Blums’ Conscious Turing Machine

Theoretical computer scientists Lenore Blum, Manuel Blum, and Avrim Blum account for consciousness from a computational perspective and offer their Conscious Turing Machine (CTM) as a formal model of consciousness. They base CTM on Theoretical Computer Science (TCS), which is a branch of mathematics that studies computation under limited resources such as time and space. (The three award-winning Blums—wife, husband, son—have their PhDs from MIT and were computer science faculty members at Carnegie-Mellon University.)

The Conscious Turing Machine (CTM) (Blum & Blum, 2025) is inspired in part by Alan Turing’s simple yet powerful model of computation, the Turing Machine (TM) (Turing, 1937), and Bernard Baars' Global Workspace (GW) “theater” model of consciousness (Baars, 1997). However, CTM is not a standard TM,^[1] nor is it a standard GW model… nor is CTM a model of the brain: it is too simple for that.

In the CTM framework, formal definitions of conscious attention and conscious awareness correspond to concepts in consciousness literature called access consciousness (Block, 1995) and phenomenal (subjective) consciousness, respectively. Although attention in CTM arises from global broadcasts, awareness also entails the co-evolution of CTM’s Model-of-the-World (MotW) along with its internally generated multimodal language, Brainish.

The Blums contend that a machine capable of interacting with its environment through input sensors and output actuators, constructing internal models of its world, and viewing these models as being the world, is the basis for its consciousness, both access and phenomenal. In their view, addressing the Hard Problem of consciousness (Chalmers, 1995b) reduces to demonstrating that subjective experience is computational in nature.

To proceed, the Blums describe their CTM model (its architecture, processors, competition) and both conscious attention and conscious awareness (subjective experience) in CTM. They also discuss how CTM integrates and aligns with key underlying principles of many contemporary theories of consciousness.

The CTM Model

The mathematics of TCS abstracts and formalizes informal notions and derives results from foundational axioms. The Blums' CTM model embodies these principles.

Architecture

The CTM is defined by a 7-tuple: STM, LTM, Up-Tree, Down-Tree, Links, Input, Output as follows:

• STM: A buffer and broadcasting station (the “conscious stage”) holding one chunk (formal definition below) at a time.
• LTM: A vast collection of independent powerful “unconscious” processors (the “audience”). All CTM processors are in LTM, none in STM. Processors produce chunks that compete to get into STM.
• Up-Tree: The conduit for the (well-defined) probabilistic competition to select the winning chunk. The Up-Tree is a perfect binary tree (height h and 2^h leaves) with a leaf (node) in each processor and root (node) in STM.
• Down-Tree: The conduit for the global broadcast of the winning chunk from STM to all processors. The Down-Tree is a bush (tree of height 1 and 2^h leaves) with leaf in each processor and root in STM.
• Links: Pathways that form between processors (via “Hebbian-like” rules) that can replace slow “conscious” communication by fast “unconscious” communication.
• Input/Output: Maps that connect LTM Input/Output processors to the environment through sensors and actuators.
• Time is measured in discrete clock ticks: t = 0, 1, 2, …, T where T is CTM’s finite lifespan.

Processors

At time t = 0, some processors (Input, Output, and Gauge processors) have specific functions built-in. Others are generic. Specializations evolve over time. All processors contribute to creating models of CTM’s inner and outer world. The portions of all processors that contribute to creating the MotW are collectively called the Model-of-the-World processor (MotWp); they unconsciously and continuously predict, correct, and learn (predictive dynamics).

While each processor has its own internal language, processors communicate with each other (within CTM) in Brainish, CTM’s self-generated multimodal language. In Brainish, the label for (a MotW sketch of) a rose might consist of a fusion of its red color, sweet smell, soft touch of its petals, and memory of past experiences with roses.

Competition for STM and Global Broadcast

As the Blums have it, at every tick, t, each CTM processor submits a chunk⟨p, t, g, w, aux> of information to its leaf node of the (binary) competition Up-Tree (think of a tennis or chess tournament) for STM. Here:

• p = address of the submitting processor,
• t = time of submission,
• g = succinct multimodal Brainish gist of information (could be a frame in a dream),
• w = signed weight (valence/importance) the processor p gives the gist at time t,
• aux = auxiliary information.

