The Five Stages of Artificial Intelligence
By John R. Kennedy ·
Understanding where we are, where we're going, and why the difference matters
Nearly everyone I speak with these days has some experience using AI. Some use it to write emails, summarize documents, or answer questions. Others have become remarkably skilled at prompting AI to perform increasingly complicated tasks.
I sometimes refer to these people affectionately as prompt geeks. They've discovered that the better they communicate with AI, the more useful it becomes.
But I've noticed something interesting about these conversations. We spend an enormous amount of time discussing what artificial intelligence can do and relatively little time discussing what it is becoming.
I think that's an important distinction.
Imagine living in the early days of the automobile and believing its greatest achievement would be eliminating the need to feed horses. You wouldn't necessarily be wrong, but you'd be overlooking the transformation of transportation, commerce, cities, and eventually much of modern society.
I suspect we're experiencing something similar with artificial intelligence. We're fascinated by the immediate benefits, which are substantial, but many of us haven't developed a framework for understanding where the technology might lead.
One way to begin is by examining five broad stages of artificial intelligence. They're not universally agreed-upon scientific classifications, and they don't represent a guaranteed sequence of technological development. Instead, they offer a practical way to understand how the capabilities of AI are expanding and, perhaps more importantly, how our relationship with those capabilities is changing.
For decades, we've interacted with relatively specialized forms of artificial intelligence without giving them much thought. Systems that recognize patterns, recommend products, filter unwanted email, or identify suspicious financial transactions have quietly become part of everyday life.
We might describe these systems as examples of Reactive AI, the first stage in our framework. They respond to particular inputs and perform relatively narrow tasks. Some can be extraordinarily sophisticated, but they generally don't operate as independent assistants pursuing our broader objectives.
For most people, these systems weren't particularly interesting. They were simply features of the technology we already used.
That changed dramatically when conversational AI became widely accessible.
Suddenly, we could communicate with computers using ordinary language and receive remarkably sophisticated responses. We could ask questions, explore ideas, create images, write software, and produce documents without understanding the technical processes involved.
This is the second stage, Generative AI, and it's where much of the public's attention remains today.
What's fascinating about generative AI isn't simply its ability to produce content. It's how dramatically it changes the experience of interacting with technology.
For most of computing history, people have had to learn how to communicate with machines. We learned commands, menus, programming languages, and increasingly complicated interfaces.
Now machines are becoming capable of communicating with us in our own language.
That change makes sophisticated technology accessible to people who might never have considered themselves technically inclined. Someone with an idea can begin exploring it without first acquiring years of specialized training.
I think that's extraordinary.
But there's something about this relationship that we should recognize. When we use generative AI, we're generally still directing the interaction. We ask questions, evaluate answers, and decide what to do next. Even when the technology performs complicated work, we're usually responsible for connecting that work to our larger objectives.
Now imagine that relationship changing.
Instead of asking AI to help plan a vacation, imagine asking it to manage the entire trip. It researches destinations, compares transportation options, checks hotel availability, coordinates schedules, and potentially makes reservations on your behalf.
You aren't necessarily giving it a series of individual instructions. You're describing an outcome and allowing the system to determine how to accomplish it.
This brings us to the third stage, Agentic AI.
Agentic systems can pursue objectives through multiple steps, use software tools, retrieve information, and potentially take actions on behalf of the people or organizations that authorize them.
Some of these capabilities are already emerging in products available today, although their reliability, autonomy, and safeguards vary considerably.
To me, this represents one of the most important transitions in the history of computing.
We're beginning to move from telling computers what to do toward telling them what we want to accomplish.
At first glance, that might sound like a small distinction. After all, we've been automating tasks for decades. But traditional automation generally follows processes that people have defined in advance. An AI agent can potentially determine parts of the process for itself, adapting its approach as circumstances change.
Consider what that means in everyday life.
An AI assistant that helps you understand a medical bill is useful. One that contacts the provider, negotiates a correction, and arranges payment could be considerably more useful.
But what happens if it misunderstands the bill? What if it shares information you didn't intend to disclose? What if it agrees to something you wouldn't have authorized?
The same capabilities that make the technology valuable introduce new responsibilities.
And those responsibilities aren't limited to the people developing AI. They extend to everyone who decides what information these systems can access and what actions they're permitted to take.
This is where I believe our understanding of artificial intelligence needs to evolve beyond prompting.
Being good at asking AI questions is useful. Understanding what happens when AI begins acting on our behalf is something else entirely.
Interestingly, an AI system doesn't need to possess human-level intelligence to perform actions with significant consequences. A relatively specialized agent might have permission to move money, distribute confidential information, or make purchases while lacking the broader judgment we would expect from a person entrusted with those responsibilities.
That distinction becomes especially important when we begin discussing Artificial General Intelligence, commonly known as AGI.
For most of human history, intelligence has been something we've understood through our own experience. We recognize it in our ability to learn, reason, adapt, and apply knowledge to unfamiliar situations. We also recognize that intelligence isn't distributed equally across every subject. A brilliant surgeon may know very little about engineering, while an exceptional engineer may struggle with problems that seem obvious to an experienced physician.
