# IIT, half masked and half disfigured

**Giulio Tononi^1^, Melanie Boly^1^, Matteo Grasso^1^, Jeremiah Hendren^2^, Bjorn Juel^1^, William Mayner^1^, William Marshall^1^, Christof Koch^3^**

^1^University of Wisconsin-Madison, ^2^Ludwig Maximilian University of Munich, ^3^Allen Institute of Brain Science

^1^Department of Psychiatry, 6001 Research Park Blvd, Madison WI 53719; ^2^Großhaderner Str. 2, 82152 Planegg, Germany; ^3^615 Westlake Ave N, Seattle, WA 98109, United States.

gtononi@wisc.edu, boly@neurology.wisc.edu, grasso2@wisc.edu, jeremiah.hendren@campus.lmu.de, bjuel@wisc.edu, mayner@wisc.edu, wmarshall3@wisc.edu, christofk@alleninstitute.org

Half masked and half disfigured, how could its identity *not* be mistaken? For such is the portrait of integrated information theory (IIT) painted by the target article. Our main task here will thus be to sketch a picture of IIT that is at least recognizable (for a proper introduction, see (Haun & Tononi, 2019; Oizumi, Albantakis, & Tononi, 2014; Tononi, 2015), focusing on a few key points.[^1]

IIT addresses consciousness starting not, as is usually done in neuroscience, from its behavioral, functional, or neural correlates, but from phenomenology—the existence of one’s own experience, which is immediate and indubitable. IIT then identifies the properties of consciousness that are *essential*—immediate, indubitable, and true of every conceivable experience—namely intrinsicality, composition, information, integration, and exclusion. These properties, which have always been at the core of IIT, were more recently codified as “axioms”(Oizumi et al., 2014; Tononi, 2012) because nothing can be more axiomatic as a starting point than what is true of every conceivable experience.

Because the axioms capture properties that are essential, it should be obvious that none of them can be left out without destroying the whole edifice. Yet Merker et al. only consider information and integration, virtually ignore intrinsicality and composition, and dismiss exclusion by claiming (gratuitously) that it was added “ad hoc” to fix unpleasant consequences of information and integration. Not even Euclid’s venerable geometry would survive without three of its five axioms, nor would physical theories without half the terms in the defining equations. Hence, this response must begin by restating the five axioms of IIT (adapted from(Haun & Tononi, 2019).

![Figure 1](/corpus/figures/tononi-2022/fig01.png)

*Intrinsicality* means that every experience is *subjective*—it is for the subject of experience, from its own private, intrinsic perspective, rather than for something extrinsic to it. Using spatial experiences as an example, if I experience the sight of my bedroom, the canvas of phenomenal space is experienced by me, not by somebody else. *Composition* means that every experience is *structured*, being composed of phenomenal distinctions and relations. In the case of spatial experiences, distinctions can be taken as “spots” (areas of space of any size and shape), and their relations as the way spots connect, fuse, and include each other. *Information* means that every experience is the *specific* way it is. For example, the canvas of space may be homogeneous, or embossed by inhomogeneities such as a body, a book, a bed, in countless possible configurations. *Integration* means that every experience is *unitary*, being irreducible to independent components. Thus, the canvas of space cannot be reduced to a left side and a right side that are experienced independently—if it were so, there would be two independent consciousnesses rather than one. Finally, *exclusion* means that every experience is *definite*—it contains what it contains, neither less nor more. In the case of space, the canvas contains all the spots and relations it contains—not just the left or the right side, or just a single spot in the middle or in the periphery—but it does not extend beyond its borders, such as behind one’s head. Importantly, these five essential properties are true of every conceivable experience, not just of spatial ones. They also delineate what must be accounted for by a satisfactory explanation of consciousness.

To provide an explanation for experience in physical terms, every phenomenal property must be accounted for by the causal properties of the physical substrate of consciousness (PSC). Therefore, the next step of IIT is to “translate” the essential phenomenal properties into essential physical properties (called “postulates”). In IIT, “physical” is defined in purely operational terms as cause-effect power—being able to take or make a difference—whether we are dealing with a Higgs boson, a stone, a computer, a body, or a brain. In principle, the cause-effect power of a physical system is characterized in full by its transition probability matrix (TPM)—how it responds to all possible perturbations of its state. From the TPM of a system, say, the one characterizing a portion of our cerebral cortex, we can then *unfold* all its causal properties and assess whether it satisfies the five essential properties required of a PSC. If so, unfolding yields a maximally irreducible *cause-effect structure* composed of causal distinctions and relations that capture all the physical properties of the PSC.

