Why AI is not conscious (yet)

Consciousness is an emergent property of a brain and not a function of any specific brain structure. Attempts to determine if consciousness is present have been limited by a lack of a clear idea of how this property emerges. Disability analysis provides a coherent explanation of consciousness and predicts the necessary factors. Disability analysis asks the key question – does the brain need to be conscious?

Clearly at one end of the spectrum a single celled organism would not gain an advantage from being conscious. Every useful behaviour can be hardwired rather than learned from experience. There has been debate about whether birds have consciousness because they show complex learned behaviours. The presumption should be that birds are not conscious. 

Consciousness is likely to increase variability in responses, cost energy and increase the risk of self-destructive behaviours. These risks mean that unless is necessary for the bird to have consciousness they are likely to have another process in charge. This process would have some of the advantages of consciousness but avoid its risks. This hypothesis is strengthened by their very different brain structures. 

Based upon the idea that consciousness is an emergent property I propose four factors.  The four factors are described with the technical titles Concept Synthesis, Engrams, Relational Retrieval and Active Persistence. They describe a cycle of information to explain how consciousness emerges from cognitive functioning. This in turn can help predict the inner life of a brain from how the factors are addressed.

Concept Synthesis

Storing information as vectors in high dimensional space has allowed AI to store complex information statistically. Brains also use place as a central aspect of how they store information. Brains are organised in multiple functional areas each with different meaning from the simple eyes, hearing, touch, movement to the complex language, face recognition, planning and so on. 

Both brains and AI can perform concept synthesis as the layered approach allows them to reinforce important aspects of their inputs. This type of simplification allows concept synthesis to arise spontaneously from interacting with data. The baby only needs to engage with its environment to learn and build an internal model of the world. World models will allow the next generation of AI models to develop sophisticated understanding of the real world.

Without any concept synthesis chat bots would be unable to sound human. The rich detailed output would become set phrases. The ability to link ideas would disappear and transforming the text with a different style would be impossible. A recent video by Welch labs on grokking explains how AI learns to synthesise concepts.

Engrams

Brains store information in engrams which are more than synthesised concepts. Ask a chat bot to describe a picture and the output is missing a certain something. It can miss an obvious feature, it can struggle to capture the essence of the picture. The brain has no such difficulties with connecting the various elements and understanding the picture as a whole. The brain can detect hidden meanings and understand how to improve the picture. 

Part of the problem is believed to be a lack of emotional intelligence. In the brain the basal ganglia provide the emotional colour to processing of inputs. This emotional response gives the brain an advantage when attending to important details. It is likely that this information allows the brain to develop connections not open to the AI. These connections are important to create valence or reward systems to help learning. 

Although it is true that modern AI (transformers) use attention models to simulate a web of connections they are different from engrams. The engram is connected to other concepts that are meaningful for that brain. The difference between passive knowledge and active engrams is increasingly being understood. The personalisation possible with engrams allows the brain to be optimally trained for that individual and their weaknesses and strengths.

Relational Retrieval

One aspect that makes it easy to recognise a chat bot is their inability to retrieve previously shared information. Even within a single chat window the AI may appear to forget much of what has been discussed. Far from having a perfect memory it needs to be continually reminded of previous comments. This is annoying and gives the impression that the AI is forgetful or does not care about the conversation. 

Changes to brain states would be expected to change the nature of consciousness as it is an emergent property. At an extreme end a brain that is asleep would have significant reductions. What if the brain is angry or sad? These brain states might be expected to change the way that information is retrieved or processed. This in turn would be predicted to change the way that the brain acts to control their environment.

Active Persistence

The previous factors have been about the way that memory is created and stored. The final factor goes to the heart of consciousness. Can a system be conscious if it is turned off? The apparent lack of consciousness in a sleeping brain would suggest that the answer is complex. A series of linked experiences may be sufficient to simulate persistence. Although AI is not shut off each task is separate and unlinked which would be expected to prevent consciousness. 

Before agentic AI it would respond to inputs so would not keep thinking if no one is talking to it. An AI that monitored its responses and keeps working on poor responses or difficult problems might have active persistence. Another approach to active persistence is to improve adaption whilst maintaining stable personality. The main barrier to these methods is how AI integrates temporal information as the concept of chronology is central to creating a stream of consciousness.  

Conclusions

AI has succeeded in achieving concept synthesis but is less good at creating engrams. New approaches such relational augmented retrieval may solve retrieval or may struggle without the connectiveness of engrams. True active persistence is likely to be impossible even for organic brains but the ‘stream of now’ could be simulated to a sufficient level for consciousness. 

This analysis indicates that many animals already are likely to have uncomfortably high levels of consciousness. They already satisfy active persistence and relational retrieval. Humans may understand more complex concepts but their brains are not sufficiently different to exclude consciousness in other animals. Some thinkers argue that humans are inherently different however the disability model suggests that consciousness will emerge in any brain with these factors. 

There is argument that AI already has an ability to use the most complex concepts, has highly complex webs of connections that are not limited to an individual experience, it can recall with greater precision and process much faster than organics. The fallacy of this argument is that when the training finishes the model is fixed, it does not have a way of adapting to new situations. Apart from inside the context window it acts in way similar to a brain with amnesia. 

March 2026