Content uploaded by Piers Kelly
Author content
All content in this area was uploaded by Piers Kelly on Mar 08, 2022
Content may be subject to copyright.
The Predictable Evolution of Letter Shapes
An Emergent Script of West Africa Recapitulates Historical Change
in Writing Systems
Piers Kelly, James Winters, Helena Miton, and Olivier Morin
A familiar story about the evolution of alphabets is that individual letters originated in iconic representations of real things. Over time,
these naturalistic pictures became simplified into abstract forms. Thus, the iconic ox’s head of Egyptian hieroglyphics transformed
into the Phoenician and eventually the Roman letter A. In this vein, attempts to theorize the evolution of writing have tended to
propose variations on a model of unilinear and unidirectionalprogression. According to this progressivist formula, pictorial scripts will
tend to become more schematic while their systems will target smaller linguistic units. Objections to this theory point to absent,
fragmentary, or contrary paleographic evidence, especially for predicted transitions in the underlying grammatical systems of writing.
However, the forms of individual signs, such as the letter A, are nonetheless observed to change incrementally over time. We claim that
such changes are predictable and that scripts will in fact become visually simpler in the course of their use, a hypothesis regularly
confirmed in transmission chain experiments that use graphic stimuli. To test the wider validity of this finding, we turn to the Vai script
of Liberia, a syllabic writing system invented in relative isolation by nonliterates in ca. 1833. Unlike the earliest systems of the ancient
world, Vai has the advantage of having been systematically documented from its earliest beginnings until the present day. Using es-
tablished methods for quantifying visual complexity, we find that the Vai script has become increasingly compressed over the first 171
years of its history, complementing earlier claims and partial evidence that similar processes were at work in early writing systems. As
predicted, letters simplified to a greater extent when their initial complexity was higher.
If all writing systems undergo observable historical change, is
it possible to identify the underlying evolutionary processes and
predict their trajectories? A long-discussed hypothesis holds
that new scripts become visually simpler and more systematic as
they are transmitted from generation to generation over his-
torical time (Changizi and Shimojo 2005; Dehaene 2009; Trig-
ger 2003:600–602). One of the greatest challenges in testing this
claim is the fact that we do not have a complete record of the first
writing systems, especially in the earliest years of their devel-
opment. We address this problem by drawing on data from the
Vai script of Liberia, a syllabary developed in isolation by non-
literates in ca. 1833 and still in use today. We find that Vai letters
have become more compressed over the first 171 years of their
history, and we suggest that this finding has broad applicability
to the evolution of early writing systems.
One of the earliest to articulate a theory of the evolution of
writing was Rousseau, who proposed that writing passed through
three stages, from pictures, to whole words, and finally to alpha-
bets. “The depicting of objects is appropriate to a savage people,”
he wrote in his Essai sur l’origine des langues,“signs of words and
of propositions, to a barbaric people, and the alphabet to civilized
peoples [peuples policés]”(Rousseau 1966 [1781]:17). This basic
proposition was to remain largelyunchallengedfornearlytwo
centuries, but the sensational decipherment of Egyptian hiero-
glyphics on the part of Champollion (1824) generated a new
wave of interest in the evolution of writing. Contrary to Cham-
pollion’s expectations, the Egyptian hieratic script was apparently
far cruder in its form than the monumental hieroglyphic script
that had preceded it, which lent the impression that it was re-
gressing to a more “primitive state”(Champollion 1836). In re-
sponse, the sinologist Pauthier addressed the apparent contra-
diction of style and chronology in Egyptian scripts, maintaining
that progressive schematization was a defining process in the
evolutionofwriting.InanupgradetoRousseau’s tripartite model,
Pauthier postulated a so-called figurative age of writing that was
characterized by pictorial representations of objects, while in a
subsequent transitory age, these iconic images become more
conventional and abstract. In a final pure alphabetic age, only the
simplest phonological elements were represented in graphic form.
