What is AI really? Throughout this article, I will remove the hype and get to the most honest answer ever.
Artificial Intelligence, or AI, or at least the first version of how we think of it today, was “invented” in the 1950s…a long time ago. Since then, various computer scientists and groups have worked on different iterations, often using different names, including machine learning and neural networks.
AI came into intense popular use and understanding with OpenAI’s public release of ChatGPT in October 2022. Since then, it seems that a day has not gone by without nearly every sentence I hear spoken or read having the word AI in it.
I cannot give a presentation to any group without mentioning AI. They will not hear of it. I must speak about AI. They want to be scared and mesmerized by AI. If I do not scare them a bit, they do not think they got their money’s worth.
I get it. AI is everywhere, and the improvements and threats it is bringing are substantial and real.
I do not overly love AI. I do not hate AI. I like to think of myself as an AI realist.
Last year, I was famous for saying, “Although AI is coming, how you are likely to be compromised today does not involve AI. If you get compromised, it is likely to be from the same ole threats that have worked for decades.”
Well, that is not true anymore. AI is here in a big way. Most phishing kits are AI-enabled. Most threats coming your way have some AI-enabled component involved and every cybersecurity threat we face will just grow in using more AI forevermore (more on this in the next article).
But what is AI?
What makes AI AI? What are the defining characteristics of that type of software that makes it AI?
I really struggled with the definition. I am a pro-AI type of person…but with serious, realistic expectations thrown in. I see way too much hype. Along my AI journey, I have read about a dozen books on AI (both pro and con) and what must be hundreds of articles. I have spent months arguing back and forth with my mentor (who knows a lot more about AI than I do) about the definition of AI.
Nearly every vendor in the world…certainly every cybersecurity vendor…is saying they have AI. But it is clear that many do not and are just stretching the meaning to be able to add the term to their sales copy. To figure out who has AI and who does not, you need a definition. It needs to be clear and definitive.
Most of the definitions of AI I hear, read, and have even spoken myself were not strictly defined. I have heard myself say what I thought AI was in a presentation and at the same time, heard my inner dialogue argue against what I just stated. For years, it has been a struggle.
Artificial Intelligence…as a term…is mostly a marketing term. Most people using it really cannot explain what it really is…technically.
If you are going to speak about AI in a thoughtful way, we need to speak about what it really is…and is not…beyond just some commonly used marketing name.
You will hear most people talk about AI being software or a service that simulates or even exceeds human intelligence, including human reasoning, decision-making, and even (although this is a further stretch) emotions.
Thinking that software can have emotions (or even simulate them accurately) is really pushing it, but it is the most common definition we encounter and hear about.
“It is software that acts as a human would.”
I have said it myself a hundred times without really thinking about it. But lately, I feel disingenuous saying it.
AI is not human. It is software. It is software programmed by humans to do only what it is programmed to do. It does not think. It does not think up something brand new out of thin air. It is only going to do what it was programmed to do, even if it appears that it is doing something “truly new” or appears “human”. It is not human. It cannot feel emotions. It cannot feel shame, anger, love, or jealousy. It cannot do something radically different from what it was programmed to do.
There have been some stories where AI supposedly threatened or tried to bribe someone who was telling the AI that it was going to shut the AI off. “Wow…the AI is threatening someone!” “Oh…the AI is threatening to reveal that someone is having an affair in order to prevent being shut off!” “It is human and it is evil!”
We are trying to make people think that AI is human or becoming human.
It is not.
AI has no emotions. It does not feel anything. The AI cannot feel threatened to feel the need to threaten someone back. It is not capable of feeling fear. It cannot “feel” the need to “threaten” someone. All it did was respond to some input text with some other text that is commonly sent by humans in response to the other text sent by humans. No more, no less. No human emotions involved…at least from the machine’s side.
But we humans love to anthropomorphize and put humanity on everything we interact with. We treat our pets as if they were human. We talk to them as if they completely understand our entire sentences. We give our cars and boats names. We see our AI LLMs responses through our human lens as being “human-like.”
We are not even consistent in our beliefs. We ignore the mistakes and embrace humanism. When AI makes a mistake, like putting six fingers on a hand, it is a computer program. When it says something similar to what a human might say, we think it might be human and sentient.
But AI is not human. It is just software.