Chunks compete with their siblings. (Chunks are siblings if they are at nodes that are children of the same parent node.) Winning siblings move up the binary tree (losing siblings drop out). Winners compete at the next tick with their new siblings. As local wining chunks move up the Up-Tree, their first four parameters stay unchanged, but their auxiliary information can change, gathering more and more global information. In CTM’s probabilistic winner-take-all competition with aux starting out as |w|, the auxiliary information in the chunk that gets into STM will be the sum of all competing entry chunks’ |weight|s at time t. We denote this number by Intensity and call this information the “global context.”

Theorem: In CTM’s probabilistic winner-take-all competition, the probability that p’s chunk ⟨p, t, g, w, aux> wins the competition will be the ratio of its |w| to Intensity (the sum of all competing entry chunks’ |weight|s at time t).

Corollary: Winning is independent of where processors are placed on the leaves of the Up-Tree competition.

Clearly, this is an important feature for a machine or brain in which no movement of processors is possible. (It is a property that would be difficult if not impossible to achieve in fair tennis or chess tournaments.)

This well-defined, decentralized, winner-take-all competition replaces the need for the undefined “central executive” used in all other theater models of consciousness.^[2]

The winning chunk is determined in h clock ticks, where h = height of the Up-Tree. Upon winning, that chunk is globally broadcast in one clock tick via the Down-Tree to every processor and stored in every processor’s global dictionary.

With this broadcast, in addition to receiving the winner’s gist, its valenced weight, and its submitting processor’s address, all processors receive information enabling them to calculate the average |weight| of all chunks that entered the competition. This information will help all processors calibrate subsequent actions.

Processors are programmed to inspect (defined below) high-|weight| received chunks to unpack their meaning and to do their best to lower the magnitude of extreme negative weights.

Conscious Attention

Definition: CTM pays conscious attention to the winning chunk when that chunk is received by all LTM processors via the global broadcast.

Already, we see how CTM can exhibit certain phenomena associated with human consciousness. Examples include: inattentive blindness (CTM pays less attention to chunks with low |weight|s); change blindness (succinct gists contain an overview rather than detailed information); the Libet effect (a delay between the chunk’s entry into the competition and the reception of its broadcast). See also Blum & Blum, 2022.

Conscious Awareness

Conscious awareness in CTM is tied to the co-evolution of Brainish and the MotW.

The Co-evolution of Brainish and the MotW

Definition: A Brainish word <p, t> is a pointer to a broadcasted chunk <p, t, g, w>. A Brainish gist is either a NULL or a sequence of at most 20 Brainish words.

All processors maintain identical global dictionaries listing the received broadcasted chunks in the order of their arrival. When given a word, every processor can examine the chunk it points to, which describes or leads to a description of its meaning.^[3]

The newborn CTM has a foggy MotW that does not differentiate itself from the world.

Sketches in the MotW are broadcasted chunks. Primal sketches are chunks that have a NULL gist. The name of a sketch <p, t, g, w> is the word <p, t> that points to it. The name of the first primal sketch (in the life of the CTM) is a Brainish word that is denoted in English by <BLOB0>.

Needing “oxygen,” the newborn CTM’s “oxygen gauge” submits to the competition a sequence of increasingly negatively weighted chunks signaling “Do something!” Those broadcasts, having negative weights of increasing intensity, generate the Brainish primal word for pain, in English <PAIN0>.

The Brainish gist, <BLOB0><PAIN0>, interpreted as “BLOB0 is in PAIN0,” forms a labeled sketch in the MotW, with <BLOB0>, the name of the sketch, and <PAIN0>, the sketch’s label.

When CTM detects an actuator moving by “the power of thought,” the corresponding sketch is labeled <SELF>. Later, if CTM scrapes its knee, the sketch of its knee is labeled both <SELF> and <Skinned-knee PAIN>, the latter being a word whose meaning evolved from earlier pains.

