AI Fundamentals #01: What Is "AI"?
'Crash course' basics about "AI" that everyone should know to make sense of all of the noise, explained in plain English with lots of examples.
AI usually stands for “Artificial Intelligence”. The term often gets used in inconsistent ways, though. The goal of this post is to help you know just enough about what AI is & isn’t, so you can make sense of what you read and hear about “AI”.
Audience for this post: Anyone. Zero prior knowledge of AI or ML (machine learning) is required. This article’s purpose is to help you go from zero to one.
Reading tip: To skim for the key takeaways, look for this 🗝️ key icon in the sections below.
Acronyms
Within the world of artificial intelligence, we hear a lot of acronyms and terms. “AI” is an overloaded term. Sometimes it gets used to describe products that don’t really have “AI inside”. Our Glossary covers some of the most common terms you’ll hear (AGI, AI, DL, GenAI, ML, NLP, NN, RL). Figure 1 shows how these terms relate to each other. We’ll briefly cover them in this article.
Texts in white in the diagram are specific kinds of AI / ML models (GPTs, LLMs, VAEs, GANs, RNNs, CNNs, diffusion models). We’ll explain them further in AI Fundamentals #2 (coming soon).
Example: ChatGPT is based on a foundational Large Language Model (LLM). Its name means that it uses a GPT (a Generative Pretrained Transformer) for its LLM.
What’s “AI”?
🗝️ You might hear people refer to artificial intelligence, AI, as a machine that can ‘think’ and act on its own, just like humans. That’s a specific kind of AI called “AGI”, or Artificial GENERAL Intelligence. We don’t have AGI yet. We are teaching computers how to ‘learn’, but they have not yet learned how to think.
Visions of AGI have been shown in many sci-fi movies about AI, famously including 1968’s 2001: A Space Odyssey and its ‘sentient supercomputer’ named HAL 9000. Other movies include Paper Man, Logan’s Run, Star Wars (C3PO and others), Star Trek, Tron (MCP), The Terminator, Short Circuit (Johnny Five), The Matrix (agents), A.I. (David), The Hitchhiker’s Guide to the Galaxy, Iron Man (JARVIS), Her (‘Samantha’, famously voiced by Scarlett Johansson), Ex Machina (Ava), and many more.
🗝️ Despite the hype and the movies, we’re still quite a long way from achieving AGI, even in research. And not everyone is sure that we should be pursuing AGI the way we are. There are a lot of societal and safety concerns.
Other than in movies, most online chatter about AI is not about AGI. On this site, when we talk about AI, we do not mean AGI. (If we talk about AGI, we’ll call it that.)
Most of the noise nowadays about “AI” is really about one specific subset of artificial intelligence called “generative AI”, or “genAI”. Later in this post, we’ll cover what generative AI is (and isn’t).
🗝️ KEY TAKEAWAY: Generative AI is only a small subset of “AI”. When you hear or read announcements, hype, or criticisms of “AI”, be sure to confirm whether they are talking about progress towards AGI, just the Generative AI subset of AI, or other types of AI or machine learning. It makes a big difference in terms of capabilities, risks, and ethics.
Big Picture View: The World of Data Analysis, Data Science, ML, and AI
It’s common for people to refer to all solutions with data-driven features as having “AI”. The world of analyzing and using data involves much more than AI, though.
🗝️ What people often call “AI” spans many uses of data, including Machine Learning (ML), data science (DS), and analytics.
What Is Data Science?
Data Science techniques include:
search and mathematical optimization;
methods based on statistics and economics; and
‘computational intelligence’ methods.
Examples of these techniques include a lot more new terms you may never have heard of: “data mining”, “rule-based expert systems”, “neural networks” (NN), “fuzzy logic”, “evolutionary computation”, “learning theory”, and “Bayesian” or “probabilistic methods”. Examples of statistical methods include “linear regression”, “logistic regression”, “decision trees”, or “random forests”. (If you haven’t heard any of those terms before, don’t worry about what they mean! We don’t need to “go there” in this article.)
You might be wondering how a “rule-based expert system” is different from AI or ML. Here’s an example that may help.
Example: Automatically filtering incoming email into folders.
In a rule-based expert system, a human defines the rules. It could be you (the human user of the email system). Or it could be a human who manages your email system, or a human who develops the email system software.