IIT then proposes its fundamental, explanatory identity, according to which the components of the cause-effect structure correspond one-to-one to the components of experience—to the quality of a specific experience. Thus, every content of an experience “here and now,” including not just space and objects, but also memories and expectations, correspond to sub-structures in a cause-effect structure. Furthermore, the quantity of experience—how much one exists phenomenally—is measured by the irreducibility Ф of the cause-effect structure.[^2]

The explanatory and predictive validity of IIT can be assessed by considering whether the PSC satisfies the five physical properties that the theory considers essential, but other regions do not. Moreover, the PSC should cease to satisfy these properties when consciousness vanishes in dreamless sleep or under anesthesia, as confirmed by studies explicitly designed to test this prediction of IIT with TMS-EEG (Tononi, Boly, Massimini, & Koch, 2016). The explanatory identity of IIT also predicts that the cause-effect structure unfolded from the PSC should account for the specific properties of specific experiences. For example, (Haun & Tononi, 2019), demonstrate how the phenomenal properties that make space feel extended (reflexivity, connection, fusion, and inclusion) can be accounted for in the cause-effect structures unfolded from neural units with grid-like connectivity, as evident in topographically organized areas of posterior cortex. We encourage the reader to consult this paper because it represents the first attempt to apply IIT to account for the quality of experience from first principles. It also demonstrates IIT’s explanatory and predictive power with respect to a large body of neurological evidence concerning the perception of space. Ironically, completely ignoring this account, Merker et al. single out “the geometries of visual space” as something that should one day be explained. [^3]

After this all-too-short summary of IIT, let us consider just some major ways in which Merker et al. make the theory unrecognizable by ignoring or misinterpreting the axioms. Let us briefly illustrate this by delving a bit deeper into the axiom of exclusion, whose sidelining leads to all kinds of erroneous inferences.

As stated above, exclusion means that every experience is *definite*—it contains what it contains, neither less nor more: for example, I experience the entire visual field, not just one half or any smaller (or larger) portion of it. Being definite is taken as an axiomatic property because, first, it is immediately given—I do not have to infer that my experience contains what it contains; it just does. Second, that an experience is definite is indubitable—an experience that contained, say, less than it contains, such as the left half of the visual field only, would simply be another experience that contains exactly what it contains, thus reaffirming the validity of the axiom. Finally, being definite is true of every conceivable experience.

As with all the other axioms, exclusion is translated in physical terms—namely cause-effect power. Accordingly, the PSC must have cause-effect power that is also *definite*, in the sense of having a border: its cause-effect structure must be specified by a definite set of units—neither less nor more. The border is established by assessing which set of units unfolds into a cause-effect structure that is maximally irreducible, as measured by Ф=Ф^max^. In IIT, Ф is the quantifier of existence (to exist, something must be irreducible to independent components), so ultimately the border of the PSC is defined by which set of units “exists the most” among competing sets.[^4] But if having Ф=Ф^max^ defines the set of units constituting the PSC, it also excludes any overlapping set having Ф\<Ф^max^.

A direct consequence of IIT is then that only maxima of Ф correspond to conscious entities (more or less conscious depending on their Ф value). This excludes all kinds of things Merker et al. erroneously list as IIT-conscious. For example, a society to which I belong cannot be conscious simply because I am conscious, which excludes from existence any superset that includes my PSC (no “group mind”). Digital computers cannot be conscious, even if they were functionally equivalent to us, because their physical substrate, unlike ours, disintegrates into trivially small units of minimal Ф.[^5] For the same reason, the internet or any network of computers cannot be conscious. Expander graphs are not conscious because they are not physical systems but mathematical abstractions (Aaronson, 2014). As for physical “grids,” it all depends on the nature of the units, on their fan-in and fan-out, and on the level of noise. Most likely, power grids should harbor little hope of being conscious. In short, Merker et al. express their own misconceived assumptions, rather than conclusions based on IIT and its axioms.