Alphabets were thus doubly compressed. Not only were they
simpler in the outward shape of their scripts but also their un-
derlying orthographic systems were reduced to the bare phone-
mic essentials. Illustrating this progression, Pauthier referred to
documented changes in both Egyptian and Chinese writing sys-
tems toward greater abstraction and simplicity. In his estimation,
the process could even be invoked as a method for dating pa-
leographic inscriptions since it was a law that “suffered no ex-
ception”(Pauthier 1838:567). Later, the classicist Ritschl (1869)
Piers Kelly is a Deutsche Forschungsgemeinschaft Fellow at the University of New England (Armidale, New South Wales 2350, Australia [piers
.kelly@gmail.com]). James Winters is an Assistant Professor at the School of Collective Intelligence, Mohammed VI Polytechnic University
(Lot 660, Hay Moulay Rachid, BenGuerir 43150, Morocco). Helena Miton is a Postdoctoral Fellow at the Santa Fe Institute (1399 Hyde Park Road,
Santa Fe, New Mexico 87501, USA). Olivier Morin is Research Group Leader of the Mint Research Group at the Max Planck Institute for the
Science of Human History (Kahlaische Strasse 10, 07745 Jena, Germany), and Chargé de Recherches CNRS at the Département d’études cognitives
ENS, EHESS, CNRS, PSL University, UMR 8129. This paper was submitted 17 X 18, accepted 10 II 20, and electronically published 3 I 22.
Current Anthropology, volume 62, number 6, December 2021. q2021 The Wenner-Gren Foundation for Anthropological Research. All rights reserved. Published by
The University of Chicago Press for the Wenner‐Gren Foundation for Anthropological Research. https://doi.org/10.1086/717779
would seek to formalize a new “scientificpaleography”that rec-
ognizedthathistoricalscriptchangewasnotarbitrarybuta
matter of “internal development”governed by “certain laws and
guiding impulses.”
1
Independently of the French and German tradition, Morgan
and Tylor were to reinvent Rousseau’s evolutionary stages of
writing in the context of their more ambitious teleological
model of human progress. For Morgan (1877), a society’suseofal-
phabetic writing, as opposed to other systems, was diagnostic of
its ascension to the highest status of civilization. Tylor (1878
[1865]:374) was even more explicit in linking technological changes,
such as a shift from stonework to metalwork, with transitions
from “picture-writing”to the alphabet. For the progressivists,
therefore, there was an implied inverse relationship between the
(perceived) complexity of a society’s material culture and the
outward simplicity of its writing system.
Although progressivism was to be roundly rejected from within
anthropology, it found fertile ground in the field of gramma-
tology. Taylor (1883a, 1883b), Clodd (1900), Ege (1921), Mason
(1928), Gelb (1963 [1952]), and Cohen (1958) gave the social
evolutionist framework a renewed impetus in their stadial for-
mulas for the evolution of writing. Their chief theoretical differ-
ences lay only in the number of discrete stages that writing was
supposed to pass through on the road to alphabetic abstraction.
Schematization as a General Law
Nineteenth-century European scholars often wondered whether
the proposed principle of progressive schematization might be
applicable to symbolic behavior more generally. Pitt-Rivers and
Balfour assessed graphic artifacts collected in colonial settings—
including New Ireland paddles, Peruvian pottery, and decorated
spear shafts from the Solomon Islands—and attempted to arrange
them in sequences from the most representational to the most
schematic. This exercise was intended to illustrate how figurative
forms become conventionalized through constant reproduction,
with their superfluous features falling away in the process:
Thus, the drawing of a stag would be made to convey infor-
mation to people at a distance that there was a herd of deer in
the neighbourhood to be hunted; and as the object of the
drawing was no longer to depict truthfully the peculiarities of
the beast, but merely to convey information, the amount of
labour expended upon it would be the least that could be
employed for the required purpose. All writtencharacters have
originated in this way; and no one now requires to be told
how pictographic representations developed into hieroglyphic
and subsequently into phonetic characters. (Pitt-Rivers 1906
[1875]:40)
Graphic communication systems, be they cave paintings or
writing systems, were therefore seen to followa principle of least
effort. One of the presumed mechanisms for schematization was
to be found in the imperfect reproduction of a graphic figure
over successive iterations. Accordingly, Balfour (1893) empha-
sized the role of “unconscious variation”whereby “the changes
are not intentional, but are due to want of skill or careless
copying, difficulty of material, or reproducing from memory”
(77). Reprising Pitt-Rivers, he maintained that figurative art
must surely have been ancestral to modern alphabetic writing
(Balfour 1893:98).
