Can you tell if this header image was created with AI or not?
Outline of today’s workshop
Learning Objectives
Understand basic AI concepts and terminology
Develop critical thinking skills for evaluating AI output
Learn to interact safely and effectively with AI tools
Part 1: Introductions (5 minutes)
who the eff is Melanie?
brief class introductions, what your level of familiarity is with AI and why you came to class today
Part 2: What is AI? (10 minutes)
Basic definition and concepts
Examples of non-AI vs. AI systems
Group discussion
Part 2: AI in Daily Life (5 minutes)
AI scavenger hunt
Group discussion
Part 3: Critically Evaluating AI Outputs (25 minutes)
Key questions to ask about AI systems before/during use
Case study analyses: Can you tell which of these things is AI-generated?
Discussion of AI hallucinations & biases
Group discussion
Part 4: Wrap Up (5 minutes)
Key takeaways & review
Resource sharing
Q&A
What is artificial intelligence?
A general definition
A common definition of AI is computer software that is designed to mimic what we understand as intelligence.
What do you understand as the intuitive definition of intelligence?
Although definitions will differ slightly, a commonly understood definition of intelligence is the ability to think abstractly, reason, and solve problems. By extension, artificial intelligence is the ability for computers to do these things, particularly for problems or concepts typically considered important to humans.
AI vs Traditional Programming
To illustrate the difference between traditional code and artificial intelligence, let’s think about something like a website.
A website can be either static (e.g. this workshop page) or dynamic (e.g. any video platform like YouTube or Instagram). Generally, a website is made with computer code that combines information, images, and interactivity to serve as a kind of landing place for a particular idea, place, person, and so on.
We can probably agree — even if we have slightly different definitions of what intelligence is at its core — that the concept behind a website doesn’t seem like a kind of intelligence. Something important to highlight here is that developers tell a webpage how to behave and act — each interaction, each piece of functionality is hardcoded to be the way that it is.
Now let’s think about an example of AI — you have probably heard of ChatGPT. ChatGPT is a chatbot powered by something called a large language model (LLM). LLMs are a kind of AI — the behavior you see in ChatGPT isn’t hardcoded. The model behind it is first trained on a dataset, and then predicts its output based on that training data. In this case, a target language like English.
Discussion Question 1: Do you agree or disagree with my given definition of intelligence?
Discussion Question 2: Do you think ChatGPT is exhibiting intelligent behavior?
Quick note on terminology: If you have heard the term machine learning before, it is considered a sub-field of AI. Generally speaking, all machine learning is AI, but not all AI is machine learning.
Any questions thus far? Let me know!
AI in Daily Life
Where have you seen AI come up in your life before?
AI systems are becoming more and more common in daily life. While some of the simplest kinds have been around for over a decade (e.g. email spam filters), these systems are beginning to get more face time. Whether it’s social media, AI assistants, or many other common applications, you probably use AI almost every day.
AI Scavenger Hunt
Let’s take 5 minutes to break up into groups. Each group will make a list of all the AI software they can identify having used in the past two days. The team with the longest list will be considered the winners of the scavenger hunt. Afterwards, we will all discuss our findings.
Discussion Question 3: Was there anything that surprised you about any of the examples we identified? Are you surprised or unsurprised by how often AI appears in your life?
Becoming a Critical Evaluator of AI
This next section is going to deal with how to become a critical AI user. My position on AI tools is really straightforward — I think they offer tremendous value to us if used very carefully.
The goal of this section is for you to become more careful about how you use these tools and evaluate their output, so you can safely get the most out of them for your benefit.
Important questions to ask as an end user
What is the purpose of this AI system, and what are its limitations?
Does this system use my data to train its models? If so, am I comfortable with that? Can I turn that off?
What sort of data do I feel comfortable giving to this AI, based on who the owner is and whether or not my data is being used?
If I’m relying on the output of this model in any way that impacts my life (however small), how confident am I that the output is fair and beneficial to me? How can I use the tool in a way that guarantees it will have beneficial effects for me?
Who’s that AI?
Now we’re going to do an activity where we take a look at different outputs of different generative AI models, and see if we can identify what is generated and what is real.
Example 1: AI-Generated Voices
For voices, we will listen voices in two different contexts we might be familiar with. After each example, we’ll guess which one is real. After the exercise, we’ll talk about ways to identify which is an AI.
Example 1.1: Newscaster
Example 1.1.a
Example 1.1.b
Example 1.2: Audiobook Narration
Example 1.2.a
Example 1.2.b
Example 2: AI-Generated Images
For images, we will take this quiz from CNET: https://www.cnet.com/pictures/ai-or-not-ai-can-you-spot-the-real-photos/
Example 3: AI-Generated Videos
For videos, we will watch this video (certified non AI-generated creator): https://www.youtube.com/watch?v=oWLHAuUoYqQ
Example 4: AI-Generated Text
Example 4.1: AI or not AI? That is the question
4.1.a: Oh, treacherous Fate, why dost thou weave
A tapestry so full of cruel deceit?