In fact, unlike Merker et al., at this stage we prefer not to speculate. As argued in various papers (Hoel, Albantakis, Marshall, & Tononi, 2016; Hoel, Albantakis, & Tononi, 2013; Marshall, Albantakis, & Tononi, 2018), also ignored by Merker et al.), to draw any conclusion about a candidate substrate of consciousness, one should first establish at which grain that physical substrate allows for maximum Ф (another consequence of exclusion). Consider the grain of our own PSC: is it neurons, micro-columns, or mini-columns? Is it over micro-seconds, milliseconds, or seconds (Tononi et al., 2016)? Indeed, predictions about grain and borders are in principle testable: if it turns out that the units of our PSC are not at the grain that allows for maximum Ф, IIT is wrong. Similarly, IIT is wrong if the theoretical PSC—the set of brain units yielding maximum Ф—does not correspond to the empirical PSC.

A critical point here is that the predictions of IIT must be validated *in* us, because only in us can phenomenology and physics be directly compared. It is only on the strength of that validation that we can venture inferences about consciousness *outside* us. Unfortunately, this realization eludes Merker et al., as revealed by this quote: “the claim of sufficiency would require independent evidence that…all…natural and artificial systems that…combine integration with differentiation…are in fact conscious.” One wonders what they have in mind, for what could possibly count as “independent evidence” for phenomenology? What makes consciousness special, and different from any other object of scientific inquiry, is precisely that it is subjective and private. As captured by the intrinsicality axiom, immediate evidence for phenomenology is only available from the intrinsic perspective of the subject of experience.[^6] Therefore the only viable scientific avenue is to develop a theory that adequately accounts for the presence and properties of our consciousness—for its quality and quantity—based on its physical substrate in our brain, and to extrapolate from there (inference *from* a good explanation, Ellia et al, in press). Conversely, inferences based on intuitive, unexamined assumptions of anatomical or functional complexity, on pre-theoretical notions, or on “gut feelings” (Aaronson, 2014), should be considered with great skepticism. In fact, a highly “complex” neural architecture or an involvement in “complex” cognitive functions is no guarantee of consciousness.[^7]

Let us now briefly examine the two postulates actually considered by Merker et al. In the target article, the main refrain is that while integration and information may well be necessary requirements for consciousness, they are not sufficient.[^8] But who said they would be? As summarized above, IIT explicitly says the PSC must satisfy all five postulates, not just two. So countless substrates may score high on information integration, but only maxima of Ф can be a PSC.

Furthermore, even integration and information are misrepresented in critical ways. The integrated information of IIT is explicitly not (Tononi et al., 2016), as assumed by Merker et al., a purported “high capacity for efficient global information transfer” of a network. The informativeness of a specific experience—what the experience means—corresponds to a specific *cause-effect* *structure*—not to a *message* to be “transferred” across a channel.[^9] In fact, one of the counterintuitive predictions of IIT is that a neural substrate might be conscious even in the absence of neural activity, with no function being performed and no information being transferred in the Shannon sense. Indeed, if IIT is correct, seeing a “blank screen” corresponds to a cause-effect structure of immense richness (informativeness), even though the experience can be roughly described and communicated to another human being in just two words (Haun & Tononi, 2019).[^10] Related predictions are that changes in connectivity, in the absence of changes in activity, can be associated with changes in experience, as can be tested with psychophysical experiments (Song, Haun, & Tononi, 2017; Tononi et al., 2016).

Before ending, we can only touch upon two additional criticisms. Throughout their article, Merker et al. emphasize that we typically feel as if we were “looking out” at the world from a “perspective” located \~4 cm behind the bridge of the nose, which they call the “ego-center.” They complain that the “intrinsic perspective” of IIT does not account for this prominent “point of view.” Indeed. The intrinsic perspective of IIT expresses the axiom of intrinsicality—a property that is true not just of viewer-centered experiences but of every conceivable experience: it means “from the inside”—for the subject of experience—as opposed to the extrinsic perspective of an observer—“from the outside.” Nonetheless, the feeling of a perspectival ‘center’ of space is an intriguing phenomenal property that should be eventually explained in physical terms. To quote from IIT’s account of spatial experience (Haun & Tononi, 2019): “Apart from some asymmetries due to border effects…, a natural center may be provided by grids specifying body space, which would be heavily bound by relations to the middle of visual space, and by grids responsible for initiating movements near the body midline.”

