For a far more distant case, someone's hair color canaffect conceptual structure through the spatial–conceptual
interface, and, if expressed in language, have an indirect effect on produced phonological structure through the chain
of interfaces (the speaker may sayHey! You dyed your hair!). Even the output of the semicircular canals may have an
effect on phonological structure (I'm dizzy!), but the effectis distant, mediated by a long chain of intermediate levelsof
structure and interfaces.
On the other hand, the motion of the speaker'slipsmay have adirecteffect on the hearer's phonological structure,
through the “McGurk” interface. So here, in a very small domain, gerrymandered in both input and output
components, we have more informationflow, and more direct informationflow, than in rando maspects of vision and
phonology.
In other words, the presence of an interface between two levels of structure is what makes themnotinformationally
encapsulated fro meach other. The richness of the infor mationflow depends on the degree to which the interface
module is more domain-specific than the structures it connects. The phonology–syntax interface is relatively rich; the
“McGurk”interface is not.
In principle, an interface module might be precisely as domain-specific as the structures that serve as its input and
output: every distinctionin theinput structure could make a differencein theoutput structure, and every distinctionin
the output structure could be affected by something in the input structure. In such a case, though, it would hardly
makesensetospeak oftwodistinct“modules”anymore,sincetheirinteractions wouldbemore or less unconstrained.
We might as well see the union of their domains as a single larger domain, and we might as well see the interface
module just as embodying some of the principles involved in integrating this larger domain.
It therefore begins to make sense to speak of“degrees of modularity”rather than absolute modularity. Two domains
connected by a narrow“information bottleneck”will be relatively modular: not very many parts of each domain can
affect the other. As the interface becomes richer, more parts can interact.If communication between the twodomains
is wide open, it is impossible to say where one leaves off and the other begins. Given the gradual nature of this
transition from relatively modular to nonmodular, it is impossible to draw a precise boundary on modularity.
A cynic might say therefore that the issue of modularity is dissolved. I would disagree. Rather, it developsinto a more
nuanced set of questions: What families of distinctions form richly interconnected and well-integrated domains, and
where are there more restricted“information bottlenecks”? The correlates in processing are as they always have been:
Can one identify particular stages of informational integration, enhanced by priming or disrupted by concurrent