2
Balfour was also the first to develop experiments for testing
the effects of cultural transmission on items of symbolic culture.
In Balfour’s tasks, an experimental participant was asked to
create a sketched reproduction of a picture, and this sketch was
then copied by a second participant, and so on. According to his
contemporary, the French philosopher Philippe (1897), these
kinds of experiments revealed a “law of economy,”by which
“unnecessary details fall and disappear to make room for what is
necessary for the whole,”adding that such transformations must
scale up to broader cultural changes at the population level
(493).
Bartlett would later design similar pencil-and-paper tasks
with the aim of testing the dynamics of human memory. No-
ticing that figurative images tended to become simpler and more
conventionalized over successive “generations,”he perceived a
real-world parallel in the historical evolution of writing sys-
tems whereby “realistic pictures”eventually became “non-
representational conventional signs”(Bartlett 1995 [1932]:95).
So taken was he by this observed correspondence that he was
inspired to use Native American pictograms and Egyptian hi-
eroglyphics as the primary input for a number of his experi-
mental tasks (Bartlett 1995 [1932]:97, 180).
3
In a more recent
1. “The fact that historical changes within a script are not a matter of
chance or arbitrariness, but rather are associated with an internal develop-
ment that progresses according to certain laws or guiding impulses is surely
indisputable: it is on this insight that the whole concept of scientific‘pale-
ography’is based”(“Dass die geschichtlichen Veränderungen einer Schrift
nicht Sache des Zufalls oder der Willkür sind, sondern viel-mehr im Zu-
sammenhange einer innern Entwicklung stehen, die nach gewissen bestim-
menden Gesetzen oder doch leitenden Trieben vor sich geht, wird wohl im
allgemeinen von niemand verkannt: wie denn auf dieser Einsicht der ganze
Begriff einer wissenschaftlichen ‘Paläographie’beruht”; Ritschl 1869:1).
2. Perhaps the greatest champion for this theory was the epigrapher
Arthur Evans, whose Scripta Minoa (1909) proposed a much deeper lineage
for writing than any of his predecessors. While the Paleolithic archeologist
Henri Breuil (1905) was prepared to suggest that stylistic transitions in
Magdalenian rock art were analogous to the evolution of writing, Evans
hinted at a much more direct historical pathway. Referring to Breuil’s
sketches, he wrote that “signs of curiously alphabetic aspect—at times even
in groups—are seen engraved on reindeer-horns or ivory, or on the surface
oftherockitself....Certainsignscarvedonafragmentofreindeerhorn
are specially interesting from the primitive anticipation that they present of
the Phoenician alef ”(Evans 1909:3–4). Eventoday, a presumed relationship
between geometric signs in rock art and modern writing continues to be
entertained (see, e.g., von Petzinger 2016).
3. Bartlett (1995 [1932]) noted that in transmission chain experiments,
“some element of an original complex gradually attains a more and more
important position and comes to stand for all the rest. The long story of the
development of the common alphabetical formsis a case in point”(271).It is
000 Current Anthropology Volume 62, Number 6, December 2021
example, Garrod et al. (2007) developed Pictionary-style graphic
tasks that detected a transition toward noniconic simplicity,
leading them to speculate that similar processes must have un-
derpinned the loss of iconicity in the cuneiform and Chinese
writing systems over deep timescales. One preliminary attempt
to compare complexity rates in natural writing systems was made
by Hegenbarth-Reichardt and Altmann (2008), who applied a
hand-coded method to show that a specific historical instan-
tiation of the hieratic script was quantifiably simpler than the
hieroglyphic script used in the same period.
A Revised Hypothesis on the Evolution
of Letter Forms
A unifying argument across these disparate fields, from nineteenth-
century paleoanthropology to contemporary transmission chain
experiments, is that graphic forms proceed through a process of
simplification that transpires in the course of their recollection and
reproduction by individual writers and their transmission from
one individual to another. However, the term “compression”is a
more accurate descriptor than “simplification”for this process. In
other words, through repeated interactions, a system of signs will
become compressed so that the same amount of information is
expressed with less descriptive effort. Although this proposition is
well confirmed under controlled experimental conditions (Ta-
mariz 2017), it not always clear how it might apply to the dynamics
of real-world historical change in writing systems.
