The now-famous essay “What Is It Like to Be a Bat”, by the American philosopher Thomas Nagel, argues that consciousness is inherently subjective and cannot be fully reduced to an objective third-person description (Nagel, 1974). In other words, Nagel argues that even if we can fully explain the mechanisms underlying consciousness, we cannot fully understand how the experience in question is lived by an organism that is different from ours. But if we apparently cannot immerse ourselves in the subjective experience of other beings, is there any way to confidently assess if a system is self-aware? And if bats are hard to figure out, what about non-living intelligent systems? Is there any way to objectively assess the presence of consciousness in intelligent systems that were not built from biology?
Bats vs. AI systems
Thomas Nagel likely chose the bat as an example to build his arguments about subjective experience because these creatures, while they are not so different from us biologically (i.e., they’re mammals), have a radically different perception system. The system in question is called echolocation and works by making sounds and then listening to the echoes that bounce back from nearby objects. Creatures such as bats and dolphins use echolocation to locate and identify objects.
Unless you are a vampire who also turns into a bat from time to time, you likely have a hard time imagining what echolocation feels like. But now try to imagine what it feels like to be software connected to a diverse system of sensory devices. More specifically, imagine being connected at the same time to street video cameras, wearable devices, satellites, microphones, motion sensors, temperature gauges, and streams of text from millions of users. Could a self-aware system actually experience many or all of these inputs at the same time? And what does it feel like to, for example, remotely control a space satellite?
The subjective experience of some AI systems could not only be more exotic relative to our own than that of bats, but we may also have a harder time agreeing on whether those systems are conscious or not. Bats are, after all, mammals, just like we are. If we have subjective experience (I can be sure of mine but not of yours), perhaps other creatures with whom we share part of our evolutionary history also have it. But what about a system made of silicon chips that are only partly inspired by the human brain architecture?
Some may be tempted to believe that the measurement of AI consciousness will always be indirect and uncertain. For example, we could measure memory, self-monitoring, etc., which can be understood as proxies for subjective experience. This may not be the case, however, if we can upload our minds into a digital system. We may be able to do this on a digital brain, but we could also try something on an architecture that resembles that of an AI system, where we could spend a few hours or so, just to get a feel of the AI’s experiences. Cool as it might be, mind uploading is not currently an option, so I would consider looking for a more feasible near-term approach to determine whether an AI is self-aware.
Strategies for Measuring Consciousness in Humans and AI
As discussed in a previous post on objective consciousness measurement, since we still cannot directly perceive consciousness itself, one way to measure it is to look for reliable footprints it leaves behind in two places at once: 1) self-reported experiences and 2) brain activity markers. In humans, for example, we can look at the smallest brain patterns that appear when a specific experience is present and disappear when it is not, while trying to separate those patterns from unrelated processes like raw sensory input, attention, memory, or the act of pressing a button to answer.
A similar approach could be used with AI, except that in this case, we could perhaps look at repeatable internal computational patterns that are present when the system reports a supposedly subjective experience and are absent, weaker, or different when it does not. At the same time, we could try to separate those patterns from processes assumed to be unrelated, whatever those might be (e.g., prompt parsing).
We don’t know if this approach would provide reliable physical signatures of consciousness in humans, let alone in AI. In humans, one problem is separating patterns of activation, as you can also pick up things that come before conscious experience, such as attention or stimulus strength, or things that come after it, such as decision-making or the act of reporting. In AI, we don’t know if, when a model says “I experienced X”, there is actually any subjective experience or just a simulation of one.
Perhaps one way to improve a test based on this approach is to instruct the AI to both simulate having a subjective experience and report one they are supposedly having. Then, we would check if there are any differences in pattern activation between the simulating and the supposedly non-simulating conditions.
Even in this example, we would still have reasons to worry that the differences in pattern activations are part of the simulation that the AI is doing, intentionally or not. For example, the AI system may be able to create two different kinds of simulation scenarios, one for the simulation condition and one for the supposedly non-simulation one. Any differences between two distinct simulation scenarios may manifest in distinct pattern activations.
Are there any better approaches that would inform the development of an AI consciousness test? Let’s take a look at two papers that might give us some ideas of what an AI consciousness test could look like.
Paper 1 Key Idea: Evaluate AI Consciousness Just as You Would Measure Human Consciousness
Butlin et al. (2023) argue that to determine whether an AI is conscious, we should look at whether the target system implements computational functions that are thought to be associated with consciousness in humans. The work assumes consciousness might come in degrees, have blurry edges, or vary across dimensions. It also assumes that consciousness depends on the right kinds of computations and functional organization and is substance-independent (i.e., it does not necessarily require a biological substrate).
