AI 102 if you want to learn more, it’s super cool, but the important part is that it created huge advances in the field

Why now is (not) different

upper limit on benefits from data
better compute still has a practical limit ¹²
putting all of the eggs in one or two baskets funding and hype-wise¹³

Practical Concerns with AI

What do you think are some practical concerns when you are using AI? What do you try to keep in mind when you are using AI?

Genuine Dangers

Malware¹⁴
Legal Grey Areas
- AI Notetakers¹⁵
- State laws
- Mostly applying existing laws to AI
Lost of trust with communities
Loss of quality

Generative AI is subject to unique ways of hacking that traditional software is not, and cybersecurity experts are just starting to get a sense of the scope of the problem. One area of particular interest is with coding- if you use generative AI to code there are several ways it can pose cybersecurity risks that are elevated compared to coding by hand. Even if you don’t open CoPilot and ask it to code for you, if you use low or no code tools, most of those use generative AI. The legal realm around AI is currently rapidly changing and uncertain. A lot of where we’re currently seeing these changes is with figuring out how current laws apply to new(er) AI tools. For example, multiple AI Notetakers are in court trying to determine things like if collecting biometric information without proper consent is breaking California privacy law, or if sending an AI notetaker to a meeting you don’t attend might be breaking federal wire tap law. You could potentially use AI in a way that leads to a loss of trust with our communities, for example if you use generative AI to create incorrect statistics and use those in a grant and don’t get the grant as a result, our communities might not want you to apply for grants any more. Due to automation bias, it’s really easy to produce worse-quality material (e.g. a grant application) that you feel is better because it came from a more automated source.

Ethics

Data Collection
- Privacy Rights¹⁶
Copyright
- Potential loss of models¹⁷
- Currently not always able to copyright AI generated content in the US¹⁸
Public Opinion
- Plagiarism¹⁹

In Data Science, a common refrain is garbage in garbage out. At an extreme, if a model was trained on 4chan, you’re not going to get professional emails out of it. For a more realistic example, if a transcript generating model was only trained on men with british accents, it won’t do a good job transcribing meetings we’re in. Data collection can also affect how the company uses your data- e.g. most AI note takers use your data to train their models, which makes them not appropriate for meetings where someone might say something you don’t want in a model you can’t control.

There are currently several court cases working to determine if training AI on copyrighted works without permission is considered fair use, and if there are instances where it might not be. Those cases could significantly impact many of the popular models today.

Right now, public opinion is still highly polarized on AI. Some people have serious ethical concerns with how a lot of the larger LLM models have been trained or are being used (e.g. to replace workers), while others consider any use of AI to be inherently unethical. Navigating this landscape will require us to be aware of opinions and considerate of them even if we disagree with them (both pro- and anti- AI).

Right model for the job

Probabilistic vs deterministic
Is AI the best solution?
Resource use
- Less obvious when it’s “on the cloud”, but still worth considering

If I create a website for online shopping, and you judge it based on whether or not it lists states with the letter “R” in their name, that’s not really fair unless my website is the “states with the letter”R” in their name shop”

If you need specific wording or existing data, probabilistic AI is not going to be appropriate. If you want to organize pdfs, probabilistic AI is more likely to work (unless you have a good file naming scheme) because it can create rules to determine which files are similar to one another.

There are times where AI is not the best solution. If you want to check the accessibility of a website, there is a lot you can do with automation (e.g. color contrast), but some of it is easier to just check than try to create AI for (full screen-reader compatibility).

Certain AI models do require more resources to use, you should balance how much you benefit from those models with the resources required to use them.

You are the one using the tool

Automation Bias²⁰
Responsibility

As a general rule, people tend to trust or rely on things that are ‘automated’ more than they should. This can include even things like trusting information you get from AI more than an expert because most LLMs are really good at sounding confident even when they’re wrong (like a lot of people on the internet), but it can also include not being as thorough when you fact check things you get from AI because it feels more trustworthy. This is part of why you have to check any information with verified sources (also because the little robots don’t understand what a ‘fact’ is). If a model produces a bad result (e.g. bad image, inaccurate transcript, inaccurate information), you are responsible for anything that happens if you use it. AI is a tool and you are using it, it is not autonomous (yet)

Quick Practical Advice

Don’t use AI search results without verifying them
If you use chat bots:
- Give them context in your prompt
- Try different prompting strategies (e.g. say “you are an X”, be nice)
Don’t forget the sunk cost fallacy
Sometimes older methods are better

The AI search is unreliable, you need to verify any information with reliable sources (like in our policy)

Currently, there is not clear evidence that certain strategies are universally helpful with all tasks for any given LLM model, and the models are changing at a fast enough pace that a strategy that works now might not a month from now. Instead, try different strategies to see which seems to get you what you want. Some strategies include being polite in how you word the prompt, including in your prompt something like “You are a planner,”, and prompting it to “think step by step”. This last one will use more resources, and is not always effective, so use it sparingly.