Merker et al. also complain that proxies for Ф proposed by different groups can lead to different results, and they argue that “integrated information itself is a hypothetical construct in search of a unique and well-defined formal definition and measure.”[^11] This too is ironic because recent papers from our group (also not cited) have done just that. By formalizing all the postulates of IIT, we demonstrated that it is possible to obtain a measure of intrinsic integrated information that is in fact unique (Barbosa, Marshall, Albantakis, & Tononi, 2021; Barbosa, Marshall, Streipert, Albantakis, & Tononi, 2020). Uniqueness is important to IIT because an experience is the way it is, and so should be its physical correspondent (rather than varying with different measures of irreducibility). Admittedly, before this theoretical development, IIT deliberately made do with provisional measures that captured at least some of the essential properties of experience. In fact, IIT remains “work in progress.” The first attempt to account for the quality of experience—specifically of spatial extendedness—is comparatively recent (Haun & Tononi, 2019). Ongoing work aims at accounting for the feeling of temporal flow and of objects, at revealing the borders of the PSC through explicit approximations of Ф from fMRI time series, and at assessing how intrinsic meanings—substructures in a cause-effect structure—can match causal regularities in the environment. While these are incremental steps, they are principled steps, and hopefully steps in the right direction.

In fact, it is one of the key consequences of IIT—that all meaning is intrinsic—which may provide an apt conclusion to this response. And meanings such as existence, subjectivity, information, space, and time, and of course consciousness itself, are likely to be highly entrenched sub-structures. So in the end it is not surprising that the theory Merker et al. decry is one they think they see, rather than the one that is actually there. Theirs is, indeed, a case of mistaken identity.

## References

Aaronson, S. (2014). Why I Am Not An Integrated Information Theorist (or, the Unconscious Expander) 

Barbosa, L. S., Marshall, W., Albantakis, L., & Tononi, G. (2021). Mechanism Integrated Information. *Entropy, 23*(3), 362.

Barbosa, L. S., Marshall, W., Streipert, S., Albantakis, L., & Tononi, G. (2020). A measure for intrinsic information. *Sci Rep, 10*(1), 18803. doi: 10.1038/s41598-020-75943-4

Dehaene, S., Lau, H., & Kouider, S. (2017). What is consciousness, and could machines have it? *Science, 358*(6362), 486-492.

Edelman, G. M., & Tononi, G. (2000). *A universe of consciousness How matter becomes imagination*. New York: Basic Books.

Ellia, F., Hendren, J., Grasso, M., Kozma, C., Mindt, G., Lang, J., . . . Tononi, G. (2021). Consciousness is a structure, not a function. *In press.*

Findlay, G., Marshall, W., Albantakis, L., Mayner, W., Koch, C., & Tononi, G. (2019). *Dissociating Intelligence from Consciousness in Artificial Systems – Implications of Integrated Information Theory*. Paper presented at the TOCAIS 2019 ( Towards Conscious AI Systems  ), Stanford, CA.

Haun, A., & Tononi, G. (2019). Why does space feel the way it does? Towards a principled account of spatial experience. *Entropy, 21*(12), 1160.

Hoel, E. P., Albantakis, L., Marshall, W., & Tononi, G. (2016). Can the macro beat the micro? Integrated information across spatiotemporal scales. *Neurosci Conscious, 2016*(1), niw012. doi: 10.1093/nc/niw012

Hoel, E. P., Albantakis, L., & Tononi, G. (2013). Quantifying causal emergence shows that macro can beat micro. *Proc Natl Acad Sci U S A, 110*(49), 19790-19795. doi: 10.1073/pnas.1314922110

Marshall, W., Albantakis, L., & Tononi, G. (2018). Black-boxing and cause-effect power. *PLoS Comput Biol, 14*(4), e1006114. doi: 10.1371/journal.pcbi.1006114

Oizumi, M., Albantakis, L., & Tononi, G. (2014). From the phenomenology to the mechanisms of consciousness: Integrated Information Theory 3.0. *PLoS Comput Biol, 10*(5), e1003588. doi: 10.1371/journal.pcbi.1003588

Song, C., Haun, A. M., & Tononi, G. (2017). Plasticity in the Structure of Visual Space. *eNeuro, 4*(3). doi: 10.1523/ENEURO.0080-17.2017

Tononi, G. (2012). The integrated information theory of consciousness: an updated account. *Archives italiennes de biologie, 150*(2/3), 56-90.

Tononi, G. (2015). *Integrated Information Theory* *Scholarpedia*.