It is remarkable, after all, that the world’s earliest independent
writing systems have all exhibited extensive iconicity and that in
many instances this iconicity diminished in later forms or
derivatives of the script. As a counterobservation, one could
point out that simple and nonfigurative symbolswere already in
wide use at the dawn of writing in Mesopotamia or that iconicity
in Egyptian hieroglyphics was retained over the entire history or
the script. Meanwhile, the Olmecoid scripts of Mesoamerica are
apparently less figurative than the Maya script that postdates
them archeologically (Downey 2014; Mignolo 1989), while the
Maya script itself retained its complexity and iconicity. These,
however, are not falsifying examples. We do not yet know
whether the pre-Maya scripts of Mesoamerica were glottic
writing (representing language) or whether they have a direct
historical relationship to the better-attested Maya script, just as
the abstract markings on neolithic Chinese pottery cannot be
proven to be ancestral to early Chinese writing. As for Egyptian
writing, there is no doubt that the more schematic hieratic script
was derived from Egyptian hieroglyphics and not the other way
around (Altenmüller 2005; but see Goedicke 1988:vii–viii).
Moreover, hieroglyphics existed in a rich graphic context that
included not just hieratic and demotic scripts but also an artistic
tradition with a well-defined repertoire of figurative forms
(Baines 2007).
In short, a minimal hypothesis that early scripts evolve in the
direction of visual simplicity is not, on the face of it, contro-
versial. What could cause this bias is not quite as evident. We
propose that it is driven by a least effort principle (inspired by
Zipf 1949). This phrase is usually linked to the fact that spoken
languages tend to minimize the total amount of effort that
speakers invest in producing words by shortening frequently
pronounced words, which are also the least informative (Pian-
tadosi, Tily, and Gibson 2011; Zipf 1949). We endorse the general
notion that symbols should minimize the effort of producing
them, to the extent that this does not harm their information
content. This principle is in line with the broader view that
communicative signals face a trade-off between simplicity and
informativeness. Applied to lexicons, the principle successfully
predicts that fine-grained categorizations apply to the more sa-
lient, relevant, or frequently communicated elements, while
coarse-grained categorizations are reserved for less important
traits, in both natural languages and artificial codes evolved in the
lab (Carr et al. 2017; Regier, Kemp, and Kay 2014).
We apply this principle to writing (departing from Zipf in
doing so). Our version of the least effort principle states that
writers are motivated to reduce their cognitive and motor ex-
ertion when tracing a letter, up to the point where further
simplification would make the letter undistinguishable from
other letters. This principle does not predict a continuous prog-
ress of increasing simplification, nor does it assume that all
scripts should reach the same level of complexity over time.
Quite the contrary. Simplification should stop around the point
where it imperils legibility, making letters too difficult to dis-
tinguish from one another, and it predicts this because a de-
crease in legibility entails a potential increase in future effort.
Where this point lies is likely to differ from script to script;
letters are distinctive for various reasons in various scripts. The
very pressure to stand out as a distinctive letter would be less
intense in a small alphabet, compared to a 200-letter syllabary.
Our hypothesis, thus, is quite compatible with the fact that the
scripts used by the world’s writing systems vary considerably in
complexity. It can also accommodate the possibility that most
scripts are not currently becoming simpler.
Character complexity is driven not solely by the cognitive
and motor factors listed above but also by variable technolog-
ical factors (e.g., means of letter inscription, etc.) or historical
contingencies (weight of precedent; contact with other writing
systems). We make no specific predictions about these cul-
turally and historically contingent factors, except to stress that
they may complicate the account given here in unpredictable
ways, making the data noisier, so to speak. For this reason, our
theoretical framework does not predict that two culturally
distinct writing systems should necessarily evolve to reach the
exact same degree of letter complexity over time (even as-
suming that both systems encode the exact same language in
the exact same way). In particular, we cannot rule out the
possibility that a script’s initial complexity could cast a long
in this lightthat we should read his choice Egyptian owl hieroglyphic /m/
as the starting stimulus in a transmission chain, noting that it “may have
been the basis of the form of our letter M”(180), as derived from the M-like
beak and face. In fact, his participants transformed it into cat.