It is not to say that AI cannot come back with things that “feel” human. It is not to say that future AIs cannot simulate emotions close enough that the distinction does not matter. But AI is not human. It is software. It is pre-programmed software.
So, what makes AI AI?
Some people say it is capable of making new decisions that it was not pre-programmed to do. I have even said that. There are some anecdotal stories…around chess, Go (e.g., a very popular logic game first popularized in Asian cultures), and protein folding, that make it seem as if the AI made new decisions that humans did not program it to do or would not have done. But that is not true.
I have read that the Go-playing AI that first beat the human champions made a move that none of the Go experts had thought of or were expecting, and that pivot, supposed “mistake”, won the game. So, is that not proof that the AI was capable of inventing some new move that it was not pre-programmed to do or that exceeded human intelligence?
Not really.
All that happened was the “AI” saw a new pattern for winning Go where humans did not see one before. It was pattern matching. It saw a pattern that humans did not see before. But it is not like the AI was not programmed to follow specific rules and look for new winning patterns. The AI did what it was programmed to do.
The AI did not become self-aware, write new programming instructions, or exceed what it was programmed to do. It is not like the Go-playing AI suddenly decided to make up new ice cream sundaes or plan a fabulous Sicily vacation. It only did what it was told to do, even if the actions appeared unexpected or exceeded expectations.
AI is very good at seeing patterns…even better than humans. This is a strong clue as to what AI is.
What Is AI?
If I try to figure out what makes AI AI and everything else not AI, it is this.
All AI (AFAIK) is a general-purpose probabilistic pattern-matching engine.
Let me explain more.
All AI (AFAIK) is based on finding patterns from data that has been inputted. Based on the input, the AI uses probabilities to determine how to respond with output.
For example, if an AI LLM consumes the Internet and you ask it a question (i.e., prompting), it will build its response one word (i.e., token) at a time by looking at the input and deciding what is the most likely first word to include in the response, and what second word is the most likely second word of a particular response, and so on. It does not understand language or speech. It is simply able to figure out what words seem to be next to what other words when responding to a particular prompt.
If I ask AI to complete the English sentence, “I am going …” it has thousands of words it can choose from, including home, work, school, crazy, and so on. And based on what data it “consumes,” it will rank the probabilities of the different words finishing that sentence.
That is what AI does. It consumes data and provides answers based on probabilities. The answer could be words, pictures, audio, or videos. But it does not truly “understand” what it is providing. The Internet is full of examples of very silly-looking AI “hallucination” mistakes. It has a hard time making hands and feet in pictures. It has a hard time showing a glass of wine filled up to the brim. It has a hard time adding two plus two. It has a hard time showing a winning poker hand with four hearts in it. I do not understand poker. It does not “understand” anything. It is just providing answers based on probabilities.
If you ask AI about the “birds and the bees”, even if you truly mean real birds and real bees, you will likely get a response about sex, attraction, and reproduction. That is because in all the writings about the birds and the bees in the same sentence, most deal with those subjects. The AI does not think about sex, attraction or love. It is not human. It is just software that can predict, based on probabilities, what words and responses are common to particular inputs. No more, no less.
It does pattern matching…based on probabilities. I add the qualifier “general purpose” to indicate that it is pretty good at pattern matching across a wide range of subjects. My mentor suggested this addition, and I liked it.
Now, I am not trying to diminish advanced pattern-matching. AI has already changed the world and will change the world in wonderful ways, both good and bad. AI will fundamentally change most things we do today, improving them. It will make you more productive. It will get you answers faster. It will make products better. It will better automate tasks. It will fill in gaps where we have shortages. It will cure diseases. It will cure cancer. It will make better answers. It will solve previously not-understood mysteries, all because it is able to see patterns where humans did not previously see patterns. Not in everything, but with a bunch of things.
I think humans will still be very much needed and excel at many tasks, including true creativity and inventions. I think humans will still find solutions and inventions where AI does not. Humans will excel where culture, emotions, and true understanding are needed.
AI will also allow attackers to improve their attacks. It is coming very quickly (I will talk about this in my next article).
But I first wanted to start by setting the definition for what AI really is. I think I have it:
AI is a general-purpose probabilistic pattern-matching engine.
I could be wrong. AI could be more than a probabilistic general-purpose pattern-matching. But so far, that is the single base requirement I can think of that defines AI more than any other defining trait.
Am I wrong? Is there more? Is there an AI that does more than advanced pattern-matching?
