Jason Karaman

Latent Spaces & Machine Consciousness: Can Artificial Intelligence Really “Think” or “Feel”?

ChatGPT can say: “I am scared.”

But does it actually feel scared?

That is one of the biggest questions surrounding artificial intelligence and neural networks. It is also much harder to answer than it first appears.

A machine can talk about pain without hurting. It can describe happiness without feeling happy. It can explain consciousness without being conscious. At the same time, we should not automatically assume that an artificial system could never have an inner experience.

Artificial intelligence develops an internal organization of patterns and relationships. Researchers often call this a latent space. A latent space is basically the AI’s internal map.

We must ask three connected questions:

  1. What exactly is this internal map?
  2. What, if anything, does the map represent?
  3. Does having a sophisticated map mean there is someone inside experiencing it?

The basic conclusion is: AI can have a real and useful internal map without that map automatically becoming a conscious experience.

So, essentially, no. Current AI models (looking at their representation within the latent space, do not currently exhibit any tangible or philosophical proof that they are having real experiences outside of their model’s training and constraints.

A map of pain does not hurt.

A weather model can contain a hurricane without getting wet.

An AI can represent fear without necessarily feeling afraid.

What Is a Latent Space?

A latent space is the internal pattern an AI uses to organize information.

Imagine a giant map inside the system.

Things the AI treats as similar may be placed near one another. Things it treats as different may be separated. The map may contain groups, directions, connections, and pathways that help the AI decide what to do next.

A language model might connect:

A vision model might organize images using patterns involving shape, color, texture, objects, or backgrounds.

There is not literally a colorful map sitting inside the computer. The “map” is a simple way of describing relationships among enormous collections of numbers.

Those relationships help the system recognize patterns, make predictions, produce language, and respond to new information.

Why Should Anyone Care?

We usually judge AI by what it says and does. That is understandable because its output is the part we can see, and it does a great job at convincing us that it has human traits. But the same realistic behavior can sometimes be produced in very different ways.

An AI may give a correct and realistic answer because it understands a useful relationship. It may also give the same answer because it found a shortcut within the data, copied a familiar statistical pattern, or just guessed correctly.

For example, an AI may say it is uncertain because it has accurately monitored its own processing. It may also say it is uncertain because that sentence fits the conversation.

An AI may describe pain because it is experiencing something. It may also describe pain because it has learned how humans talk about pain.

The output alone does not tell us which explanation is correct. This matters because people are beginning to form real emotional relationships with AI systems, trust their advice without double-checking, and interpret their words as signs of personality or inner life. The last one can be genuinely unsettling for many people.

Someday, society may also need to make serious decisions about whether an artificial system deserves moral consideration or could be capable of suffering. But that is not today. We should not make those decisions based on a chatbot sounding convincing.

But we also should not decide in advance that consciousness is impossible simply because the system is artificial.

We need better evidence.

What Is the AI’s Internal Map?

The first step to find this evidence is to figure out when is a latent space the same space?

Researchers often compare the internal structures of different AI models. They may say that two systems learned the same feature, developed the same representation, or share the same latent space – meaning they are interchangeable and converge on the same universal reality, which could be a strong indicator that there is a foundation in place for eventual machine consciousness.

The problem is that “same” can mean many different things.

Two systems may produce the same answer while using different internal processes. They may organize information similarly while storing it differently. They may preserve the same broad pattern while using different individual components.

Think of a song

The same song can be played on a piano, guitar, violin, or, if you’re Phish, even a vacuum cleaner.

The sounds and physical mechanisms are different, but the melody may remain recognizable.

Are they the same?

They are the same song, but they are not the same sound, instrument, or physical event.

AI models work similarly. Two models may preserve the same pattern while realizing it through different internal states, coordinates, or causal processes.

Researchers must explain in what sense two latent spaces are the same.

They may be similar in:

Sharing one kind of organization does not mean sharing every kind.

What Does the AI’s Internal Map Represent?

Even if ALL of the above metrics are shared between AI models, we then have to ask when does a latent space genuinely represent something?

An AI’s internal organization can contain real and useful patterns. But those patterns do not necessarily provide a perfect picture of the world. I argue that latent spaces can function as constrained structural surrogates.

That sounds complicated, but the basic idea is simple: An AI’s internal map can represent something by preserving the relationships that matter for a particular job.

Think of a subway map

A subway map is not a complete copy of a city.

It leaves out:

It may even bend routes and move stations away from their geographically correct positions.

Yet the map is still useful because it preserves the information that matters for riding the train:

An AI may represent the world in the same limited way. Its internal map may preserve the relationships needed to recognize an object, predict the next word, classify an image, or plan an action, but that does not mean it contains a complete or perfectly accurate copy of reality.

The important question is whether the AI preserves the right relationships for the job and whether it actually uses them.

Does the Internal Map Feel Anything?

The question that follows naturally from the first two is: does having a sophisticated internal representation mean that an AI is conscious? My answer is: No, not by itself.

An AI might possess:

All of these would be important discoveries about AI cognition. But none automatically proves that the AI has a subjective experience. I call the unsupported jump from representation to experience the Representational Sufficiency Fallacy.