For instance, you might specify that any email from yahoosports.com or nfl.com, or with certain words in the title (e.g. “Pittsburgh” and any of (“Steelers”, “Penguins”, “Pirates”, “Panthers”)), goes into a folder called “sports”.
In an AI / ML system for filtering email, you don’t have to define the rules in advance. The AI / ML model may be pre-trained on other people’s emails. Or it may look at the way you’ve already filtered your emails to date.
In either case, it uses math to automatically figure out patterns which indicate that certain new emails should probably always move into the “sports” folder. This is called ‘learning’.
This pattern learning might be imperfect: for instance, it could sort an email about the Pittsburgh Zoo’s penguin exhibit into your sports folder.
Natural Language Processing (NLP) includes some non-AI methods that fall under Data Science. (That’s why NLP isn’t entirely inside the AI circle in Figure 1.)
This second diagram shows many techniques that are part of Data Science, and predate what’s considered AI today, but are sometimes called AI as part of a comprehensive vision for using data.
What Isn’t Data Science?
So what’s Data Science, and what’s the difference between Data Science and AI / ML? One view is that “Data science brings structure to big data, while machine learning focuses on learning from the data itself”. That’s a bit simplistic, though. Real data science involves analyzing data, not just structuring and preprocessing it.
GOFAI (“good old-fashioned AI”) or just “AI” is sometimes used to refer to Data Science and statistical methods. Basically, data science and good old-fashioned AI exclude generative AI (GPTs, LLMs, etc.) and much (but not all!) of machine learning.
🗝️ Let’s now move past what Data Science is & isn’t. The key is to know that the main feature of AI / ML is learning.
A Brief History of Artificial Intelligence
🗝️ AI (“artificial intelligence”) is not new. It’s over 80 years old!
The first work that is now recognized as AI dates back to 1943 (81 years ago).
AI research was founded as an academic discipline in the mid 1950’s.
In 1955, Stanford Professor John McCarthy defined AI as “the science and engineering of making intelligent machines” 1.
By the mid-80’s, courses in AI theory and applications were already available in graduate computer science curricula. (I know because I took them.)
AI techniques initially included traditional symbolic AI as well as many of the techniques mentioned under Data Science. Machine Learning (ML) focuses on statistical algorithms that can “learn from data and generalize to unseen data, and thus perform tasks without explicit instructions”.
ML includes neural networks (NN), deep learning (DL), and reinforcement learning (RL). There are many types of Neural Network models (CNN, RNN, GAN) and other generative methods (VAEs, diffusion models). ML methods can also be divided into unsupervised, semi-supervised, or unsupervised learning.
If you’re curious to learn more, AI Fundamentals #2 will cover these ML methods and ML ‘supervision’. The specifics aren’t important for this article.
Where Does an AI or ML Tool Get Its Data?
This, right here, is THE trillion-dollar question about AI.
🗝️ The two most important things to know about AI and ML are:
AI and ML are already everywhere in our daily lives. Not convinced? See: “But I Don’t Use AI": 8 Sets of Examples of Everyday AI, Everywhere
AI and ML run on data - yours, mine, everyone’s. Knowing where and how an AI / ML system got its data is important. Humans create the data an AI- or ML-based system uses. Those humans are morally and ethically entitled to “consent, control, credit, and compensation”2 for use of their data and content. See: “Giving credit for the 3Cs / 4Cs where due: an origin story”
When is “AI” not Artificial Intelligence?
🗝️ You may see or hear chatter about an “AI-based” or “AI-driven” tool or product. It might use one of those “good old-fashioned AI” (data science or analytics) methods under the hood instead. That could be a good thing. More on this later!
What AI Is Today
There is a saying that “AI is whatever hasn't been done yet.” Some technologies which used to be considered as AI are now so common that we consider them normal and non-magical, and not “AI”. This is called the “AI Effect”.
Example: People learning ML often try to create models to recognize hand-written digits. But optical character recognition (OCR) for printed documents is pretty standard nowadays. Most people don’t think of OCR as AI any more; it’s no longer magical.
Today, AI is considered to include algorithms for ‘learning’ that cause computer systems to simulate some aspects of human intelligence. Data scientists and software engineers use these algorithms with data to develop AI / ML models and integrate the models into software applications (tools, services, or products).