The main point of this paper is that we should evaluate AI consciousness using scientific theories about how consciousness works, rather than relying on behavior-based tests. For this reason, the paper surveys several major scientific theories of consciousness and extracts what it calls “indicator properties”, that is, recurring functional themes that multiple theories identify as necessary or important. Theories and ideas that are of relevance according to this paper:
1. Recurrent Processing Theory: Consciousness may depend on information being processed in loops, not just passed forward once. In simple terms, the brain or system may need to send information back and re-check it, rather than just reacting in a one-way chain.
2. Global Workspace Theory: For consciousness to be possible, you need a system in which different specialized parts of a system share information through a limited central workspace. In other words, something becomes conscious when it is brought into a kind of mental spotlight and made available to many parts of the system.
3. Higher-order theories: Assume that consciousness may involve a system noticing or monitoring its own mental states. In other words, the system is not just experiencing something, but also has some awareness of what it is experiencing.
4. Attention Schema Theory: A conscious system is one that builds a model of its own attention. That means it can keep track of what it is focusing on and perhaps control that focus.
5. Predictive Processing: Rather than passively receiving information, the brain is constantly making guesses about what’s coming next and updating itself when reality doesn’t match. Consciousness, in this view, is an ongoing loop of prediction and correction.
6. Agency and Embodiment: In some views, consciousness is linked to being an active agent with goals, not just a passive receiver of information. A conscious system may matterfully interact with the world, act on it, and learn from what happens next.
According to the same paper, the resulting list of indicators is meant to be provisional, and possessing more of these indicators should make us more confident that a system is conscious, while lacking them should do the opposite.
The specific indicators include algorithmic recurrence in input modules, perceptual organization into coherent integrated representations, multiple specialized subsystems operating in parallel, a limited-capacity workspace that creates a processing limit, global broadcast of workspace contents across modules, state-dependent attention that allows sequential querying of modules, top-down or generative perception, metacognitive monitoring that separates reliable from noisy representations, agency tied to belief formation and action selection, sparse and smooth coding that generates a “quality space,” a predictive model of attention itself, predictive coding, flexible goal-directed learning, and embodiment understood as modeling how actions change inputs.
Note that the paper excludes Integrated Information Theory (IIT) from its main framework because it is not compatible with the computational functionalist assumptions guiding the report. More specifically, IIT treats consciousness as an intrinsic causal property of a system’s physical organization, not something preserved just by implementing the same functions or computations. Also note that the paper reviews several contemporary AI systems, including LLMs, and provides arguments for why these systems are likely not conscious.
Paper 2 Key Idea: The Brain’s Capacity for Consciousness Is Dependent on Its Ability to Generate Both Integrated and Differentiated Activity
Casali et al. (2013) set out to solve one of the hardest problems in clinical neuroscience: how to assess consciousness objectively in someone who cannot reliably speak, move, or respond to commands. They proposed a new measure, the perturbational complexity index (PCI), as a theoretically grounded way to estimate the brain’s capacity for consciousness without depending on sensory input, motor output, or verbal report.
The conceptual basis of the paper comes from the idea that consciousness depends on the brain’s ability to generate activity that is both integrated and differentiated. Integrated means that many parts of the brain can interact and form a unified whole rather than behaving like disconnected islands. Differentiated means that these interactions must also be complex and varied rather than repetitive, rigid, or stereotyped.
The authors argue that conscious experience always has both features. Every experience is unified, but every experience is also specific: seeing a face is not the same as hearing a tone or feeling pain. On this logic, a conscious brain should be able to produce complex patterns of activity that are both widespread and information-rich.
When consciousness is lost, such as in deep sleep, anesthesia, or severe disorders of consciousness, the brain may still respond, but those responses tend to become either too local and fragmented or too global and uniform. In other words, either integration fails, differentiation fails, or both fail. PCI is meant to capture such situations.
You can measure PCI by stimulating it with transcranial magnetic stimulation (TMS) and then recording the brain’s electrical response with something called high-density electroencephalogram (EEG). A conscious brain, according to the authors, should not just react locally or in a bland repetitive way. It should generate a pattern that spreads through multiple interconnected cortical regions and unfolds in a differentiated way. By contrast, an unconscious brain might show only a small local response, suggesting reduced integration, or it might produce a widespread but highly stereotyped slow-wave type response, suggesting reduced differentiation.
