Intro to AI: What Is AI?
What Is AI?
Over time, researchers have approached AI from various perspectives. Some define intelligence based on how closely it mimics human behavior, while others focus on an abstract, formal concept of intelligence known as rationality—essentially, the ability to consistently do the “right thing.”
Even the understanding of rationality differs: for some, intelligence is about internal cognitive processes and reasoning, while others emphasize observable intelligent behavior.
It’s important to clarify that when we talk about rationality, we aren’t implying that humans are irrational in the sense of lacking mental clarity. Rather, we’re acknowledging that human decisions are not always mathematically flawless.
This blog provides an introduction to the different approaches to defining and understanding Artificial Intelligence (AI) and rationality. I try to explains how AI research has historically been divided into four main categories based on two dimensions:
- human vs. rational
- thought vs. behavior.
The Four Approaches to AI:
1. Acting humanly (Turing Test approach):
The Turing Test, introduced by Alan Turing in 1950, was a thought experiment aimed at avoiding the philosophical complexity of the question, “Can a machine think?” A computer passes this test if, after a series of written questions, a human interrogator is unable to distinguish whether the responses are from a person or a machine.
To succeed in a rigorously applied Turing Test, a computer would need to demonstrate several key capabilities:
- Natural language processing to effectively communicate in human language,
- Knowledge representation to store and manage information it acquires,
- Automated reasoning to answer questions and draw conclusions,
- Machine learning to adapt to new situations and recognize patterns.
Turing believed that physically simulating a human was not necessary to demonstrate intelligence. However, other researchers have expanded on this by proposing the “Total Turing Test,” which requires interaction with objects and people in the real world. To pass this more comprehensive test, a robot would also need:
- Computer vision and speech recognition to perceive the environment,
- Robotics to manipulate objects and navigate its surroundings.
2. Thinking humanly (Cognitive modeling approach):
To claim that a program thinks like a human, we first need to understand how humans think. This understanding can be gained in three primary ways:
- Introspection—observing and reflecting on our own thoughts as they occur,
- Psychological experiments—studying human behavior in action through controlled observations,
- Brain imaging—examining the brain’s activity while a person is engaged in various tasks.
Once we develop a sufficiently detailed theory of how the mind works, we can translate that theory into a computer program. If the program’s input-output behavior mirrors human behavior, it provides evidence that some of the mechanisms within the program may be similar to those that operate in the human brain.
This approach seeks to understand how humans think and replicate that process in AI systems. It relies on methods like introspection, psychological experiments, and brain imaging. If a program can emulate human thought processes, this suggests it might operate similarly to humans.
3. Thinking rationally (Laws of thought approach):
The Greek philosopher Aristotle was among the first to attempt to formalize “right thinking,” or the process of irrefutable reasoning. His syllogism established patterns for logical arguments that would always lead to correct conclusions if the premises were accurate.
A classic example begins with the statements “Socrates is a man” and “all men are mortal,” concluding that “Socrates is mortal.” (Although this example is more likely attributed to Sextus Empiricus than Aristotle.) These laws of thought were intended to govern how the mind operates, laying the foundation for the field of logic.
However, traditional logic relies on certain, definitive knowledge about the world—a condition rarely met in reality. We don’t possess absolute rules for complex fields like politics or warfare the same way we understand the rules of chess or arithmetic. Probability theory helps fill this gap by enabling rigorous reasoning with uncertain information. It allows for constructing models of rational thought, which move from raw sensory input to an understanding of how the world works and predictions about the future.
Yet, probability theory alone does not produce intelligent behavior. For that, we need a theory of rational action because rational thought, on its own, is insufficient for generating intelligent behavior.
4. Acting rationally (Rational agent approach):
An agent is simply something that acts. While all computer programs perform actions, computer agents are expected to go beyond that—they must operate autonomously, perceive and respond to their environment, function over extended periods, adapt to changes, and pursue set goals.
A rational agent is one that acts to achieve the best possible outcome or, in cases of uncertainty, the best expected outcome.
In the “laws of thought” approach to AI, the focus was on making correct inferences. While making correct inferences can be part of rational behavior—since deducing the best action and then taking it is one way to act rationally—there are also instances where rational action doesn’t involve inference. For example, reflexively pulling your hand away from a hot stove is often more effective than waiting to carefully analyze the situation before acting.
The rational-agent approach to AI offers two key advantages over other approaches. First, it is more general than the “laws of thought” approach, as correct inference is just one of several possible ways to achieve rational behavior. Second, it lends itself more easily to scientific progress, since rationality can be mathematically defined and applied across a broad range of contexts. From this well-defined standard, we can often work backward to design agents that are guaranteed to behave rationally, a task that is much harder if the goal is to simply mimic human behavior or thinking.
In AI, the focus has been on building agents that act in ways we define as “doing the right thing.” What constitutes the “right thing” is determined by the objectives set for the agent. This approach is widespread, forming a kind of standard model not only in AI but also in fields like:
- Control theory, where controllers minimize a cost function,
- Operations research, where policies maximize rewards,
- Statistics, where decision rules minimize loss functions,
- Economics, where decision makers maximize utility or some measure of social welfare.
In this approach, an AI system is expected to operate autonomously, adapt to changes, and make decisions to achieve the best possible outcome, even under uncertainty. This approach is broader and more scientifically rigorous than the laws of thought approach. It is closely related to fields like control theory, operations research, and economics.
Key Points:
- AI can either seek to emulate human intelligence or aim for an abstract, rational form of intelligence.
- The Turing Test focuses on imitating human behavior, while the rational agent approach is about achieving the best outcome, regardless of human-like behavior.
- The rational agent model is pervasive across various disciplines like control theory, economics, and statistics, where decision-making aims to optimize outcomes.
Summary
In summary, AI has developed across different methodologies depending on whether the goal is to replicate human behavior or achieve optimal decision-making through rational actions.