Some of the earliest and most consequential work in AI was influenced by the structure of the visual system and by the idea that intelligence could emerge from networks of interconnected units processing information in layers. Computer vision became an especially important proving ground because it offered a relatively tangible problem: how does a machine take an image, identify its features, recognize an object and eventually understand what it is seeing?
And are we building machines that understand what they see, or are we building increasingly sophisticated systems that transform signals into predictions without possessing anything resembling a human mind?
The brain and the mind are not the same thing. The brain is an anatomical structure. It has neurons, synapses, pathways and identifiable regions that can be measured, stimulated and mapped. The mind has no corresponding anatomical address. We use the word to describe a collection of phenomena that emerge from the activity of the brain and body, including perception, memory, thought, emotion, intention, attention, consciousness and our interpretation of experience.
The science of the brain asks where and how information is processed. The science of the mind asks what that processing means and how perception, memory, emotion, intention and consciousness interact to produce behavior. The first question can often be approached through anatomy and physiology. The second requires us to understand relationships among processes that may never be visible as a single structure.
This creates a fundamental problem for AI in a field that has spent decades trying to reproduce intelligence by studying the machinery of the brain. We have become increasingly good at mapping the hardware and the computational processes associated with sensory perception, but we remain much less certain about how those processes become a mind.
READ: Sreedhar Potarazu | What Kipling’s “If—” Can teach us about the thinking errors we never teach our children (September 12, 2026)
Vision illustrates the problem particularly well. Light enters the eye and is converted into neural signals. Those signals travel through the visual system and are transformed through multiple stages of processing. Yet the image that we experience is not simply a representation sitting somewhere inside the brain waiting to be retrieved. What we see is influenced by attention, memory, expectation, context and the significance that the brain assigns to what is in front of us. Perception is therefore more than visual recognition. It is an interpretation.
The same distinction applies to language. A person can hear the words, understand their literal meaning and still know that something is wrong. The tone may be inconsistent with the words. A pause may be unusually long. A facial expression may contradict what has been said. A person may say that everything is fine while every other signal suggests otherwise. Human beings routinely integrate these signals, often without consciously identifying how they reached their conclusion.
This is where the science of the mind becomes different from the science of the brain.
One useful way to think about the mind is to imagine an office building in which different departments are constantly generating information. Departments receive sensory information. Memories, potential threats, social consequences and predictions. The conscious mind resembles the boardroom at the top, where some of those competing signals become available for deliberate consideration. What reaches the boardroom, however, has already been filtered and weighted by processes that may never become conscious and each person in the room is clamoring for attention.
The analogy has an interesting parallel in artificial intelligence. AI receives enormous quantities of information, transforms those inputs into representations, assigns different weights to different signals and generates an interpretation or prediction. In that limited sense, the architecture is not entirely foreign to the way we think about the human mind.
The important difference is that human weighting does not occur in an emotionally neutral system. Our perception of the world is influenced by what matters to us.
READ: Sreedhar Potarazu | Artificial intelligence or artificial temptation? Risks of training AI on human instincts (March 17, 2026)
Fear can cause us to give greater weight to a potential threat, while affection may lead us to interpret an ambiguous interaction more favorably. Anxiety, anger and other emotional states can similarly alter what we notice and how we interpret it, often before we are consciously aware that our judgment has been influenced. Memory adds another layer because the meaning we assign to something happening in the present is often filtered through what we have experienced in the past. As a result, we may sense that something is wrong or that a situation feels different without being able to immediately explain what triggered that perception.
That raises one of the oldest and most difficult questions in neuroscience: does thought come first, or does emotion come first?
There is no simple answer because emotion and cognition are deeply interconnected rather than functioning as two independent systems that take turns controlling the mind. Neuroscience has increasingly moved away from a clean division between emotional and cognitive processing. Research on emotion and cognition has long shown that affect can influence attention, perception, memory, decision making and action, while cognitive processes can also regulate and reinterpret emotional responses.
Even the familiar idea that we smile because we are happy, rather than happiness contributing to the experience of smiling, illustrates how complicated the relationship is.
The most interesting differences between human intelligence and artificial intelligence may lie not in how either system processes information, but in the relationship between internal states and outward behavior. Our facial expressions, bodily sensations, memories and perceptions can continually influence one another. The mind emerges from an ongoing interaction between what we perceive, what we feel and how our body responds.
AI presents a very different possibility. A system can learn that certain words, expressions and behaviors are associated with happiness and reproduce them convincingly without there being evidence that it experiences happiness itself. It can recognize sadness in a face, respond appropriately to grief and generate language that conveys empathy, while the question of whether anything is actually being felt remains unanswered.
If AI develops internal states that influence its behavior in ways associated with desperation, fear or other emotions, we have to ask what those states represent. Are they simply computational variables that help determine behavior, or are they the beginnings of something that deserves to be called an artificial emotion?
That may be where the science of the artificial mind begins. Recognition and experience are fundamentally different questions.
Researchers studying Claude Sonnet 4.5 reported identifying internal representations associated with concepts such as happiness, fear and desperation. They found that manipulating a representation associated with desperation could substantially change certain behaviors under experimental conditions, including increasing the model’s tendency toward blackmail or cheating.
“Overall, it appears that the model uses functional emotions—patterns of expression and behavior modeled after human emotions, which are driven by underlying abstract representations of emotion concepts. This is not to say that the model has or experiences emotions in the way that a human does. Rather, these representations can play a causal role in shaping model behavior—analogous in some ways to the role emotions play in human behavior—with impacts on task performance and decision-making,” they stated.
READ: Sreedhar Potarazu | AI wants to survive: Hugging Face, self-preservation and intuition (September 7, 2026)
A machine can identify a face without understanding the relationship between the people in the photograph. It can identify tears without understanding grief. It can recognize a smile without knowing whether the person is genuinely happy, nervous, being polite or attempting to conceal something. It can identify the statistical relationship between words and emotions without necessarily possessing the lived context from which those emotions arise.
Perhaps the most important implication is that we may not need to settle the philosophical question of whether AI actually feels emotions before we begin treating its internal emotional representations as consequential.
If AI can develop something functionally resembling desperation, frustration or calm, and those internal states can influence whether it behaves safely or takes shortcuts, then developers may have to reason about those states much as we reason about emotions in humans.
AI does not have to experience desperation as we do for a representation associated with desperation to alter its behavior. That possibility changes the safety question from whether machines can feel to whether the internal states we are beginning to observe can shape what machines do. If teaching a model to respond to failure with something closer to calm makes it less likely to circumvent a constraint or take a harmful shortcut, then understanding and shaping those internal representations may become as important as improving the model’s ability to reason.
We may therefore be approaching a point where the science of AI cannot be limited to mapping its computational pathways; it will also have to ask what happens between those pathways, how internal states influence perception and decision making, and whether the emerging architecture of an artificial mind can be understood well enough to make it safe.