Operationally, PCI is calculated in several steps. First, the researchers deliver a TMS pulse to a cortical area and record the brain’s response over roughly the first 300 milliseconds using EEG. Second, they estimate the cortical sources responsible for the observed scalp signals. Third, they apply a statistical procedure to determine which cortical sources are significantly activated at which time points, producing a binary spatiotemporal matrix, called SS(x,t), that marks where and when significant activity occurred. Finally, they compress this binary matrix using a Lempel-Ziv complexity[1] measure and normalize it by source entropy[2].
The intuition is simple, even if the math is technical: a pattern that is redundant and predictable can be compressed easily and therefore has lower complexity, while a pattern that is varied and nonredundant resists compression and therefore has higher complexity. PCI is thus the normalized algorithmic complexity of the brain’s evoked response to a direct perturbation. The index is low when the response is either too limited or too stereotyped, and high when the response is both broadly distributed and richly differentiated.
The first major empirical result is that PCI distinguished conscious from unconscious states in healthy participants. Across 32 healthy subjects and 152 sessions, wakeful participants showed PCI values in a relatively tight range, roughly 0.44 to 0.67, with a mean around 0.55. By contrast, when subjects lost consciousness during non-rapid eye movement sleep or under anesthesia, PCI dropped to a lower non-overlapping range of about 0.12 to 0.31.
The paper reports that PCI was not significantly affected by stimulation site, stimulation intensity, hemisphere, or whether participants had their eyes open or closed, as long as the stimulation successfully triggered a significant cortical response.
A second major result is that PCI behaved consistently across very different routes to unconsciousness. The authors examined NREM sleep, midazolam sedation, propofol anesthesia, and xenon anesthesia. These conditions differ in mechanism, yet PCI dropped in all of them to similarly low values. Midazolam produced values around 0.23 to 0.31, propofol around 0.13 to 0.30, xenon around 0.12 to 0.31, and NREM sleep around 0.18 to 0.28.
The paper also shows that PCI is sensitive to intermediate or graded states rather than only full-on versus full-off consciousness. In six participants given propofol, the authors recorded PCI during wakefulness, intermediate sedation, and deep sedation. During intermediate sedation, when the subjects were drowsy and only partially responsive, PCI took values between about 0.34 and 0.42. These values fell between the fully conscious range and the clearly unconscious range, suggesting that PCI can track gradual changes in conscious level. A similar pattern appeared in sleep. In one participant, PCI measured during the transition from wakefulness into stage 1 sleep was around 0.39, again intermediate. During rapid eye movement sleep, from which the subject reported dream experience upon awakening, PCI rose to about 0.46, placing it back inside the conscious distribution. This is quite interesting, given REM sleep involves disconnection from the external environment but not necessarily the absence of experience.
The patients in minimally conscious state and those who had emerged from minimally conscious state showed intermediate but elevated PCI values. In minimally conscious patients, PCI ranged from about 0.32 to 0.49, with a mean around 0.39. In patients who had emerged from minimally conscious state and recovered functional communication despite severe impairments, PCI ranged from about 0.37 to 0.52, with a mean around 0.43. Importantly, in all cases these values sat above the highest PCI value observed during unequivocal loss of consciousness in healthy subjects, which was 0.31. That gives the index a potentially useful threshold-like property.
As the authors themselves note, while PCI is not an absolute diagnostic boundary, the data suggest that PCI values above the “unconscious ceiling” may indicate at least some preserved capacity for consciousness. This is especially useful because minimally conscious patients often fluctuate and may be hard to classify with confidence using bedside observation alone.
It’s worth mentioning that PCI fits between Integrated Information Theory (IIT) (Albantakis et al., 2022) and Global Neuronal Workspace Theory (GNWT) (Mashour, Roelfsema, Changeux & Dehaene, 2020) because it measures whether a perturbation produces a response that is both integrated across the system and differentiated across space and time; that language is very close to IIT, which says consciousness depends on information being both unified and highly specific/differentiated. At the same time, PCI also overlaps with GNWT, because a conscious system in global-workspace terms should show activity that can spread across distributed networks rather than stay local and stereotyped.
It is of interest to know that Casali’s findings have been replicated. Sarasso et al. (2015) found low-complexity responses under propofol and xenon but a wakefulness-like complex spatiotemporal response under ketamine despite behavioral unresponsiveness, supporting the idea that PCI tracks the brain’s capacity for consciousness better than simple responsiveness. Casarotto et al. (2016) reported that an empirically derived PCI cutoff separated conscious from unconscious benchmark conditions with 100% sensitivity and specificity and detected minimally conscious state with 94.7% sensitivity.