It can be easy to spend more time trying to do something an ‘easy’ way than you would doing it the ‘hard’ way. I have a tendency to do this, and what I’ve found helpful is setting a timer for a quarter or half as long as I think it’d take to do the hard way and switching to the hard way if I’m not done in that time.

AI is the new cool shiny toy that’s suddenly everywhere, and it can be exciting to use new tech (or old tech with a new coat of paint). It’s not always the most effective or easiest way to do things even if it feels like it is.

Putting It In Action / Hypotheticals

Let’s say you want to predict the future population of Knoxville. A company is offering to use AI to forecast the population for the next ten years. What could you ask them to evaluate if their methods work?

Would you feel confident in your answer, or would you want to ask someone who is more familiar with forecasting?

Let’s say you want to write Planning 101 guide for ETDD to give to local planners to help them make better choices (and maybe sell later), but you’re on a time crunch so you want to use a generative AI to help you.

How could it help you?

What concerns would you keep in mind using it?

Thank you

References

2024. https://foundationcapital.com/ideas/has-ai-scaling-hit-a-limit.

2025. https://news.bloomberglaw.com/litigation/wiretap-suits-pit-old-privacy-laws-against-new-ai-technology.

n.d. https://commons.wikimedia.org/wiki/File:Email-spam-sample_(cropped).png.

n.d. https://owasp.org/www-project-citizen-development-top10-security-risks/content/2022/en/CD-SEC-01-Blind-Trust.html.

In Computerworld. n.d. https://www.computerworld.com/article/4041849/enterprise-note-taking-apps-face-legal-scrutiny-as-otter-hit-with-privacy-suit.html.

In Google Cloud. n.d. https://cloud.google.com/discover/gpu-for-ai.

In NVIDIA Technical Blog. 2025. https://developer.nvidia.com/blog/how-code-execution-drives-key-risks-in-agentic-ai-systems/.

Bergmann, Dave. The ELIZA Effect: Avoiding Emotional Attachment to AI Coworkers | IBM. 2025. https://www.ibm.com/think/insights/eliza-effect-avoiding-emotional-attachment-to-ai.

Bringsjord, Selmer, and Naveen Sundar Govindarajulu. “Artificial Intelligence.” In The Stanford Encyclopedia of Philosophy, Spring 2025, edited by Edward N. Zalta and Uri Nodelman. Https://plato.stanford.edu/archives/spr2025/entries/artificial-intelligence/; Metaphysics Research Lab, Stanford University, 2025.

Chan, Cecilia Ka Yuk. “Students’ Perceptions of ‘AI-Giarism’: Investigating Changes in Understandings of Academic Misconduct.” Education and Information Technologies 30, no. 6 (2025): 8087–108. https://doi.org/10.1007/s10639-024-13151-7.

Namco. English: Pac-Man Approaching a Row of Three Dots. 2024. https://commons.wikimedia.org/wiki/File:Pac-Man_eating_dots.svg.

Office, United States Copyright. Copyright and Artificial Intelligence Part 2: Copyrightability. 2025. https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf.

Office, United States Copyright. Copyright and Artificial Intelligence Part 3: Generative AI Training. May 2025. https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf.

Patel, Dwarkesh. 2026. https://www.dwarkesh.com/p/francois-chollet.

Programmers at Work: Interviews Wit 19 Programmers Who Shaped the Computer Industry. Repr. with new additions. Microsoft Pr, 1989.

published, Craig Hale. “Microsoft Admits an Office Bug Exposed Confidential User Emails to Copilot.” In TechRadar. 2026. https://www.techradar.com/pro/security/microsoft-admits-an-office-bug-exposed-confidential-user-emails-to-copilot.

Romeo, Giuseppe, and Daniela Conti. “Exploring Automation Bias in Human–AI Collaboration: A Review and Implications for Explainable AI.” AI & SOCIETY 41, no. 1 (2026): 259–78. https://doi.org/10.1007/s00146-025-02422-7.

San José’s Information Technology Department: Digital Privacy Office, City of. AI Handbook. 2024. https://www.sanjoseca.gov/home/showpublisheddocument/109904/638463850657330000.

Tabassi, Elham. Artificial Intelligence Risk Management Framework (AI RMF 1.0). National Institute of Standards; Technology (U.S.), 2023. https://doi.org/10.6028/NIST.AI.100-1.

“Tesler’s Theorem and Other Adages and Coinages.” In Larry Tesler. n.d. https://www.nomodes.com/larry-tesler-consulting/adages-and-coinages.