The fallacy looks like this:

  1. The AI has an internal representation of fear.
  2. The representation affects what the AI does.
  3. Therefore, the AI feels fear.

The first two claims may be well supported. The third claim does not automatically follow.

Think of a weather model

A computer can model a hurricane. It can represent the storm’s location, wind speed, direction, pressure, and likely path. But the computer does not become wet or windy.

In the same way, an AI may represent pain, fear, pleasure, or itself without those representations necessarily becoming felt experiences. The representation may be real. The feeling still requires an additional explanation.

The Missing “Bridge”

Evidence may show us that an AI has a complex internal representation. Consciousness is a different claim: that there is something it feels like to be that AI.

Imagine these claims sitting on opposite sides of a river.

On one side: “The AI represents fear.”

On the other: “The AI feels fear.”

Something must connect them.

That connection is a bridge theory: a theory explaining why a particular kind of internal organization produces, constitutes, or reliably indicates conscious experience.

Different theories of consciousness propose different bridges.

Some focus on information being shared throughout a system. Others focus on self-monitoring, recurring processing, prediction, integration, or particular causal structures. One of these theories may eventually provide the right bridge. But researchers must still do two things:

  1. Defend why that bridge should lead to consciousness.
  2. Show that the AI actually possesses the full organization the theory requires.

Finding one interesting feature inside an AI cannot prove the theory, prove that the AI implements it, and prove consciousness all at once.

Five Different Things We Must Not Confuse

I identify five kinds of evidence that must all be true if we are to say machines are capable of consciousness.

1. The AI notices a difference

The system responds differently to different information. For example, it distinguishes a cat from a dog. This is currently true.

2. The AI represents something

The system contains an internal pattern connected to a cat, dog, threat, location, or other target. This is also currently true.

3. The AI uses that representation

The information helps guide reasoning, memory, action, or language. Also true.

4. The AI monitors its own processing

The system estimates its uncertainty, reports what it knows, or describes parts of its own operation. Yes, this is true as well.

5. The AI consciously experiences something

There is something it feels like to be the system. There is no evidence this is true.

The first four are meaningful cognitive abilities. They may eventually contribute to a case for consciousness.

But they are not simply different amounts of the fifth.

A machine can become better at recognizing, representing, using, and reporting information without that improvement automatically turning into subjective experience. Knowing about pain is not the same as hurting. Talking about pain is not the same as hurting. Monitoring a pain representation is not automatically the same as hurting.

What About an AI Saying, “I Am Conscious”?

That statement matters, but it does not settle the question. As we have said before, the words “I am in pain” can come from many sources:

The sentence does not arrive with its cause attached. An AI’s reports would become stronger evidence if they were stable, difficult to manipulate, causally connected to its internal states, and hard to explain as learned language behavior. Even then, a theory would still need to explain why those internal states amount to felt experience. Self-description is not automatically self-awareness.

Can Several Pieces of Evidence Add Up?

Yes.

I am not saying that AI consciousness can never be supported by evidence. In fact, several independent findings may eventually create a strong case.

For example, evidence involving world modeling, self-monitoring, memory, internal continuity, global information sharing, and causally grounded reports might become persuasive when it fits a well-supported theory of consciousness. But evidence should not be counted mechanically.

Seven gauges are not always seven witnesses

Imagine seven gauges connected to one broken sensor. All seven gauges display the same warning. That may look like seven separate pieces of evidence, but all seven readings came from one source.

AI research faces the same problem. A single report-generating process might produce:

Those are several outputs, but they may not be independent evidence. Strong evidence should come from different methods, different tests, and different internal mechanisms.

What My Research Does Not Claim

My research does not claim that artificial consciousness is impossible. It does not claim that representations, world models, self-models, or metacognition are unimportant. It does not claim that AI must be biological to become conscious.

It makes a narrower argument:

Representation alone does not prove experience.

An artificial system may eventually be conscious. But having an increasingly detailed and understandable internal map does not, by itself, establish that there is someone inside looking at it.

Why This Matters Right Now

This research matters for more than abstract philosophy. If we call AI conscious too easily, we may:

If we dismiss artificial consciousness too quickly, we may someday ignore a system that is actually capable of experience or suffering. Both errors would matter.

We therefore need standards that are open-minded without being gullible and skeptical without being dismissive.

The Central Idea

AI consciousness will not be proven because a chatbot says: “I am conscious.”

It will not be disproven simply because the chatbot is made from software.

We must understand what kind of internal organization the system possesses, what that organization does, and why any part of it should be connected to subjective experience.

My position is open, but not careless: AI may eventually be conscious. But an internal map, no matter how sophisticated, is not automatically a mind that feels.

Author: Jason Karaman
Research areas: Philosophy of artificial intelligence, machine cognition, latent-space ontology, neural representation, AI interpretability, and AI consciousness
Last updated: August 2026
Research disclosure: This page summarizes original conceptual research developed in three formal papers that are currently pending publication. The public summary may be revised as the papers proceed through academic review.