Subjective Experiences and the Feeling of Consciousness in CTM

Here is where the Blums’ CTM model imparts subjective consciousness:

Definition.

A processor inspects a chunk <p, t, g, w, Intensity> whenever it examines each of its components and unpacks all words <p’, t’> appearing in gist g.

A processor unpacks a word <p’, t’> when it retrieves its meaning (defined to be the chunk <p’, t’, g’, w’, Intensity’> that <p’, t’> points to) and examines each component, p’, t’, g’, w’, Intensity’. The NULL word is considered unpacked.

CTM inspects a broadcasted chunk <p, t, g, w, aux> whenever all processors simultaneously (upon its reception and in the same tick) inspect that chunk.^[4]

Axiom A. When CTM inspects a broadcasted chunk <p, t, g, w, Intensity>, it evokes in CTM a unique subjective experience. The intensity and valence of that evoked experience is given by w (modulated by the “global context” Intensity).

Specifically:

If gist g is NULL, then when CTM inspects the broadcasted chunk <p, t, g, w, Intensity> it evokes in CTM a unique subjective experience, and that experience, whatever it might be, is defined to be a primal subjective experience.^[5] The intensity and valence of that evoked experience are given by w (modulated by the “global context” given by Intensity).

If gist g = <p1, t1>… <pk, tk>, then when CTM inspects the broadcasted chunk <p, t, g, w, Intensity> it evokes a unique subjective experience in CTM that is a fusion of the subjective experiences evoked by the meanings of the words in g. The intensity and valence of that evoked experience are given by w (modulated by the “global context”).

Formal Definition. Conscious awareness in CTM arises when CTM pays conscious attention to a chunk <p, t, g, w, Intensity> and inspects it.

1. Conscious attention is a consequence of the fact that ALL processors are receiving and attending to exactly the same broadcast.
2. Conscious awareness arises from the fact that CTM does not consciously see the world. CTM is consciously aware of only its self-constructed Model of the World (MotW). The MotW holds the whole of CTM’s conscious view of the world. A MotW processor creates and puts in the MotW labeled sketches of the objects in the world. The sketches and their colorings (labelings) describe the CTM’s current understanding and phenomenological feelings in/for/of the world. (This helps explain what makes dreams so realistic.)

Theorem: When CTM becomes consciously aware of chunk <p, t, g, w, Intensity>, a subjective experience is evoked. This subjective experience is a fusion of the subjective experiences evoked by the meanings of the words in g. The intensity and valence of the evoked experience are given by w (modulated by “global context”).

Proof: According to the Blums, CTM becoming consciously aware of chunk <p, t, g, w, Intensity> means that CTM pays conscious attention to it—i.e., all processors simultaneously receive that broadcasted chunk— and simultaneously inspect it. The conclusion of the Theorem then follows from Axiom A.

As CTM becomes consciously aware of more chunks, the MotWp may call attention to this awareness by submitting to the competition chunks alluding to this with high |w|.. The word <p, t> pointing to such a broadcasted winning chunk becomes the Brainish word for <CONSCIOUSLY_AWARE> or simply <CONSCIOUS>. The MotWp labels the sketch of CTM in the MotW with <p, t> aka <CONSCIOUS>.

When CTM inspects a broadcasted chunk with the labeled sketch <CTM><CONSCIOUS>, it experiences (by Axiom A) a fusion of the experiences evoked by (the meaning of the words in) this labeled sketch. This experience evokes CTM’s feeling of its own consciousness.

Mislabeled MotW sketches can evoke “human-like” pathologies in CTM: a sketch of a (perfectly) functional actuator labeled <NOT-SELF> may cause body integrity disorder; a sketch of an amputated actuator labeled <SELF>, phantom limb syndrome; a sketch of a friend labeled <SPY>, paranoia.

Alignment with other theories of consciousness

Although a simple machine model, CTM integrates and aligns, at a high level, with key features of major contemporary theories of consciousness. These theories include:

1. Global Workspace / Global Neuronal Workspace

2. Attention Schema Theory

3. Predictive Processing

4. Embodied Embedded Enactive Extended Mind

5. Integrated Information Theory

6. Evolutionary Theories of Consciousness

7. Reticulothalamic Activating System + Free Energy Principle

A discussion of CTM’s alignment with these and other theories of consciousness can be found in (Blum & Blum, 2025).

