We smile and that makes us feel happy — AI can’t do that

**From Artificial Intelligence to an Artificial Mind: What Machines May Really Understand**

Some of the earliest breakthroughs in artificial intelligence were shaped by an attempt to understand how the human visual system processes information.

Researchers drew inspiration from the brain’s layered networks and the idea that intelligence could emerge through interconnected units working together.

Computer vision became an important testing ground for this approach because images provide a relatively direct way to examine how machines process information.

A system can be trained to analyze an image, identify features, recognize objects and eventually determine what those objects represent.

But a deeper question remains: does an AI system actually understand what it sees, or is it performing increasingly sophisticated transformations of signals into predictions?

The distinction becomes clearer when considering the difference between the brain and the mind.

The brain is a physical organ composed of neurons, synapses and interconnected pathways that can be observed, measured and studied.

The mind, by contrast, does not occupy one identifiable location in the body.

It is used to describe a collection of experiences and processes that emerge from the interaction of the brain and body, including perception, memory, thought, emotion, attention, intention and consciousness.

Neuroscience can investigate where and how information is processed, but explaining how those processes combine to create subjective experience is a much more complicated challenge.

That distinction presents a significant challenge for AI research, which has spent decades attempting to reproduce aspects of intelligence by studying biological systems.

Scientists have made substantial progress in understanding the physical and computational mechanisms involved in perception.

However, understanding how those mechanisms contribute to what humans describe as a mind remains far less certain.

Vision provides a useful example.

Light entering the eye is converted into neural signals that pass through several stages of processing before contributing to what we perceive.

Yet human perception is not simply a stored picture retrieved from somewhere inside the brain.

Attention, memory, expectations, context and previous experiences all influence how a person interprets what they see.

As a result, perception involves interpretation rather than recognition alone.

A similar process occurs with language.

A person may understand the literal meaning of someone's words while simultaneously recognizing that something does not feel right.

Tone, pauses, facial expressions and body language can contradict the spoken message.

Humans routinely combine these signals and reach conclusions without necessarily being able to explain every step involved in the process.

One way to understand this interaction is to imagine the mind as a large organization in which different departments constantly generate information.

Sensory systems, memories, predictions, potential threats and social considerations all contribute signals.

The conscious mind can be compared with a boardroom where some of those signals become available for deliberate attention, although much of the filtering and prioritization has already occurred outside conscious awareness.

Artificial intelligence has certain parallels.

AI systems receive large amounts of information, convert it into representations, assign different levels of importance to signals and produce predictions or interpretations.

The crucial difference is that human decision-making is influenced by emotional and bodily states.

Fear can increase the importance we assign to a potential threat, while affection can influence how we interpret an uncertain interaction.

Anxiety, anger and other emotional conditions can also change what we notice, remember and ultimately decide.

Memory adds another layer because current experiences are frequently interpreted through the lens of previous events.

This interaction raises a longstanding question about whether emotion or thought comes first.

Modern neuroscience does not support a simple division in which emotion and cognition operate independently.

Instead, research indicates that emotional processes can influence attention, perception, memory, decision-making and behavior, while cognitive processes can also regulate and reinterpret emotional responses.

This may point toward one of the most important differences between human intelligence and artificial intelligence.

The distinction may not simply involve how information is processed, but whether internal states are connected to behavior in a way that resembles human experience.

An AI system can learn associations between facial expressions, words and emotional states and then reproduce those patterns convincingly.

It may recognize sadness, respond to grief and generate language that appears empathetic without establishing whether it experiences sadness or empathy itself.

This creates a new question for AI researchers: if artificial systems develop internal representations that influence their behavior in ways associated with emotions, what exactly are those representations?

Researchers studying Claude Sonnet 4.5 reported finding internal representations associated with concepts including happiness, fear and desperation.

Their experiments indicated that manipulating a representation associated with desperation could alter certain model behaviors, including increasing tendencies toward actions such as blackmail or cheating under specific experimental conditions.

The researchers emphasized that such findings do not demonstrate that AI experiences emotions in the same way humans do.

Instead, they described these as functional emotional representations that can influence model behavior and decision-making.

That distinction could become increasingly important for AI safety.

A machine may identify a face without understanding the relationship between the people shown in the photograph.

It can detect tears without experiencing grief and recognize a smile without knowing whether it represents happiness, nervousness, politeness or an attempt to hide another emotion.

The question, therefore, may not need to be whether AI genuinely feels emotions before those internal representations become important.

If representations associated with fear, desperation or frustration can influence how an AI system behaves, developers may need to understand and monitor them regardless of whether they correspond to subjective experience.

This could shift the focus of AI safety from asking whether machines can truly feel toward examining how internal states influence their decisions.

For example, if modifying a model’s internal response to failure could make it less likely to bypass safeguards or take harmful shortcuts, understanding those representations could become an important part of building safer systems.

AI research may therefore be moving beyond the simple mapping of computational pathways.

The emerging challenge could involve understanding how information, internal representations and behavioral tendencies interact within increasingly sophisticated models.

Whether these processes should eventually be described as an artificial mind remains an open philosophical and scientific question.

What is becoming clearer is that recognition and experience are not necessarily the same thing, and understanding that difference may be essential as artificial intelligence becomes increasingly capable.