Vaswani, Ashish, Noam Shazeer, Niki Parmar, et al. Attention Is All You Need. no. arXiv:1706.03762 (August 2023). https://doi.org/10.48550/arXiv.1706.03762.

Policy

Footnotes

According to Tesler, he originally said “Intelligence is whatever machines haven’t done yet”, however Tesler’s Theorem is more well known (source: “Tesler’s Theorem and Other Adages and Coinages,” in Larry Tesler, n.d., https://www.nomodes.com/larry-tesler-consulting/adages-and-coinages.)
Selmer Bringsjord and Naveen Sundar Govindarajulu “Artificial Intelligence,” in The Stanford Encyclopedia of Philosophy, Spring 2025, ed. Edward N. Zalta and Uri Nodelman (https://plato.stanford.edu/archives/spr2025/entries/artificial-intelligence/; Metaphysics Research Lab, Stanford University, 2025).
Elham Tabassi Artificial Intelligence Risk Management Framework (AI RMF 1.0) (National Institute of Standards; Technology (U.S.), 2023), NIST AI 100–1, https://doi.org/10.6028/NIST.AI.100-1.
City of San José’s Information Technology Department: Digital Privacy Office AI Handbook (2024), https://www.sanjoseca.gov/home/showpublisheddocument/109904/638463850657330000.
Programmers at Work: Interviews Wit 19 Programmers Who Shaped the Computer Industry, Repr. with new additions (Microsoft Pr, 1989)., see page 266. Image from: Namco English: Pac-Man Approaching a Row of Three Dots. (2024), https://commons.wikimedia.org/wiki/File:Pac-Man_eating_dots.svg.
Image from: n.d., https://commons.wikimedia.org/wiki/File:Email-spam-sample_(cropped).png.
Namco English.
Image from:
in Google Cloud, n.d., https://cloud.google.com/discover/gpu-for-ai.
Ashish Vaswani et al. Attention Is All You Need, no. arXiv:1706.03762 (August 2023), https://doi.org/10.48550/arXiv.1706.03762.
Dave Bergmann The ELIZA Effect: Avoiding Emotional Attachment to AI Coworkers | IBM, 2025, https://www.ibm.com/think/insights/eliza-effect-avoiding-emotional-attachment-to-ai.
2024, https://foundationcapital.com/ideas/has-ai-scaling-hit-a-limit.
Dwarkesh Patel 2026, https://www.dwarkesh.com/p/francois-chollet.
n.d., https://owasp.org/www-project-citizen-development-top10-security-risks/content/2022/en/CD-SEC-01-Blind-Trust.html.; in NVIDIA Technical Blog, 2025, https://developer.nvidia.com/blog/how-code-execution-drives-key-risks-in-agentic-ai-systems/.; owaspLLM
in Computerworld, n.d., https://www.computerworld.com/article/4041849/enterprise-note-taking-apps-face-legal-scrutiny-as-otter-hit-with-privacy-suit.html.; 2025, https://news.bloomberglaw.com/litigation/wiretap-suits-pit-old-privacy-laws-against-new-ai-technology.
,Craig Hale published “Microsoft Admits an Office Bug Exposed Confidential User Emails to Copilot,” in TechRadar, 2026, https://www.techradar.com/pro/security/microsoft-admits-an-office-bug-exposed-confidential-user-emails-to-copilot.
United States Copyright Office Copyright and Artificial Intelligence Part 3: Generative AI Training, May 2025, https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf.
United States Copyright Office Copyright and Artificial Intelligence Part 2: Copyrightability, 2025, https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf.
Most of the studies right now are focused on academia, but we can extrapolate from these studies that what is or is not plagiarism with regards to AI is not settled. Here is one study Cecilia Ka Yuk Chan “Students’ Perceptions of ‘AI-Giarism’: Investigating Changes in Understandings of Academic Misconduct,” Education and Information Technologies 30, no. 6 (2025): 8087–108, https://doi.org/10.1007/s10639-024-13151-7.
Giuseppe Romeo and Daniela Conti “Exploring Automation Bias in Human–AI Collaboration: A Review and Implications for Explainable AI,” AI & SOCIETY 41, no. 1 (2026): 259–78, https://doi.org/10.1007/s00146-025-02422-7.

AI 101

What we’ll cover:

Goals for the Training

Defining AI

Defining AI

Defining AI

Examples of AI

How AI Works Conceptually

Real Life Examples of Each Kind of AI

Why AI Is Gaining Attention (again)

Why not before now:

Why now is different (technical):

Why now is different (social):

Why now is (not) different

Practical Concerns with AI

Genuine Dangers

Ethics

Right model for the job

You are the one using the tool

Quick Practical Advice

Putting It In Action / Hypotheticals

Thank you

References

Policy

Footnotes