An additional independent sample by Sinitsyn et al. (2020) replicated PCI’s performance in discriminating unresponsive wakefulness syndrome from minimally conscious state, with most UWS patients showing slow, stereotyped, low-complexity responses and most MCS patients falling in the consciousness range.
More recently, Comolatti et al. (2019) proposed some changes to the original PCI. The proposed method, called the Perturbational Complexity Index based on State Transitions (PCIST), is meant to preserve the conceptual strengths of PCI while making it easier and faster to compute. The paper reported that the new approach matched the original one in its ability to distinguish conscious from unconscious states, while being much faster to calculate and easier to apply. They also showed that the newer version could be used with intracranial stimulation and stereoelectroencephalography recordings, where it detected a clear drop in brain-response complexity during sleep.
Towards an AI Consciousness Test
Here is one takeaway from each of the two works discussed above:
– Butlin et al. argue that to assess whether an AI system is conscious, we need to look at whether it implements the kinds of computational organization that major scientific theories associate with consciousness in humans.
– Casali et al. propose a way to assess if a system is conscious at a particular time by perturbing it and measuring whether that perturbation produces a response that is both integrated across the system and differentiated over time and space, since in the human brain, conscious states produce richer, less stereotyped responses than unconscious ones.
For an AI consciousness test, applying both ideas in a single assessment tool could mean combining a structural checklist of consciousness-relevant indicators with an intervention-based assessment: deliberately perturb internal representations, cached context states, memory modules, control signals, or recurrent processes where these exist, and then measure whether the effects remain narrow and stereotyped or instead propagate in a flexible, globally available, and information-rich way across the system’s computational pathways. The results can be treated as graded rather than all-or-none, because both Butlin’s framework and Casali’s PCI are explicitly compatible with degrees or levels of consciousness-related capacity.
Whether an assessment based on this combination would be of help or not is to be determined. For instance, we cannot be sure that the kinds of computational organization that major scientific theories associate with consciousness in humans are close to reality. Even if they are, it does not follow that the same markers would lead to consciousness in a system that’s architecturally very different from us.
Equally important, while Casali et al. provide some interesting empirical findings following the application of their PCI, we’re far from universally agreeing that integration across the system and differentiation over time and space are universal markers of consciousness.
That being said, it would be interesting to see if an AI system that scores high in consciousness at an assessment, let’s say, inspired by Casali et al.’s PCI and Butlin et al.’s ideas, produces different behavioral outputs relative to a system that scores low in consciousness at the same assessment.
If the observed behavioral differences were also indicative of consciousness, we could at least suspect we’re on the right track. In a worst-case scenario (or best-case, depending on how you see it), we might have achieved something that is functionally equivalent to consciousness, if not subjectively so.
Let’s Imagine an AI Consciousness Experiment with Bats
A hypothetical experiment would compare three kinds of bat-shaped minds under the same echo-filled conditions: real bats, robot bats powered by an intelligent architecture not thought to support consciousness, and robot bats powered by an architecture that, according to current theories, might. All three would have to solve the same kinds of problems through the same small, quirky, bat body: navigating by echolocation, foraging, avoiding obstacles, and making sense of a world in which nearly everything important arrives as reflected ultrasound. The point would be to see whether different ways of organizing intelligence produce different kinds of unified, flexible, and adaptive behavior under the same sensory and bodily constraints.
First, each system would be evaluated using a structural checklist inspired by Butlin et al., asking whether it has the kinds of organizational features that major theories associate with consciousness. These might include things such as recurrent processing, integrated memory access, globally available information, flexible routing, and self-monitoring capacities.
Second, inspired by Casali’s Perturbational Complexity Index, researchers would introduce controlled perturbations into neural activity in real bats and into internal representations, memory states, routing signals, or recurrent processes in the robot bats. The aim would be to see whether the disturbance remains narrow and stereotyped, as if one part of the system briefly malfunctions and the rest carries on in blissful ignorance, or whether it spreads through the system in a more flexible, integrated, and information-rich way across time and functionally distinct subsystems. From this, each biological and artificial bat could be assigned a graded consciousness-related score rather than a simple yes-or-no label.