Kelly et al. The Predictable Evolution of Letter Shapes 000
shadow over its subsequent evolution, making it likely that
initially complex scripts stabilize at a higher level of complexity
than ones that were initially less complex.
This still does not tell us why letter shapes should be overly
complex from the outset. In theory, the mechanism just de-
scribed could result in a process of complexification, where
overly simple letters get enriched until they become sufficiently
distinctive. We do think that, in theory, such a process should
happen and that some actual examples may be found. Scholars
have pointed to potential complexification processes at work in
the Aztec and Mixtec scripts (Boone 1994; Downey 2014),
certain Linear A inscriptions (Steele 2017:165), and later forms
of Ogham (McManus 1996). Complexification, or inertia, may
set in when writing systems are employed primarily for sym-
bolic, aesthetic, or religious purposes or when the inscriptions
are intended to be displayed for long periods of time.
Yet we also concur with the literature just reviewed: the trend
usually goes from complex to simple. Why would this be? Here
is one possible answer. The need for letters to be distinctive is
immediately apparent to the eye of a script inventor. To esti-
mate, and then to reduce, the precise amount of effort that goes
into tracing a letter would presumably take longer. The differ-
entials are minute, becoming significant only when a letter is
repeatedly traced millions of time. It would thus make sense for
script inventors to overlook the discrete costs of letter produc-
tion and favor distinctiveness oversimplicity. A related pressure
leading to initial complexity is the fact that, from an inventor’s
point of view, ease of imagination and recall (pushing for dis-
tinctiveness) may matter more than ease of production.
In some ways, the process we are positing resembles cursiv-
ization, another well-known evolutionary trend that affects
writing systems, the Egyptian and Latin scripts being the clearest
examples (Coulmas 2003; Parkes 2008). Cursivization occurs
when a script starts being written much more frequently and
rapidly than it used to be—by copyist monks in the Latin case—
linking letters together, producing rounded shapes and liga-
tures. Cursivization may or may not produce simplified shapes
(it did in the case of Egyptian hieratic; Hegenbarth-Reichardt
and Altmann 2008). In any case, it has much in common with
simplification. It is a gradual process driven by a principle of
least motor effort that may come into conflict with letter dis-
tinctiveness; in the Latin case, cursivization produced such
ambiguous pairs as pand qor nand m. Cursivization is com-
mon, and its effects are recognizable in various unrelated scripts,
but it is neither a universal process nor one that fashions every
script into the same identical shapes.
The view that cultural change is somehow geared toward
particular states has recently been defended by cultural attrac-
tion theory (Morin 2016; Scott-Phillips, Blancke, and Heintz
2018; Sperber 1996), a school of thought working within the
broader framework of cultural evolution (Boyd and Richerson
1985; Cavalli-Sforza and Feldman 1981). Its goal is to identify
attractors, that is, states that cultural dynamics tend to evolve
toward unless they are in that state already. Attraction dynamics
are well explored experimentally (Fehér et al. 2009; Kalish,
Griffiths, and Lewandowsky 2007; Miton, Claidière, and Mer-
cier 2015). We contend that letter shapes satisfying a principle of
least effort are such an attractor: scripts should evolve toward
that state unless they are already there.
Iconicity versus Visual Complexity
If, therefore, the material shape ofwriting issubject toprocesses
of schematization, from iconic to abstract, howis such change to
be measured or predicted? The problem faces a number of
immediate challenges. First, the property of iconicity is hard to
capture and measure objectively given the cross-cultural di-
versity of figurative traditions and their context-specificper-
ceptions of verisimilitude. This inherent subjectivity of iconicity
judgments is succinctly expressed by Morphy (1991), a visual
anthropologist who has defined figurative representations as
those that are “intended to ‘look like’the object represented and
to be interpreted as such by those familiar with the iconographic
code”(52; emphasis added). Added to this is the problem that
even within a single visual tradition, there is no robust way of
determining whether a given sign is more, or less, iconic than
another (but see McDougall, Curry, and de Bruijn 1999). We
argue, however, that the tendency for graphic images in trans-
mission experiments to lose iconicity over multiple generations
is subordinate to a more fundamental transition from less com-
pressible to more compressible. That is to say that the images
(iconic or otherwise) become graphically optimized in terms of
information storage, retrieval, and reproduction by human
agents. We do not claim that graphic iconicity itself is an en-
cumbrance with no role or affordance (see, e.g., Turoman and
Styles 2017) but rather that the diachronic measurement of
iconicity is neither feasible nor desirable, let alone the mea-
surement of its presumed effects. In other words, whether an
insider observer perceives a loss or maintenance of iconicity
across the full history of a writing system is of far less relevance
to whether the writing system, both the individualitems and the
set of items, is becoming more or less compressed over time or
remaining stable.