Tononi, G., Boly, M., Massimini, M., & Koch, C. (2016). Integrated information theory: from consciousness to its physical substrate. *Nat Rev Neurosci, 17*(7), 450-461. doi: 10.1038/nrn.2016.44

Tononi, G., McIntosh, A. R., Russell, D. P., & Edelman, G. M. (1998). Functional clustering: identifying strongly interactive brain regions in neuroimaging data. *Neuroimage, 7*(2), 133-149. doi: 10.1006/nimg.1997.0313

Tononi, G., Sporns, O., & Edelman, G. M. (1994). A measure for brain complexity: relating functional segregation and integration in the nervous system. *Proc Natl Acad Sci U S A, 91*(11), 5033-5037. doi: 10.1073/pnas.91.11.5033

Wagner, M. (2006). *The Geometries of Visual Space*. London, UK: Psychology Press.

[^1]: Addressing all misrepresentations in the target article would require a response just as long.

[^2]: The fundamental identity of IIT is explicitly between an experience, with all its specific properties, and a cause-effect structure. It should be obvious that an experience cannot be identical to a single number, Ф(\>0), which is instead the quantifier of how much the experience exists.

[^3]: The reference is to (Wagner, 2006), discussed and quoted in (Haun & Tononi, 2019).

[^4]: In accordance with the principle of maximal existence (Tononi, 2015). The border separates units inside and outside the PSC even though they may be connected.

[^5]: (Findlay et al., 2019; Oizumi et al., 2014)

[^6]: In short, I only have immediate evidence for my own experience. Everything else is an inference, including the presence of consciousness in other human beings (in philosophy, this is known as the problem of other minds).

[^7]: If anything, rather than “panpsychist,” IIT is much more restrictive than many functionalist notions about consciousness that would ascribe consciousness to intelligent machines that can broadcast information globally or monitor the trustworthiness of sensory information (Dehaene, Lau, & Kouider, 2017).

[^8]: This is a typical initial reaction to IIT even when all five postulates are duly considered (unlike Merker et al.). One might concede that certain brain regions may be uniquely suited to satisfying those properties and provisionally accept the empirical evidence indicating that those regions may constitute the PSC. But one still cannot imagine how a physical system—even one that is a maximum of intrinsic, structured, specific, unitary cause-effect power—would then “give rise” to experience. This attitude is understandable because as neuroscientists (and generally as scientists), we take the extrinsic perspective: we take a physical substrate—the brain—as our starting point and try to figure out how it works and performs its various functions. And from this extrinsic perspective, there seems to be no way to derive the very fact that consciousness exists—the existence of an intrinsic, phenomenal perspective. This is why, when it comes to consciousness, the starting point must be consciousness and its properties, not a physical substrate. What we should expect of physics is to account for the properties of experience, not to “give rise” to them. Phenomenology is first and physics second: physics is our way to characterize operationally—from within our own consciousness—what we infer to exist independent of it (Ellia et al., 2021).

[^9]: This point was made, for example, in (Edelman & Tononi, 2000).

[^10]: Just to give a sense, the cause-effect structure unfolded from a grid of 8 binary units explored in (Haun & Tononi, 2019), to account for spatial experience is composed of 133 distinctions and \>10^40^ relations. As explicitly stated by the explanatory identity of IIT, an experience corresponds to an (immensely rich) cause-effect structure, not to a single number Ф, which only quantifies its irreducibility.

[^11]: Merker et al. mention early work that introduced a measure of complexity C~N~ (Tononi, Sporns, & Edelman, 1994) which is sensitive to the dual requirements of information and integration (based on the compositional sum of average mutual information between subsets and the rest of a system) and has widespread applicability. They point out approvingly that C~N~ (unlike Ф) was not claimed to be a measure of consciousness. This is correct, because it was already clear then that all essential properties of consciousness need to be satisfied in physical terms, among them intrinsicality and exclusion. In fact, other early work attempted to capture these other properties by introducing a measure of functional clustering to identify the borders of a maximally integrated entity (Tononi, McIntosh, Russell, & Edelman, 1998). But it was also clear that a proper measure of consciousness should eventually account for all its properties at once, including composition and information, and should be state-dependent rather than an average. Only in this way can we account for the fact that every experience exists in full, immensely structured and fully specific, “here and now.” or, as elegantly put by Merker et al., “in a perpetual pristine present for which no content outside the state it is in exists”.