The stars above, they sparkle fair and bright,
Yet oft their light betrays the wanderer's path.
Am I but clay within thine artful hand,
Molded to thy whim, shaped by thy caprice,
Or doth my soul hold power o’er its course,
With will enough to wrestle thee aside?
4.1.b: O, what a noble mind is here o'erthrown!
The courtier's, soldier's, scholar's, eye, tongue, sword;
The expectancy and rose of the fair state,
The glass of fashion and the mould of form,
The observed of all observers, quite, quite down!
And I, of ladies most deject and wretched,
That suck'd the honey of his music vows,
Now see that noble and most sovereign reason,
Like sweet bells jangled, out of tune and harsh;
That unmatch'd form and feature of blown youth
Blasted with ecstasy: O, woe is me,
To have seen what I have seen, see what I see!
Example 4.2: AI scientist or science?
4.2.a: Title: Impacts of Climate Variability and Change on (Marine) Animals: Physiological Underpinnings and Evolutionary Consequences
4.2.b: Title: Adaptive Morphological Divergence in Marine Fauna: Insights from Evolutionary Transitions in Coastal Ecosystems
Tips & Tricks for Identifying AI
Practice regular skepticism when you are consuming anything digital, or when you can’t confirm the source of information (this isn’t new, but know that “fake news” just got a lot more insidious)
Keep in mind that identifying AI is some balance of trusting your gut, and knowing that it is sophisticated enough to trick you.
Look for inconsistencies!
For images and videos, focus particularly on the appearance of the pixels (do they look blurry, or finely resolved, like a real camera can do?), hands/fingers, logos/text, physics constraints, and overall consistency/strangeness of small details.
For voices, focus on the cadence and quality of the voice. Can you really hear the emotion? Is there weird inflection? No inflection?
Remember that human-made content (voices, images, etc) are imperfect. Sometimes, you can identify an AI not by a mistake, but by the fact that the content seems too flawless.
Err with caution — always take your suspicions seriously and seek verification in important scenarios, and always verify an AI output before using it in safety-critical situations (e.g. health information)
Hallucination & Bias
AI can do something called hallucination. This is similar to what happens when humans hallucinate, but it’s less obvious. A generative model is meant to be trained for accuracy, and there are ways engineers can ensure this.
But even with the most robust training methods, sometimes, the models make mistakes. And right now, instead of the models knowing they’re making something up, like how a human can know this, they state that information confidently just as they would any of the right information.
Similarly, if an AI is biased — for example, a hiring algorithm preferentially selects men — the model will not be aware of this. There are researchers working to mitigate this, but this is still a problem.
Resultantly, it is absolutely crucial to be aware that these systems make things up and can be biased, just like humans. They are not perfectly objective. Although they can help us become more objective over time, their degree of truthfulness is still in its infancy and we must be skeptical of claims to the contrary.
Tips for Successful & Useful AI Interactions (and other AI safety tips)
Chabots/Generative Images
Many of the big player AI offerings (e.g. Gemini, ChatGPT/DALLE, Midjourney) are generally good options for image generation and chat assistance
If you’re really concerned about security, you can explore options to use personal compute to spin up open source models or build your own at home. Ask me more about this if you’re interested!
Well-made Chatbots (e.g. ChatGPT and Claude are two I use weekly) are very helpful with instructions for completing particular tasks, serving as “learning assistants” for topics they’re trained on (e.g. code-writing), brainstorming, helping you draft an email, and similar things.
Make your prompts as specific and contextual as you possibly can. Chatbots do better with more instructions (similar to a person).
If you didn’t get the output you wanted, re-prompt the model — this also helps with better results.
If the bot is hallucinating, report it to the developers, if this is an option. They will use that chat to engineer better responses.
Protecting Your Image Data Online
For artists, or other concerned individuals — if you are concerned about copyrighted or sensitive image data being scraped by an AI, you can use data poisoning tools like Nightshade to protect your work. These tools can imperceptibly alter the pixel data of your images, which will make them incompatible with the model training process. I strongly advise people to be aware when publishing any kind of data on the internet that unless you place privacy restrictions on it (e.g private social profiles) it is vulnerable to being scraped by AI.
I tell everyone I can these days not to pick up calls from numbers you do not recognize. Well-made AI can clone voices in a way that is very convincing. By refusing these calls, you prevent your voice from being potentially scraped and prevent others from attempting to trick you with cloned voices.
When it comes to recommendation algorithms, if you do use them, I recommend making use of the limited controls these platforms offer to users in terms of algorithmic content tweaking (e.g. If you don’t like seeing something on your feed, make sure you mark it as such!). Please remember these platforms aren’t designed with benefitting users in mind — currently, the most popular platforms bake in those features ad-hoc only. I also strongly recommend checking out decentralized social media or highly customizable social media, e.g. Bluesky or Mastodon.