Emergent Writing Systems and What They
Might Tell Us
Probing for compression effects in ancient primary inventions
of writing is difficult since there is no way of reconstructing the
very early phases of the development of writing as it unfolded
year by year from individual to individual.
4
In short, the sur-
viving archeological record does not permit us to observe how
incremental processes of change and variation, witnessed over
very short time spans, might scale up to larger patterns over
4. As Silvia Ferrara (2014) put it, “Incipits of all script are invisible, as
our points of departure are irrevocably not ‘the first’attestations, but the
‘earliest’that the vagaries of archaeological preservation has made
available to us”(75).
000 Current Anthropology Volume 62, Number 6, December 2021
generations, centuries, and millennia. And if we turn instead to
later and better-documented periods in the history of writing, it
becomes hard, if not impossible, to disentangle the effects of
inheritance from other processes.
5
One proposal, by no means original, is to turn instead to
secondary inventions of writing generated via stimulus diffu-
sion. These are systems created in near isolation by nonliterate
inventors who borrowed the idea of writing but did not directly
acquire its principles directly from literate teachers. Terminol-
ogy for this special class ofsecondary scripts varies; however, we
prefer “emergent scripts”(by analogy with “emergent lan-
guages”)anddefine them as functional writing systems devel-
oped by nonliterates from minimal stimulus.
6
Both Tylor (1878 [1865]) and Kroeber (1940) recognized the
potential of emergent systems to illustrate general processes
of cultural borrowing, diffusion, and change. Other scholars,
meanwhile, argued that they might also serve as heuristics for
probing the origin and evolution of writing itself across deep
timescales (Dalby 1967; Diringer 1968 [1948]; Ferguson 1995;
Gelb 1963 [1952]; Kotei 1972). Isolated reinventions of writing
serve as naturalistic transmission experiments in script change
since the nonliterate inventors taught their systems to new
generations of nonliterates who have in turn passed their knowl-
edge to subsequent generations. Writing has been reinvented un-
der these conditions at least seven times in recent history. The
best-known example is that of the Cherokee script developed by
Sequoyah in 1821. This was followed soon after by the Vai script
(Liberia, ca. 1833), the Bamum script (Cameroon, ca. 1896), the
Alaska script (1901–1905), the Caroline Islands script (1905–
1909), the Masaba script (Mali, 1930), and Pahawh Hmong
(Vietnam, 1955–1971).
Of these scripts, we contend that the Vai script of Liberia
(fig. 1) is the best candidate for historical analysis. The script
was created by at least eight men (Forbes 1851) who had prob-
ably been exposed to the Arabic and Romanscriptsbut were not
literate in either of them and certainly did not apply them as
models, unlike the Cherokee syllabary that took direct inspira-
tion from the Roman alphabet.
7
While Roman and Arabic
scripts were certainly used in West Africa by a literate minority
in the early nineteenth century, no influence is discernible in
early or later forms of the Vai letters. Indeed, a powerful indi-
cation that the inventors of Vai were not literate in the seg-
mental Roman and Arabic scripts is the fact that the Vai writing
system is a syllabary (Daniels 1992:88). Although the Vai script
was sporadically taught in local schools, its transmission has
remained largely informal and noninstitutionalized for most of
its history. Until recently, those who wished to learn the lan-
guage had to seek out a literate individual to teach them on a
voluntary basis (Bai Leesor Sherman, personal communication;
Scribner and Cole 1981). The teaching itself follows no pre-
scribed method: there is no recitation order or any common
reference syllabary, meaning that apprentices learn-by-doing
(Konrad Tuchscherer, personal communication). Thus, if cer-
tain characters are not needed in a communication, there may
never be occasion to learn them at all.
