TL;DR: We can propose “consciousness tests” only if we imply the existence of a task that only a genuinely conscious system can solve. Therefore, if we invent an internally uncontradictive solution to that task, we’ll create consciousness.
I won’t waste my – and yours – time arguing for the importance of consciousness research. You already know it. Everything we know is known through the lens of conscious experience. Nor will I dwell on the familiar difficulties: the hard problem, the explanatory gap, the proliferation of theories, or the fact that comparisons between the most promising two has not shown a clear winner.
Instead, I’m here to share something that gives me hope. But first, a simple map of the labyrinth, with its dead ends and unexplored passages.
What do we do when we try to understand what consciousness is?
The first instinct is to take something conscious, take something unconscious, and compare them. This approach, which I call the comparative approach, has appeared in many forms since it was first proposed. Examples include comparisons between:
wakefulness and anesthesia, coma, or deep sleep;
perceived and unperceived stimuli, as in binocular rivalry or inattentional blindness;
animals that pass consciousness tests and those that do not;
intact and damaged brains, including lesions, surgical interventions, and neurodegeneration;
developmental stages before and after the presumed emergence of consciousness, both ontogenetically and phylogenetically.
The underlying logic is:
System A is conscious and possesses property α. System B is unconscious and lacks property α. Therefore consciousness depends on α.
This approach is incredibly productive in identifying the possible neural correlates of consciousness, but it is simultaneously an important limitation. Even if α is a perfect correlate, it does not explain why consciousness exists or what causal role it plays.
At first glance, the limitation may seem easy to overcome (neurobiologists are especially prone to simplifying it all, in my experience). Once α has been identified, one might simply turn α off and see whether consciousness disappears. But this does not increase the explanatory power of the result. “Consciousness depends on α” is very different from a statement such as “DNA is a double helix of nucleotides”. The former does not give us information on lower-level structures or processes that determine the properties of subjective experience, and does not explain the rules of that game.
The next tempting approach is reverse engineering. Suppose we have identified a correlate. We can then ask how, exactly, that correlate produces experience. How does the thalamocortical loop explain the redness of red? If this approach ever succeeds, it would constitute a genuine explanation.
Reverse engineering can be combined with phenomenological approaches. These begin by asking what properties consciousness possesses and what properties a system must have in order to generate subjective experience. Integrated Information Theory is the most prominent example. Ideally, approaches (2) and (3) should eventually meet in the middle.
Both approaches treat consciousness as something to be invented, be it in reverse or in the initial direction. We understand things most deeply when we can disassemble them and put them back together again. A clear answer to the question “How do we make a system conscious?” would go much further toward resolving the hard problem than any number of correlates.
There is, however, another approach—one that treats consciousness not as the destination but as a means of reaching it; not as something to be invented, but as an evolutionary invention that has emerged for an adaptive purpose.
Under this perspective, we stop searching for a universal neural correlate of consciousness in the same way that we do not search for a universal physical correlate of swimming that both algae and whales must possess. Instead, we search for a problem that consciousness evolved to solve across multiple evolutionary lineages, potentially through different, analogous mechanisms
Suppose there exists some task X that cannot be solved by an unconscious system. If we construct a system capable of solving it, then we may have reconstructed the essential mechanism of consciousness. The challenge is to define X rigorously. The challenge is to define X rigorously.
A weak version would simply require that consciousness helps solve X. A strong version requires something much more interesting: solving X is identical to being conscious. The analogy would be the relationship between cardiac muscle contraction and heartbeat. The process is not merely correlated with the function; it constitutes the function. Under this stronger criterion, a philosophical zombie capable of solving X becomes incoherent. The solution itself would be the experience.
One of examples for the ‘stronger’ task I like to think of: let’s assume that consciousness exists for preprocessing of the external stimuli in form of their classification into internally valuable categories and further calculations using the internal signs for these categories instead of the objectively registered signals.
This does not eliminate the possibility of multiple competing hypotheses. Different candidates for X will produce different designs. But unlike many existing theories, these proposals can be evaluated as complete mechanisms rather than collections of correlates. A successful proposal should satisfy at least one of two criteria: converge with known facts about brains; generate novel empirical predictions. If it does neither, it can be discarded.