Low-risk applications can be designed to take direct actions based on AI / ML model results or recommendations. This may be called “autonomous intelligence” or “automated intelligence”.
Example: An email application moves an email flagged by ML as suspicious to your spam folder. If you decide the ML is wrong, you simply move the email back to your inbox. It’s low risk, and the consequences of the ML being wrong are minor.
Higher-risk practical applications with ML and AI models often have a human in the loop to review and decide before any actions are taken.
Example: A bank’s machine learning system may issue a recommendation for whether to approve or disapprove a mortgage application. The consequences of the ML being wrong are serious for the applicant. A human mortgage lender reviews the application and the ML recommendation before making a decision.
However, in some cases (such as self-driving cars or safety features), actions recommended by AI or ML need to be taken immediately, without time for a human to review. AI / ML models which are developed for these safety-critical cases require extraordinary care.
Example: A car with a safety feature detects when a pedestrian steps off the curb right in front of the car, or walks behind it while the car is backing up. The car’s safety system has to automatically apply the brakes immediately to avoid hitting the person. There’s no time for human review in that loop.
Research in so-called "intelligent agents" is active. These agents can ‘perceive’ their environments and take actions to maximize their chances of achieving the defined goals. An intelligent agent may also use learning to improve its performance. Defining the goals correctly is critical!
Generative and Non-Generative AI
Generative AI is simply “a subset of AI that creates new content”.
🗝️ To determine if an AI tool is generative, look at whether the AI is focused on:
understanding or analyzing what exists (that’s “non-generative”), or
creating something specific that did not exist before, based on what did exist before (that’s “generative”).
Generative AI nowadays can be used on many “modes” of data, such as: text, images, audio, video, or source code. GenAI can even be used to generate “synthetic data” (for more on this, see Lakshmi Veeramani’s 6P article on synthetic data). GenAI can also be “multi-modal”, which means that it can generate an output with more than one of these modes at once, from the same input.
Example prompt: “Explain AI to people without technical backgrounds in software”.
A single-mode genAI text tool could generate a few paragraphs or song lyrics.
A single-mode genAI video tool could generate an explanatory video with animated images.
A multi-modal genAI tool could generate a video with song lyrics, sung by simulated voices; simulated musical instruments; and simulated dancers or actors in front of AI-generated scenery images.
To clarify the difference between generative and non-generative AI, let’s look at text.
Natural Language Processing (NLP) is a branch of Machine Learning focused on understanding speech or text. In brief, NLP analyzes existing text content. Generative AI builds on the NLP analysis results by adding the ability to generate new text content. (AI Fundamentals #2 has more about NLP.)
Examples:
Google or Bing search - You type in a question. A NLP model in the search tool analyzes your question and uses it to select search results to show you. (In the newest versions of Google or Bing search, by the way, you’ll see “AI Overviews”. Those may be generative.)
Apple’s Siri voice assistant, Amazon Alexa device, or “OK Google” voice interface - You speak a command or question to the device. A NLP model in the device analyzes your speech to convert it to text. Then it uses the text to determine how to respond to execute your command or to answer your question.
Website chatbots - You type in a support request or question. A NLP model on the website tries to ‘understand’ your request. It then uses its ‘understanding’ to look up the right options to present to you to help solve your problem.
🗝️ NLP does not create texts which did not exist before. GenAI and its large language models (LLMs) can.
If all a chatbot does is look up keywords in a list, and show you the top 5 matches for you to choose from, that chatbot is not using GenAI - it’s conventional NLP.
If the chatbot appears to generate new unique text as its response to you (perhaps along with those top 5 choices), then it may be using GenAI. (Or it could just be NLP, automated with a template).
Large Language Models (LLMs) are now being “trained” on massive sets of data (many terabytes), and can learn patterns and structures. A LLM will still try to interpret speech or text, just as non-AI NLP does. And it will try to apply the patterns and structures it learns, to create (or ‘generate’) new content.
What Can Go Wrong With AI
AI tools are, at best, as good as the data they were trained on. When faced with input that doesn’t match up well to its training data, AI and ML systems flounder.
Example: A self-driving car system trained on southern California weather won’t recognize snow-covered roads and will not perform well.