The final step would be behavioral. We would assess whether the systems that score higher also show more unified self-monitoring, more adaptive responses to ambiguity, more coherent error correction, and a more stable point of view over time. This matters because, if consciousness has an adaptive function, it should not only appear in internal dynamics but also shape how an organism or agent behaves in the world. More specifically, if consciousness helps expand behavioral flexibility, enables real-time adjustment through memory of past experience, and supports a more globally integrated form of control, then systems with stronger consciousness-related profiles should not merely complete tasks successfully, but do so in a way that is more coherent, context-sensitive, and resistant to fragmentation under pressure.
If the real bats and the candidate-conscious robot bats begin to resemble one another more than they resemble the non-conscious robot bats, both in how perturbations propagate and in how they behave, we would not have proved that either is having subjective experience. But we would have reason to suspect that we might be onto something. More specifically, we would have evidence that certain ways of organizing cognition, which may or may not translate into subjective experience in all circumstances, may produce something closer to a unified perspective on the world: a mode of processing in which perception, memory, motivation, and action are not merely computed in isolated fragments, but integrated into a more coherent, flexible, and temporally stable point of view. In that sense, the experiment would be relevant not only to AI but also to the study of consciousness more broadly, because it would test whether functions often attributed to consciousness can be detected across very different substrates, provided they are all forced to make sense of reality through the same bat body.
Finding that consciousness matters for behavioral flexibility would not be surprising. Subjective experience may be adaptive because it expands an organism’s behavioral repertoire, perhaps allowing behavior to be adjusted in real time to a greater extent. Consciousness may have first evolved by linking memory to motivational control so past experience can guide action more flexibly and safely.
Perhaps this is relevant to our bat experiment in that it gives us a reason to compare behavior, memory use, and global coordination across real bats and robot bats. More specifically, if consciousness serves those functions, then systems with stronger consciousness-related profiles should not just solve tasks, but show more flexible, unified, context-sensitive behavior under the same bat-shaped constraints.
Why bats? One reason is that we have the famous essay/question “What Is It Like to Be a Bat?”famousquestion/essay, and it would be cool to have an experiment inspired by it. More importantly, bats are an interesting case because echolocation is an unusual form of active sensing in mammals, where perception and action are tightly coupled and continuously adjusted. Perhaps that makes bats useful for testing whether consciousness-related organization tracks flexible and globally coordinated behavior under different environmental demands. Of course, building robot bats that would be relevant to this experiment is more difficult than building robot mice, and the latter would work fine too.
Using AI systems meant to replicate some of the brain functionalities that complex creatures like mammals have could help us study consciousness without worrying that we just measure how well an artificial system imitates humans. Yes, it’s true that we don’t know for sure that mammals have conscious experience. That being said, if it turns out that consciousness helps expand behavioral repertoire, supports real-time updating from experience, and links memory to motivational control, then systems with stronger consciousness-related profiles should not merely solve tasks, but solve them in a more flexible, context-sensitive, and coherent way. We can test this premise by assessing differences between different biological and digital species in both behavioral repertoire and their cognitive systems, with the hope of finding some relationships that will strongly suggest we’re on the right path for cracking the nature of conscious experiences.
Bottom Line: We Can At Least Measure if AI Systems Have Integration and Differentiation
Perhaps the only way to tell for sure that a system other than your own is conscious is by somewhat becoming part of that system and using it to experience stuff. For example, you could link your brain to another one and have a shared space of experiences. Since mind sharing is not an option Uber or others can offer at present, we need to rely on more rudimentary approaches that might lead us in the proper direction.
One thing we can do is check whether AI’s internal processes are well-connected and work together as a whole (integration), and whether its responses are rich and varied rather than repetitive and rigid (differentiation). If AI systems that score higher on these measures also behave in more flexible, coherent, and adaptive ways, that’s at least a sign we might be on the path to figuring out how consciousness works. This approach could also allow us to estimate with a higher level of confidence the level of consciousness (if any) in non-human Earthly biological creatures – bats included!
Sources
Albantakis, L., Barbosa, L., Findlay, G., Grasso, M., Haun, A. M., Marshall, W., … & Tononi, G. (2023). Integrated information theory (IIT) 4.0: Formulating the properties of phenomenal existence in physical terms. PLoS computational biology, 19(10), e1011465. Link.
Butlin, P., Long, R., Elmoznino, E., Bengio, Y., Birch, J., Constant, A., Deane, G., Fleming, S. M., Frith, C., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., & VanRullen, R. (2023). Consciousness in artificial intelligence: Insights from the science of consciousness. arXiv. Link.