On June 11, the Fourth Animal Consciousness Conference began in the Qingcheng Mountains. The previous two conferences produced an organized list of reliable tests of animal consciousness by these criteria:
Species generality — applicability across phylogenetically distant species;
Sensory/motor compatibility — independence from specific sensory or motor abilities;
Training dependence — minimal requirement for prolonged training;
Theory dependence — independence from specific theoretical frameworks;
Confidence — interpretability as evidence of consciousness;
Invasiveness — minimal disruption of physiological processes;
Neurobiological applicability — suitability for neurobiological investigation.
The website of Anyminds project emphasizes the importance of evaluating the existing tests and developing new ones“ for investigating underlying neurobiological mechanisms, with the goal of integrating behavioral and neural approaches within a comparative framework”. This, to me, sounds like working within comparative approach. But nevertheless I find it promising, because every test implies:
1) Existing of task impossible for zombies;
2) Evolutionary significant role of consciousness;
3) Difference between the role of consciousness and such features as intelligence, memory, attention, etc.
4) Possibility of documentation of performing conscious processes from third person
Every test is ultimately built around a proposition of the form: “Consciousness serves to...”
In other words, each test proposes a candidate X. We can then design a system that performs X and analyze the mechanism required to do so. That mechanism becomes a candidate mechanism of consciousness, provided that the claim “X is a causal role of consciousness” is correct.
This allows us to generate numerous internally coherent theories that bridge the explanatory gap, as follows:
X cannot be performed without consciousness.
Process N results in successful performance of X.
Therefore, process N is consciousness.
Each theory can then be tested against neurobiological reality in the search for the objectively correct one.
Once again, I want to emphasize the difference between the proposed engineering approach and the more widely-spread comparative approach.
The comparative approach begins with existing biological or artificial systems. It relies heavily on the assumption that we can identify which systems are conscious and which are not. It implies comparing conscious system one with other and searching for similarities or comparing conscious with unconscious and searching for differences. Ideally, it identifies a process or structure present in all conscious systems and absent from all unconscious ones – one whose addition produces consciousness and whose removal eliminates it. The resulting candidate can then be reverse-engineered and compared with predictions from phenomenological theories.
The engineering approach begins with tasks identified for existing systems. It relies heavily on assumption ‘this cannot be made without consciousness’. It seeks artificial or theoretical solutions to those tasks and generates multiple internally coherent candidate mechanisms of each of which is a possible mechanism of consciousness, although not necessarily mammalian-like consciousness. The result can then be compared with existing biological systems.
If consciousness has a unique causal role, identifying that role should be one of the highest-priority open problems at the intersection of neuroscience, philosophy, and AI research.The motivation is not only to create artificial consciousness. It is also to avoid creating it accidentally. If consciousness is the solution to some computational problem X, then sufficiently capable optimizers for X may become conscious whether or not we intended them to. I believe this is an invitation to a kind of research that should interest many potential readers.
TL;DR: We can propose “consciousness tests” only if we imply the existence of a task that only a genuinely conscious system can solve. Therefore, if we invent an internally uncontradictive solution to that task, we’ll create consciousness.
I won’t waste my – and yours – time arguing for the importance of consciousness research. You already know it. Everything we know is known through the lens of conscious experience. Nor will I dwell on the familiar difficulties: the hard problem, the explanatory gap, the proliferation of theories, or the fact that comparisons between the most promising two has not shown a clear winner.
Instead, I’m here to share something that gives me hope. But first, a simple map of the labyrinth, with its dead ends and unexplored passages.
What do we do when we try to understand what consciousness is?
The first instinct is to take something conscious, take something unconscious, and compare them. This approach, which I call the comparative approach, has appeared in many forms since it was first proposed. Examples include comparisons between:
The underlying logic is:
System A is conscious and possesses property α.
System B is unconscious and lacks property α.
Therefore consciousness depends on α.
This approach is incredibly productive in identifying the possible neural correlates of consciousness, but it is simultaneously an important limitation. Even if α is a perfect correlate, it does not explain why consciousness exists or what causal role it plays.
At first glance, the limitation may seem easy to overcome (neurobiologists are especially prone to simplifying it all, in my experience). Once α has been identified, one might simply turn α off and see whether consciousness disappears. But this does not increase the explanatory power of the result. “Consciousness depends on α” is very different from a statement such as “DNA is a double helix of nucleotides”. The former does not give us information on lower-level structures or processes that determine the properties of subjective experience, and does not explain the rules of that game.
The next tempting approach is reverse engineering. Suppose we have identified a correlate. We can then ask how, exactly, that correlate produces experience. How does the thalamocortical loop explain the redness of red? If this approach ever succeeds, it would constitute a genuine explanation.