Among all of the current types of AI, Generative AI is the most prone to this. To quote
, “GenAI sucks at outliers” 3. Sometimes new genAI content is simply nonsensical. This is often called “hallucination” or “confabulation”. Here are some of the mistakes a genAI tool could make:Respond to a research question with a citation of a paper or article that was never written;
Tell you “facts” about a person which aren’t accurate;
Make up a new quote of things someone never said, or create a new video showing them doing things they never did;
Provide a code snippet that’s invalid and won’t compile,
Create an image of a human with too many or too few fingers, arms, legs, or teeth.
All 5 of these mistakes by genAI tools have already happened many times this year.
Scammers are increasingly using #3. Such AI-generated audio recordings or videos showing someone doing or saying things they didn’t do or say are called “deepfakes”.
Example: You get a phone call that is supposedly from a family member. The voice says it is an emergency situation and they need you to wire them money right away. The call is a deepfake. It’s not really from your family member.
Most genAI tools are not transparent and do not mark the creations from their tools as AI-generated. This is slowly changing.
In the meantime, it’s hard for us to know if what we are seeing or hearing is real or a deepfake.
🗝️ The 2 most important things to know about the risk of AI-generated fakes or hallucinations are:
If you’re the one using the AI tool, verify what it gives you. Don’t assume it’s accurate. (And be transparent about labeling what you created as ‘generated by AI’.)
If you’re seeing or hearing content created by somebody else, be skeptical. Check it out before you share it or act on it. (Think about establishing a family code word in advance, to use for validating if an emergency call or message is real.)
What Makes AI or ML “Good”?
🗝️ Goodness of an AI / ML solution depends on many factors. Ideally, it meets 6 criteria of goodness:
Solves a real problem or meets a genuine need.
Solves the problem well — better than other non-AI / ML solutions.
Was developed ethically (for instance, does not allow use of, and was not trained on, stolen or scraped data; and it prevents and mitigates biases).
Is efficient and uses minimal resources (computing, energy, water).
Appropriately involves humans in the loop and handles risks.
Has a ‘closed loop’: captures feedback, uses it to improve quality, and is monitored by its developers for drift, biases, and outlier scenarios.
Model Accuracy
Accuracy is one way we measure how well a statistical or AI / ML model solves a problem. For instance, there are two ways classification models can be wrong:
“False Positive”: Classify something as being X when it is not X.
“False Negative”: Classify something as not X when it is X.
Example: A ML model looks at a scan and classifies whether spots on the image are cancerous or not.
False Positive: Saying it is cancer when it isn’t. This will cause unnecessary stress, costs, and physical risks. A patient may have further tests and maybe treatments (with side effects) that they don’t even need.
False Negative: Saying it isn’t cancer when it is. This error will mean that the cancer will grow untreated. The patient will not be diagnosed until much later (if at all) when the cancer is much harder to treat, which could be fatal.
Models can be adjusted so that the more serious consequences (false positives or false negatives) are given more weight. This ‘tuning’ isn’t always easy to do, though.
In this example, what is the right weight for classifying cancer from images? Which type of error is more important, and how much more important?
Avoiding a false positive and not putting people through unnecessary tests and treatment?
Avoiding a false negative and leaving someone undiagnosed?
Model Biases
Even if they are highly accurate overall, AI and ML models can have biases that can harm some sub-groups. This can happen if they are built on poorly-selected data or aren’t properly trained.
Example: Many AI/ML models that use images are less accurate on darker skin.
People might not be recognized as pedestrians by a self-driving vehicle, which would fail to brake for them.
They could be locked out from opening a door or a computer they are entitled to use.
An injury or skin condition may not be properly recognized and treated.
So AI/ML models are also evaluated for goodness by looking at what are called “bias-variance tradeoffs”.
Models aren’t perfect the day they are released, so accuracy and bias are always concerns. Models can also become outdated quickly, or when they are used in situations that their training data didn’t cover.
🗝️ AI/ML techniques are very powerful tools. But power tools can be dangerous. And not every problem needs a power tool.
AI should not be a sledgehammer that we run around with, looking for nails.
Many Interpretations for “AI”
“Artificial Intelligence” is not the only “AI” worth considering. Many other types of data analytics can be useful in solving problems. And many people resist the idea of AI possibly replacing humans (which would be AGI), and have come up with other terms to make it clearer that AI isn’t meant to replace us. Here are some alternate “AI” terms you may hear:
Augmented Intelligence - intelligent tools aiding humans, not replacing them.