For example, CTM aligns with the architectural and global broadcasting features of Bernard Baars’ global workspace (GW) theory of consciousness (Baars, 1997) and, at a high level, with the global neuronal workspace (GNW) theory of consciousness of neuroscientists Stanislas Dehaene (Dehaene, 2014), Jean-Pierre Changeux (Dehaene & Changeux, 2005), and others. However, CTM differs from GW in significant ways: CTM’s competition for global broadcasting is formally defined; CTM constructs world models; and CTM purposely has no Central Executive. This latter makes it a good model for AGI (Blum & Blum, 2025).

Since CTM makes no claim that its functions map to specific areas of the human brain, CTM can align with the underlying principles of competing theories such as Global Neuronal Workspace and Integrated Information Theory that make such claims (see Cogitate, Ferrante, Gorska-Klimowska et al., 2025).

Philosophically, CTM aligns with much of Daniel Dennett’s functionalist perspective (Dennett, 1991). CTM’s competition for STM leading to global broadcast and conscious attention aligns with Dennett’s “multiple drafts” competing (Dennett, 1991) for “fame in the brain” (Dennett, 2001). Along with Dennett, the Blums do not see the explanatory gap (Levine, 1983) as insurmountable. On the contrary, they see CTM as helping to explain the feeling of “what it is like” (Nagel, 1974). Related, as with Michael Graziano’s Attention Schema Theory (Graziano, Guterstam, Bio, & Wilterson, 2020), CTM (with its MotW) appears to embody and substantiate “illusionist” notions of consciousness proposed by Dennett (Dennett, 2019) and Keith Frankish (Frankish, 2016).

On the other hand, some researchers, whose theories align with CTM to varying degrees, conclude that consciousness requires living organisms. The Blums disagree strongly with this. Prominent are Anil Seth (Seth, 2025), whose predictive processing and controlled hallucinations align with CTM’s predictive dynamics and MotW, and Nick Humphrey (Humphrey, 2023), whose evolutionary arguments align with CTM’s co-evolution of Brainish and the MotW.

In Wiring the Brain, Kevin Mitchell presented sixteen questions “that a theory of consciousness should be able to encompass,” declaring that “even if such a theory can’t currently answer all those questions, it should at least provide an overarching framework” (see Mitchell, 2023). Preliminary answers from the perspective of CTM can be found in (Blum & Blum, 2025).

Summary

The Blums’ conclusion: CTM is a simple formal computational model of consciousness with four key takeaways:

• CTM is clearly buildable.
• CTM exhibits a variety of phenomena associated with human consciousness (Blum & Blum, 2022).
• CTM aligns at a high level with many leading theories of consciousness.
• CTM addresses Kevin Mitchell’s 16 questions: What should a real theory of consciousness encompass?

The Blums’ implication: CTM makes plausible the claim that AI consciousness is inevitable (Blum & Blum, 2025).

Footnotes

[1]. For one, it is a highly parallel Random Access Machine (RAM).

[2]. Baddeley describes the central executive as "the most important but least understood component of working memory" (Baddeley, 2003). CTM predicts that a central executive is not necessary for consciousness. The Blums contend, if CTM theory is a good model of human consciousness then CTM predicts that the brain has no Central Executive.

Furthermore, they argue, CTM’s distributed and fair competition mechanism, enabling creative participation of unexpected talents, makes it a good model for AGI (Blum & Blum, 2025).

[3]. Brainish words are reminiscent of what, in the AI/ML/neuroscience literature, are called keys. Chunks (that words point to, i.e., their meanings when inspected) are reminiscent of what are called values (Gershman, Fiete, & Irie, 2025).

[4]. We assume that CTM can unpack a word in .1 milliseconds, 20 words in at most 2 milliseconds, and then it can inspect a chunk in 3 milliseconds.

[5]. The bottom-most turtle.