What Is Artificial Intelligence?
Artificial Intelligence (AI) is not magic—it’s a collection of tools and computational methods used to build software and machines capable of performing digital tasks that previously required human skills. Computers utilize various statistical processes and training data to mimic cognitive abilities traditionally associated with the human brain, enabling interactions that appear intelligent. While some modern manifestations of AI resemble science fiction, the models and development techniques behind them stem from straightforward, even humble, beginnings. This article explores how AI works, its historical development and applications, what differentiates AI technologies from other software, and some speculation about future AI applications.
AI technologies have existed for decades and have appeared in early consumer-facing products such as computer games (like Chess and Pac-Man), household devices (like the Roomba), and productivity tools (like Google Maps). Think about the first time you were rerouted around a traffic jam on your daily commute or when your music app suggested a new artist immediately after you bought a song. That’s AI in action—a system of rules identifying patterns in your datasets (your calendar, purchases, behaviors, and preferences) which can respond with relevant recommendations. In fact, it’s the personalization of these services that makes the computer system feel ‘intelligent.’
Software has always been capable of storing data, but reading that data and applying rules to generate personalized outcomes is what transforms software into artificial intelligence.
Early Systems: Rule-Based and Rigid
These early systems were highly prescriptive. They targeted common tasks—managing schedules like a virtual assistant or guiding users via chatbot to technical support—but weren’t actually very ‘intelligent’. They were unable to step outside the bounds of rules the programmers built into them. For example, when receiving a flight confirmation, a basic assistant would always ask if you wanted to download a map or book a hotel, no more, no less. If you asked for help updating that same airline reservation, you always got the same few options about changing or cancelling your ticket. These systems were useful when their assumptions matched your needs, but they lacked flexibility and problem-solving capabilities. Here’s an example of a common game that digitally aware users would play with chatbots:
You: I need to change my flight.
Bot: I can help with that. What’s your reservation number?
You: It’s under my emotional support platypus.
Bot: I’m sorry, I didn’t understand. Please provide your reservation number.
You: His name is Mr. Quillington. He handles all my travel.
Bot: I’m sorry, I didn’t understand. Please provide your reservation number.
You: He says he prefers aisle seats.
Bot: I’m sorry, I didn’t understand. Please provide your reservation number.
Once the inputs stepped outside of the specific keywords the system was trained to accept, it was useless. Still, these early rules-based systems laid the groundwork for the more advanced machine learning models behind today’s assistants, apps, and conversational agents that populate our online life today.
Machine Learning: Following the Rules
Early AI systems were rule-based, using a single, invariant algorithm to produce results. Some early machine learning systems—such as digital chess players or robotic vacuums—relied on artificial neural networks that mapped out all possible actions and responses. Algorithms directed how the system navigated the ‘neurons’ (or decision points) in this network to select the next move.
Three key traits defined these systems:
1. Problem Mapping: Tasks like playing chess or recognizing images were modeled as decision trees, represented by the neurons or nodes in a neural network.
2. Preferred Paths: Some paths (or moves) are more desirable than others depending on the previous moves and decisions.
3. Training Data: A set of example data was available to be mined to identify optimal good or bad moves—this is known as supervised learning.
Training involved applying weights to connections in the network, making some outcomes more likely than others. The AI works by searching ahead and selecting the move with the highest probability of success based on its training.
Despite being task-specific, early neural networks could mimic human behavior surprisingly well as long as they stayed within the bounds of their training. That task could be fairly sophisticated (playing chess, mapping a room’s shape, creating social media or performance media recommendations) but the neurons and parameters (percentage predictions of ‘right’ or ‘wrong’ moves) in the net were optimized for one task only, based around whatever kind of data was used to train the system. These systems, known as narrow AI or weak AI, were highly effective for certain tasks but not generalizable without retraining.
As you can imagine, with a lot of possible moves in a system that keeps updating, and the large amounts of data required to represent good and poor moves, this kind of decision making requires a lot of computing power. The computer science and data science ideas that specify these models trace back to classical statistical techniques like linear regression. However, real-time operation only became feasible only with the advent of faster chips and parallel computing architectures. One notable milestone was IBM’s Watson, introduced around 2011, which combined robust hardware and refined neural net architecture to tackle complex problems.