What makes human intelligence remarkable isn't that any one of us knows everything. It's our ability to encounter something unfamiliar and begin figuring it out.
Artificial General Intelligence represents the possibility of developing a machine with comparable intellectual flexibility. Rather than excelling primarily within particular categories of tasks, such a system could potentially approach unfamiliar problems, acquire new knowledge, and apply what it learns across many different fields.
Researchers disagree about exactly what would qualify as AGI, and there is no universally accepted test for determining when we've achieved it. But the underlying idea is fascinating because it challenges something we've taken for granted throughout human history: that people occupy a unique position when it comes to broadly adaptable intelligence.
I sometimes wonder how our relationship with technology would change if we could no longer assume that the person using a computer was the more generally capable intelligence in the relationship.
Today, when I ask AI to help me solve a problem, I bring my own experience and judgment to the conversation. I may not understand every technical detail of its response, but I can question its assumptions, compare its recommendations with what I already know, and decide whether its conclusions make sense.
Now imagine working with a system whose understanding of a subject is substantially greater than your own. You might recognize that its conclusions are useful without fully understanding how it arrived at them. In some respects, we already experience this when consulting specialists. I don't need to understand everything my physician knows to benefit from medical advice, but I do need some basis for deciding whether that advice deserves my trust.
With people, we rely on more than intelligence when making those judgments. We consider experience, reputation, professional responsibilities, and the consequences someone might face for making a poor decision. We also recognize that even the most knowledgeable experts can disagree.
How would we develop a comparable relationship with an artificial intelligence that possesses extraordinary capabilities but doesn't share the experiences, motivations, or responsibilities of a human professional?
It's a question I find more interesting than simply predicting when AGI might arrive.
And it becomes even more intriguing when we consider Artificial Superintelligence, or ASI, the fifth stage in our framework.
Artificial superintelligence remains hypothetical. It describes a form of intelligence that would substantially exceed human intellectual capabilities across a broad range of important domains. We don't know whether such a system will ever exist, much less what its capabilities or limitations might be.
Nevertheless, the possibility invites us to reconsider what intelligence actually means.
Throughout history, we've benefited from people whose abilities exceeded those of their contemporaries. Extraordinary thinkers have transformed mathematics, medicine, physics, philosophy, and nearly every other field of human endeavor. Yet even the greatest minds operated within the limits of human experience. They lived among us, communicated through shared languages, and participated in societies shaped by human needs.
A superintelligent machine might not share those characteristics.
It could potentially recognize patterns we cannot perceive, develop explanations beyond our ability to independently verify, or propose solutions that challenge assumptions we've held for centuries.
That possibility is exciting. It also presents a peculiar dilemma.
If a system becomes intelligent enough to consistently outperform our own reasoning, how do we determine when to trust it and when to question it?
We wouldn't want to dismiss an extraordinary discovery simply because we couldn't immediately understand it. At the same time, accepting every conclusion without question would mean surrendering something essential about our role in making decisions.
I don't know exactly how we'll resolve that tension. I suspect it will require us to develop new ways of evaluating evidence, establishing accountability, and understanding the limitations of systems whose capabilities may exceed our own.
But something about this entire progression keeps bringing me back to the present.
We don't need to achieve AGI or superintelligence before confronting these questions.
Consider the AI agents we discussed earlier. Even with today's limitations, they can be given access to information and permission to perform activities that have real consequences. A system doesn't need to be smarter than a person to make a purchase that person didn't intend, disclose information that should have remained private, or initiate a process that becomes difficult to reverse.
In that sense, the arrival of increasingly autonomous AI may change our daily lives more immediately than the arrival of increasingly intelligent AI.
I find that distinction important because it suggests that the future of artificial intelligence isn't determined solely by technological breakthroughs. It's also shaped by the choices we make about how to use the capabilities already available to us.
When I think about the five stages together, I don't see a simple progression from primitive machines to superior ones. I see an expanding relationship between human intentions and technological capabilities.
At first, we built systems to perform specific tasks. Then we developed systems capable of creating things in response to our instructions. Now we're beginning to build systems that can pursue objectives on our behalf. Beyond that lie the possibilities of general intelligence and perhaps forms of intelligence that exceed our own.
Each development changes what becomes possible, but it also changes what we need to understand about ourselves.
For a long time, the limitations of computers helped define the boundaries of our relationship with them. There were simply many things machines couldn't do, regardless of how much authority we might have wanted to give them.
Those boundaries are becoming less restrictive.
And as they change, we'll increasingly find ourselves making decisions that previous generations never had to consider.
I remain optimistic about what artificial intelligence can accomplish. The potential to expand knowledge, improve healthcare, accelerate discovery, and make sophisticated capabilities accessible to ordinary people is extraordinary.
But I don't think understanding the future of AI requires us to predict exactly how intelligent machines will become.
It requires us to recognize that our relationship with them is changing.
Perhaps the most consequential stage of artificial intelligence won't be defined by the moment a machine becomes smarter than a human.
It may be defined by the moment we begin trusting machines to make decisions we once believed only people should make.
Continue reading: When AI Becomes Your Assistant, Who Is Really in Control?