Automated Intelligence - intelligent automation to improve efficiency & reliability.
Adaptive Intelligence - tools taught or programmed to respond to changes in their environment.
Analytical Intelligence - tools that analyze data and provide insights (also sometimes called BI, or Business Intelligence).
Actionable Intelligence - software that not only diagnoses a problem, it provides recommendations on what to do about the problem.
The latest twist on “AI”: Apple introduced “Apple Intelligence” in mid-2024.
🗝️ In short, the A in “AI” can mean lots of things. Don’t get hung up on the name. The most important thing to focus on is using data ethically for effectively solving a problem and efficiently adding value.
Some Useful A’s That Don’t Need AI
1. Analytics – analyzing raw data to draw conclusions about that information. Many meaningful analytics can use ML to help steer business decision making. But analytics do not require ML or AI.
Example: A simple graph of sales data can show an unusual drop or a trend over time. The graph indicates a problem that needs to be investigated.
2. Automation – When we simplify a repetitive process or run it, we are creating automation. Automation can make tasks more reliable and more efficient. But automation does not require ML or AI.
Example: A programmable home thermostat can be set to raise or lower the house temperature at certain times of day. This keeps the home comfortable while reducing power costs, daily human effort, and risks of forgetting.
🗝️ The key point is that solutions to many problems do not require machine learning or artificial intelligence to be useful.
Always start with “the simplest thing that could possibly work”.
If you don’t need ML or AI to solve your problem, don’t use it!
Sometimes the simplest solutions are the best. They can be:
more efficient
easier to understand (what we call ‘explainable AI’)
more accurate at solving the problem (which would mean there’s really no reason to use AI or ML)
Conclusion: There’s No Free AI Lunch
There’s so much buzz about “AI”. Some people even think a product won’t sell nowadays unless the company claims the product has AI inside. But a solution doesn’t need to use artificial intelligence to be valuable.
When you hear people talk about AI, you can cut through the hype by asking:
what’s intelligent about it and
how well it solves the actual problem.
🗝️ Find out: Is it analytical? Is it actionable? Is it ethical? Does it truly learn? What data was it trained on, and when? How often does it hallucinate or make mistakes? What are its biases? How are the developers keeping the tool up to date, improving its accuracy, and reducing its biases? Does the tool explain why it came up with the conclusions or recommendations it did?
Always keep in mind that AI isn’t free (even if you can use the tool for free).
Acquiring (hopefully ethically), cleaning, and processing data to feed AI/ML systems is expensive.
Training and retraining AI/ML models is expensive.
The expense of AI includes money, time, and environmental resources for all of that computing, data sourcing, and storage.
🗝️ Don’t be afraid to ask basic common-sense questions.
How does the AI-based tool’s accuracy & benefits compare to non-AI solutions?
What’s the risk or cost of using AI to solve this problem? Is it worth it?
What’s Next?
We hope that this article has given you some grounding on what AI is and isn’t. Let us know if it helps you navigate our AI-influenced world, or if it raises new questions.
If you’re curious to learn more, please see AI Fundamentals article #2: What Kinds of AI Are There? (coming soon).
What questions do you have about AI that we can answer next?
Would another format (e.g. video) be better for you for learning about this topic?
References
Credit for the 4Cs (consent, control, credit, compensation) phrasing goes to the Algorithmic Justice League (led by Dr. Joy Buolamwini).
Credit for the original 3Cs (consent, credit, and compensation) belongs to CIPRI (Cultural Intellectual Property Rights Initiative) for their “3Cs' Rule: Consent. Credit. Compensation© 2017.”
Gary Marcus, 2024-08-01, “This one important fact about current AI explains almost everything”: “The simple fact is that current approaches to machine learning (which underlies most of the AI people talk about today) are lousy at outliers, which is to say that when they encounter unusual circumstances, like the subtly altered word problems that I mentioned a few days ago, they often say and do things that are absurd. (I call these discomprehensions.)”
“The AI world pretty much divides into two groups: those who understand why current machine learning techniques suck at outliers, and therefore struggle at things like driverless cars and high-level reasoning in unusual circumstances — and those who don’t.”
“You can’t think we are close to AGI once you realize that as yet we have no general solution to the outlier problem. You just can’t.”