From Watson to Generative AI
Watson was an early language model designed to handle natural language processing (NLP) but was still limited compared to today’s generative models. Still, it’s worth considering the original implementations of language models such as Watson that led to the generative large language models (LLMs) that we consider to be AI today because it helps us understand the ways in which these systems are optimized. Watson used supervised learning and linear regression, with human experts identifying the features and datasets for optimization and training. Whether Watson was playing Jeopardy! or analyzing healthcare documents, it did so based on human-curated data and explicit feature selection. Often a team of human experts crafted training data sets on which the system would become expert, and identified features of that data that made it a ‘good’ or ‘bad’ example.
A significant shift in AI models and AI research occurred when AI began to consider context, not just content. For example, identifying whether a review is positive requires the model to pay attention to whether the word “good” is preceded by “not”, which changes the meaning of the phrase entirely. These meaningful contextual units of language — tokens — are the building blocks of large language models (LLMs). They also provide a way for a statistical system to choose its own features rather than requiring tedious hours of human labor to define them. Models which can construct tokens in this way are called Transformer models, and form most generative LLMs in popular use. Vendors charge by token usage, making it important to understand how they affect accuracy and cost.
Natural Language Makes Its Own Rules
Supervisory machine learning achieved remarkable results in those specific domains that were important or lucrative enough to warrant the extensive training burden and computational time. Commercial applications like Netflix recommendations, Google Maps, and Apple’s Siri all used non-generative models and worked just fine. Responses from these systems were bound to the set of known answers that the system had been trained on. This is known as weak AI or narrow AI and it still represents the lion’s share of AI applications that users encountered as recently as 2024.
The revolution in the AI industry came with deep neural networks, which use unsupervised reinforcement learning to learn from massive datasets (e.g., the entire set of documents on the open World Wide Web). Deep AI models contain hundreds of thousands of nodes and millions of parameters; they independently determine the best length, content and context of token structures, relationships between nodes, and optimal pathways—continuously improving by comparing outputs to known inputs.
While the inner hidden layers of nodes in these deep learning models are not represented by a known, supervised set of training rules or weights, a generative ai system can handle an unrestricted set of questions and instructions in almost any colloquial syntax or language, and they can generate responses that sound like a human produced them. Though their inner workings are not fully interpretable, they exhibit behavior that seems intelligent and adaptable, and they are capable of continuously training and refining themselves.
A New Programming Paradigm
With the advances in computing power in recent years, and the notion that a transformer or other self-determining model can learn, interpret, and even generate new content, we have begun a new era in computing. Perhaps the most prominent example is OpenAI’s ChatGPT—short for Generative Pre-trained Transformer. The transformer architecture enables the model to analyze long text sequences and identify important patterns, allowing it to not only understand natural language but to generate it as well.
This exchange between a human and the machine gives the appearance of artificial general intelligence (AGI), but in fact it is really just a style of computer program that takes human text as an input – termed ‘prompting’. While these remarkable systems do tend to understand most kinds of phrasing and can produce a wide variety of useful outputs also in human readable language, prompting is just another syntax for programming.
Like any computer language, there are techniques to good prompting that will result in better results. The best prompts specify the context in which we would like our questions answered and the voice in which they should be answered. Prompts work best when they are broken down into incremental requests for response. All of these prompting techniques are really just ways for the human to modify their requests so they are better understood by the GPT.
Because ChatGPT and other Generative Transformers such as Anthropic’s Claude, Google’s Gemini and others work best with expert prompting and sometimes a more specific data context that may not be readily available from scraping the Web, a new class of applications has arisen called AI Assistants. These are AI agents trained on specialized, sometimes commercial data and tailored for particular domains. Assistants are sometimes freely available, such as tech support chatbots or email summarizers. In this case the AGI is meant to enhance the value of the underlying software or service. Some assistants represent specialized expertise or data however, and are a software product in their own right, such as legal research assistants or drug discovery systems.
Types of Artificial Intelligence and New Data Sources
While text-based AI models dominate the news, and have taken over every written task from school work to board room presentations, they are by no means the only strong AI systems available. Other types of artificial intelligence include different inputs, such as voice or motion, different generative outputs such as images or directions, and other use cases such as operating devices or demonstrating processes. Ultimately, AI is a technique for programming digital systems using a very broad definition of data as input and training or prompting as the programming language.
Voice Assistants like Amazon Alexa+ or Android Voice Assistant are trained on vast global voice datasets consisting of all ages and voice ranges of users, and all manner of phrasing and languages. (Why do you think these services are free? They wanted your voice data.) The increasingly fast models and near-universal access to the internet is taking us closer to a day where Douglas Adams’ Babelfish will be reality. In the same way as Google Maps ensures you are never lost, a near term reality is that you will be able to communicate with anyone around the world. Voice AI can also be used to generate entertainment content – audio books, songs, and other kinds of vocal performances. If you haven’t tried suno.com you need to.
