The match is intense, and the two sides stay tied until the final round. Kasparov decides to try a risky strategy to throw the computer off but ultimately fails, losing the match in just over an hour. The world is shocked. Artificial intelligence (albeit a comparatively primitive form) has surpassed human intelligence in a highly publicized event. Now, almost thirty years later, humans stand no chance against chess programs. Grandmaster players and world champions of the present day don't even bother trying to beat computers. In one experiment, participants mistakenly identified GPT-4.5 as a human 73% of the time after a five minute conversation. In another, patients consistently rated an AI's text responses as being more empathetic than those written by human physicians. Some people have even begun to explore friendships and romantic relationships with chatbots.

The rapid progress of computing in the 2000s has led to an equally explosive growth in the various types of artificial intelligence. But what exactly is AI? Let's dive in.

There is a lot of philosophical debate about intelligence.

You could easily write an entire dissertation on the nature of intelligence and still have an entire follow-up book series afterwards. Philosophers have debated what exactly intelligence, consciousness, and thought are and how they should be defined for eons – I'm not kidding, go look up “philosophy of mind” and spend the rest of your day reading about it, if you like. This debate doesn't stop at philosophy, though. There isn't exactly a concise definition in computer science, either.

For the purposes of this article we will borrow the International Organization for Standardization's definition:

an engineered system that generates outputs such as content, forecasts, recommendations, or decisions for a given set of human‑defined objectives, and can operate with varying levels of automation

AI isn't new

As mentioned in the beginning of this article, computer scientists have been producing various forms of artificial intelligence for decades. Deep Blue was AI according to the definition above – it could use defined objectives (the rules of chess and the given moves) and output a decision (its choice of move) and do so in an automated manner. Search engines meet this definition. 2000s kids might remember Cleverbot, and game show fans may recall the episode of Jeopardy! featuring IBM's Watson.

All these things and many more meet the criteria for AI, all the way back to the 1950s when the first chess programs were written. 2001: A Space Odyssey came out in 1968, and HAL 9000 has lived in the public imagination as a possible doomsday instigator ever since. So why has AI suddenly exploded across the world?

Generative artificial intelligence

I won't get too far into technical details here – to precisely define the techniques that go into AI, I would need to also explain a lot of high level logical and mathematical concepts. In a very simplified explanation: over the last decade or so, we have developed technology that allows computers to utilize algorithms for the purpose of “learning.” Imagine playing through the first level of an arcade game of your choice. You probably didn't grab the manual and read up on how the game worked and plan out some strategies.

More than likely, you grabbed the joystick and jumped right in. When you fell into a pit, got cornered by a ghost, or failed to dodge an obstacle, you learned a new rule. Each time the game started over, you learned a little more: I need to have some rings to stay alive; the red ghost stays right behind me; if I press ← ← → as Raiden by rolling my thumb across the d-pad really quickly I can cheat straight to Kintaro without my opponents having a chance to recover (thanks little bro).

This is essentially what people mean when they talking about training an AI model. We can give a system big sets of data and teach it how to evaluate those datasets by providing feedback on its responses. The exact methods on how to do this vary, but in short, it's a computerized version of the above. This is what the color red looks like, select things that are red. Yes. Yes. No. No. Yes. The AI can then replicate what it has “learned,” leading us to the “generative” part. Now that it can identify what red is, it can determine whether or not something is red and produce an appropriate replication. It also refines itself as it collects more data, similar to the way we learn – only it doesn't have to worry about human issues like forgetting or becoming confused, and it can process large amounts of data much faster than you. That brings us to our next point:

Large language models (LLMs)

You may have heard this term – or the related GPT, short for generative pre-trained transformers – used to describe the chatbots proliferating the internet. These models are trained on large datasets like books or online discussions and then cleaned up by having bad data removed and safeguards installed. Or, in the case of Microsoft's Tay, naively expected to learn good things from Twitter. Either way, the models can then refer back to the data they have studied to generate a response.

For an example, I asked a few search LLMs: “is the difference between a trebuchet and a catapult important?” Take a second to think about how you would answer that question. Each of the three I asked said “yes” and proceeded to launch into great detail about how the two differ and why that is historically significant. They provide similar responses to “is the difference between a grilled cheese and a melt important.” These topics are examples of something that is highly discussed online but I daresay would be found boring by most people (thank the heavens I haven't allowed comments on this site).

This is an important point and one I wish to emphasize. The models are not truly thinking in the sense that humans are. To illustrate my point: what would you do if I asked you to drawn a clock with the hands at 3:15? Well, generative AI has a breakdown. The model does not understand what the clock is the way that a human does; it does not know that squiggly, fancy looking black lines are hands or that the small hand points to the hour and the big one the minute. Thus, it cannot extrapolate and successfully create an accurate clock. A similar issue arose a while back with wine glasses – GPT programs could not depict an overflowing wine glass because the internet overwhelmingly has pictures of half-full ones.

The way these models are programmed can, however, lead to a phenomenon called hallucination. Sometimes for one reason or another generative AI will recognize a pattern where none exists or experience an issue with its programming. This can lead to the production of false results or misinformation – try asking an LLM for a highly specific book or movie recommendation and you'll find that they often invent fake titles to meet your specifications.

Why some people are angry

I won't get into the ethical debate about learning from other people's work or plagiarism, but you should be aware that it exists. A few weeks ago I answered an extremely niche question for a user on the internet involving lost media. Imagine my surprise when, later that afternoon while revisiting the topic, I used Google to search for more information and was given parts of my own answer in the AI summary! The lost media in question was so obscure that my response on a public forum became the foundation for the AI's response – you can imagine why this would be upsetting for some individuals.

And herein lays another major issue with generative AI systems: they are only as good as the data that composes them. Take the case of poor Tay (Microsoft's failed chatbot mentioned above): in less than 48 hours Twitter transformed her from an earnest teenager to a racist, inflammatory Holocaust denier with disdain for Ted Cruz. Even a decade later LLMs are running into similar issues, forcing their developers to implement safeguards against certain behaviors by having them avoid certain topics or provide hard-coded responses. There are concerns among psychologists about the impact human-computer relationships will have, and some companies have experienced major issues due to AI-related mistakes or implemented it for dystopian purposes like monitoring whether or not employees are being friendly enough.

Lastly, the high amount of computer power required for even simple LLM requests has led to the production of giant facilities that are bad for the environment in all kinds of ways, and sometimes this cost is for bizarre or pointless reasons – like simulating a conversation with your imaginary girlfriend who is also an anime dragon princess.

But it isn't all bad.

AI is showing promising results in a lot of areas. AI models can detect some diseases from scans more accurately and earlier than doctors. It allows researchers in many fields to automate boring, repetitive tasks and analyze tons of data at scale. Despite the concerns from therapists, many individuals have reported positive mental health experiences with chatbots – people with anxiety can roleplay conversations without the fear of embarrassment, for example. As a supplemental tool, AI can give children feedback on their homework responses and explain concepts in depth as needed on an individual level.

Like so many other phenomena in computing, AI will be used for purposes noble and deplorable. It will be loved and hated, eagerly embraced and feared. Regardless of how you feel about it, I hope that you found this article helpful and informative and learned a little something, even if it was only how to cheat at Mortal Kombat II.

#computing #ai #tech #explanation

Rebecca B. – BS Interdisciplinary Studies (Marketing/Computing) Return to home