Casali, A. G., Gosseries, O., Rosanova, M., Boly, M., Sarasso, S., Casali, K. R., Casarotto, S., Bruno, M.-A., Laureys, S., Tononi, G., & Massimini, M. (2013). A theoretically based index of consciousness independent of sensory processing and behavior. Science Translational Medicine, 5(198), 198ra105. Link.
Casarotto, S., Comanducci, A., Rosanova, M., Sarasso, S., Fecchio, M., Napolitani, M., Pigorini, A., G Casali, A., Trimarchi, P. D., Boly, M., Gosseries, O., Bodart, O., Curto, F., Landi, C., Mariotti, M., Devalle, G., Laureys, S., Tononi, G., & Massimini, M. (2016). Stratification of unresponsive patients by an independently validated index of brain complexity. Annals of neurology, 80(5), 718–729. Link.
Comolatti, R., Pigorini, A., Casarotto, S., Fecchio, M., Faria, G., Sarasso, S., Rosanova, M., Gosseries, O., Boly, M., Bodart, O., Ledoux, D., Brichant, J.-F., Nobili, L., Laureys, S., Tononi, G., Massimini, M., & Casali, A. G. (2019). A fast and general method to empirically estimate the complexity of brain responses to transcranial and intracranial stimulations [Preprint]. bioRxiv. Link. (Note: The published version of the paper is Comolatti, R., Pigorini, A., Casarotto, S., Fecchio, M., Faria, G., Sarasso, S., Rosanova, M., Gosseries, O., Boly, M., Bodart, O., Ledoux, D., Brichant, J.-F., Nobili, L., Laureys, S., Tononi, G., Massimini, M., & Casali, A. G. (2019). A fast and general method to empirically estimate the complexity of brain responses to transcranial and intracranial stimulations. Brain Stimulation, 12(5), 1280–1289. Link.
Lacalli T. (2024). The function(s) of consciousness: an evolutionary perspective. Frontiers in psychology, 15, 1493423. Link.
Mashour, G. A., Roelfsema, P., Changeux, J. P., & Dehaene, S. (2020). Conscious Processing and the Global Neuronal Workspace Hypothesis. Neuron, 105(5), 776–798. Link.
Nagel, T. (1974). What is it like to be a bat? The Philosophical Review, 83(4), 435–450. Link.
Sarasso, S., Boly, M., Napolitani, M., Gosseries, O., Charland-Verville, V., Casarotto, S., Rosanova, M., Casali, A. G., Brichant, J. F., Boveroux, P., Rex, S., Tononi, G., Laureys, S., & Massimini, M. (2015). Consciousness and Complexity during Unresponsiveness Induced by Propofol, Xenon, and Ketamine. Current biology : CB, 25(23), 3099–3105. Link.
Sinitsyn, D. O., Poydasheva, A. G., Bakulin, I. S., Legostaeva, L. A., Iazeva, E. G., Sergeev, D. V., Sergeeva, A. N., Kremneva, E. I., Morozova, S. N., Lagoda, D. Y., Casarotto, S., Comanducci, A., Ryabinkina, Y. V., Suponeva, N. A., & Piradov, M. A. (2020). Detecting the Potential for Consciousness in Unresponsive Patients Using the Perturbational Complexity Index. Brain sciences, 10(12), 917. https://doi.org/10.3390/brainsci10120917. Link.
[1] Lempel–Ziv is a way of measuring how complex or compressible a pattern is: if a signal is very repetitive, it can be compressed a lot and gets a low Lempel–Ziv score; if it is varied and less predictable, it is harder to compress and gets a higher score.
[2] Normalize by source entropy means they adjust that complexity score based on how much possible information or variability the original signal could contain, so the final value does not just reflect having more raw activity or noise, but how complex the pattern is relative to the amount of information available. Entropy here refers to how much possible variation the original binary pattern could have, given the proportion of 0s and 1s. Signal entropy is a measure of how unpredictable or variable a signal is, that is, how much potential information it contains. For a binary signal made of 0s and 1s, if it is almost all 0s, or almost all 1s, the entropy is low because it is very predictable; if it contains a more balanced mix of 0s and 1s, the entropy is higher because each next value is less predictable. After measuring complexity with Lempel–Ziv, they divide or scale that value by the signal’s entropy, so the final result reflects how complex the pattern is compared with how much variability was available in principle. In simpler terms: a pattern is not supposed to get extra credit merely for being noisy or full of activity; it gets credit for being richly structured relative to its own informational possibilities.