Image recognition and image generation tools are a third example of non-text-language AI systems that create graphics, maps, or predictive visualizations. They are as universally useful as natural language systems, and find utility in financial, weather, and fraud detection systems, to name a few. Image systems are a fertile area of AI research since they require different types of 2D and 3D models to accurately train and generate their content. Transformer-based image-generation systems created persistent and sometimes hilarious hallucinations in the text content of images, highlighting the importance of appropriate training data and model type; text within images (labels, signs, etc.) was treated as pixels not text, so the statistical model will create more pixels that look like text but are often not real words and often could not be placed with precision in an image. Music and Image generation systems are powerful creativity enhancers.
These AI-powered image and voice generation systems are also a topic of increasing legal and ethical debate because their training data may not belong to the engineers or companies who developed them, and was often covered by copyright restrictions in a way that many (but not all) text web pages are not. For example, AI-generated art in a known artist’s style raises questions of infringement. Image training data certainly introduces an obvious content bias into these systems – almost all of the early images created by ChatGPT’s image generator had a very distinct style because the source material was a particular set of open-source clipart.
Another emerging use case is found in robotics AG: self-driving cars learn to recognize obstacles and memorize mapped routes based on thousands of miles of driving data and lidar records of the surrounding traffic. In this case the input data isn’t text, it is the record of human interaction with a steering wheel, gas, and brake pedals correlated with a record of what the street looked like at the moment a change was made in the car’s trajectory and speed. There are assembly line robots that are trained by humans programming the robot through physical manipulation or video. These robots are attempting to automate the sometimes physically demanding or repetitive tasks of transport, machine assembly, warehouse stocking, and even shipping. As with all new technologies, the applications are really only limited to our ability to define a need and cast the solution in terms of the system engineering – in this case the models and the data.
AI Limitations
It is tempting to subscribe to the point of view that AI systems are legitimate replacements for human tasks and intelligence. This is a critical mistake. They are statistical systems and use predictive analytics to determine what actions are mostly likely based on patterns of inputs. One of the most pressing challenges in the development and use of AI systems is bias — systematic and unfair discrimination embedded in AI models. Despite their apparent objectivity, AI systems are only as fair and accurate as the data they are trained on and the assumptions made during their design. Because training datasets often reflect historical, social, and institutional inequities, AI models can inadvertently learn and amplify those biases.
For example, an AI system used in hiring might favor male candidates if its training data comes from a company’s historical hiring patterns that underrepresented women. Facial recognition systems have performed poorly on individuals with darker skin tones when they’ve been trained predominantly on photos of lighter-skinned faces. Even language models can reflect and reproduce stereotypes present in the text they ingest—subtly reinforcing gender roles, racial hierarchies, or biased representations of cultures and professions. Bias doesn’t only arise from the data itself. It can also stem from how a model is trained, which features are emphasized or ignored, and how decisions are interpreted.
Similarly, AI systems can inherit knowledge limitations from their training data. Responses may simply be made up because the system is always programmed to answer, and answer with confidence, even if it lacks examples in its data that would allow it to predict an answer or generate content correctly. This results in sometimes humorous, sometimes insidious fabricated content known as hallucinations. This terminology makes light of one of the biggest problems with AI, particularly for younger or non-technical users. Hallucinations can easily be mistaken for fact by someone who is not knowledgeable about the task they are asking AI to perform.
Addressing bias and hallucinations in generative AI is an active area of AI research and policy. Makers of reusable generative models and the applications which consume them must be vigilant and proactive about rooting out bias and incorrect responses in their systems with audits, diversification and freshening of training data, and automated or human-in-the-loop benchmarking systems that alert when a system is expressing biased or made up responses.
Combating these limitations isn’t just a technical issue—it’s also ethical and social. It requires interdisciplinary collaboration, transparency from developers and organizations, and policies that enforce accountability. As AI systems increasingly make or influence decisions in finance, healthcare, education, and justice, ensuring they do not perpetuate or worsen existing inequities or misinformation is both a technological imperative and a moral obligation. Educating as many people as possible about how these systems work, demystifying them to the point of tools, and advertising the critical analysis of results is necessary for them to become a sensible part of society.