Reverse engineering can be combined with phenomenological approaches. These begin by asking what properties consciousness possesses and what properties a system must have in order to generate subjective experience. Integrated Information Theory is the most prominent example. Ideally, approaches (2) and (3) should eventually meet in the middle.
Both approaches treat consciousness as something to be invented, be it in reverse or in the initial direction. We understand things most deeply when we can disassemble them and put them back together again. A clear answer to the question “How do we make a system conscious?” would go much further toward resolving the hard problem than any number of correlates.
There is, however, another approach—one that treats consciousness not as the destination but as a means of reaching it; not as something to be invented, but as an evolutionary invention that has emerged for an adaptive purpose.
Under this perspective, we stop searching for a universal neural correlate of consciousness in the same way that we do not search for a universal physical correlate of swimming that both algae and whales must possess. Instead, we search for a problem that consciousness evolved to solve across multiple evolutionary lineages, potentially through different, analogous mechanisms
Suppose there exists some task X that cannot be solved by an unconscious system. If we construct a system capable of solving it, then we may have reconstructed the essential mechanism of consciousness. The challenge is to define X rigorously. The challenge is to define X rigorously.
A weak version would simply require that consciousness helps solve X. A strong version requires something much more interesting: solving X is identical to being conscious. The analogy would be the relationship between cardiac muscle contraction and heartbeat. The process is not merely correlated with the function; it constitutes the function. Under this stronger criterion, a philosophical zombie capable of solving X becomes incoherent. The solution itself would be the experience.
One of examples for the ‘stronger’ task I like to think of: let’s assume that consciousness exists for preprocessing of the external stimuli in form of their classification into internally valuable categories and further calculations using the internal signs for these categories instead of the objectively registered signals.
This does not eliminate the possibility of multiple competing hypotheses. Different candidates for X will produce different designs. But unlike many existing theories, these proposals can be evaluated as complete mechanisms rather than collections of correlates. A successful proposal should satisfy at least one of two criteria: converge with known facts about brains; generate novel empirical predictions. If it does neither, it can be discarded.
On June 11, the Fourth Animal Consciousness Conference began in the Qingcheng Mountains. The previous two conferences produced an organized list of reliable tests of animal consciousness by these criteria:
The website of Anyminds project emphasizes the importance of evaluating the existing tests and developing new ones“ for investigating underlying neurobiological mechanisms, with the goal of integrating behavioral and neural approaches within a comparative framework”. This, to me, sounds like working within comparative approach. But nevertheless I find it promising, because every test implies:
1) Existing of task impossible for zombies;
2) Evolutionary significant role of consciousness;
3) Difference between the role of consciousness and such features as intelligence, memory, attention, etc.
4) Possibility of documentation of performing conscious processes from third person
Every test is ultimately built around a proposition of the form: “Consciousness serves to...”
In other words, each test proposes a candidate X. We can then design a system that performs X and analyze the mechanism required to do so. That mechanism becomes a candidate mechanism of consciousness, provided that the claim “X is a causal role of consciousness” is correct.
This allows us to generate numerous internally coherent theories that bridge the explanatory gap, as follows:
Each theory can then be tested against neurobiological reality in the search for the objectively correct one.
Once again, I want to emphasize the difference between the proposed engineering approach and the more widely-spread comparative approach.
The comparative approach begins with existing biological or artificial systems. It relies heavily on the assumption that we can identify which systems are conscious and which are not. It implies comparing conscious system one with other and searching for similarities or comparing conscious with unconscious and searching for differences. Ideally, it identifies a process or structure present in all conscious systems and absent from all unconscious ones – one whose addition produces consciousness and whose removal eliminates it. The resulting candidate can then be reverse-engineered and compared with predictions from phenomenological theories.
The engineering approach begins with tasks identified for existing systems. It relies heavily on assumption ‘this cannot be made without consciousness’. It seeks artificial or theoretical solutions to those tasks and generates multiple internally coherent candidate mechanisms of each of which is a possible mechanism of consciousness, although not necessarily mammalian-like consciousness. The result can then be compared with existing biological systems.
If consciousness has a unique causal role, identifying that role should be one of the highest-priority open problems at the intersection of neuroscience, philosophy, and AI research. The motivation is not only to create artificial consciousness. It is also to avoid creating it accidentally. If consciousness is the solution to some computational problem X, then sufficiently capable optimizers for X may become conscious whether or not we intended them to. I believe this is an invitation to a kind of research that should interest many potential readers.