8
The fact that the number
of graphemes in the system did not substantially expand or
contract from its inception up until the present day means that
the pressure for distinguishability of letters has not appreciably
changed over its history. Vai is also a wholly secular writing
system that has always been used for ordinary communicative
purposes, such as letter writing and household administration,
as opposed to display or ritual. For all these reasons we might
expect that Vai writers have used their scripts under conditions
of relative freedom, giving variation and change the best chance
to emerge. Finally, the script has been independently docu-
mented on at least 16 separate occasions between 1834 and
2005, meaningthat almost all of its history is recorded and does
not require any interpretive reconstruction.
The Vai script remains in active use today among Vai speakers;
however, the devastations of the Liberian Civil Wars (1989–
1997, 1999–2003) and the West African Ebola crisis (2013–
2016) have had some impact on its domains of use and the
distribution of its users. Prior to these events, the script had a
more prominent role in formal communications between Vai
chiefdoms, especially in the coordination of elections for the
5. It is often noted that writing systems have a bias toward inertia. The
clearest demonstration ofthis can be seen when a writing system developed
by speakers of a given language is laterappropriated for use by speakers of a
language with dissimilar phonological or morphological rules. In many
cases, features of the system designed for the first language are retained for
the new language, despite increased costs in terms of accuracy and learn-
ability. Consider how the Minoan syllabic system was appropriated for
Mycenean Greek, despite the latter’s preponderance of consonant clusters
(Mattingly 1992), or the fact that Turkish was, for centuries, represented
in the Arabic script, which is well suited to Semitic languages, with conso-
nantal roots and inflecting vowels, but not to the phonotactics of Turkish.
However, when writing systems are actively reformed to accommodate the
phonologies of their new languages, these reforms often result in the more
precise specificationof phonological units, e.g., from Sumerian cuneiform to
the more phonetic Akkadian cuneiform or in the shift from the Phoenician
script (specifying consonants) to the Greek alphabet (specifying both con-
sonants and vowels).
6. Note that the term “emergent”is not defined historically, so there is no
categorical cutoff pointdistinguishing emergent scriptsfrom primary scripts
on the basis of periodization; indeed, several ancient scripts are argued to
have been emergent, including Linear A and Old Persian Cuneiform.
Daniels introduced the term “unsophisticated grammatogenies”(Daniels
1992:85) for such scripts, which we believe is misleading since it lends the
impression that they were naively or deficiently constructed.
7. Although unsuccessful attempts have been made to detect influ-
ences from systems as disparate as Arabic, Hebrew, and Chinese (for a
summary see Kelly 2018), the Vai script is very much a closed writing
system (Houston 2004c) with no discernible interference from other
writing traditions.
8. Momolu Massaquoi (1899) complained that before he drew up his
standard chart of Vai signs, “each man . . . began the study of the char-
acters where he liked and ended where liked, few ever mastering the whole”
(578). However, Tuchscherer’sobservationsconfirm that these efforts had
limited impact on traditional ad hoc learning methods.
Kelly et al. The Predictable Evolution of Letter Shapes 000
Traditional Council of Chiefs. During the Ebola crisis, mean-
while, the Vai script was mobilized within Cape Mount as a
medium for promoting health and quarantine regulations. Vai
program officers used the script to display warnings and di-
rectives on public placards and to distribute messages among
Vai communities. Another significant change in recent years is
the fact that the script is now being taken up bylearners who are
not ethnically Vai. This has been noticed among students at the
University of Liberia campus in Sinje, Grand Cape Mount
County.
An Analysis of Compression in the Vai Script
Drawing on manuscript sources, we proposed to measure
compression in the Vai script as it has evolved between 1834
and 2005. The Vai script has roughly 200 graphemes (inventory
sizes vary slightly depending on the source and historical time
period) representing CV syllables of the Vai language and some
logograms.
The Vai script has previously been treated as a case study
for exploring synchronic distributions of complexity in writing
systems (Rovenchak, Mačutek, and Riley 2009; Rovenchak, Riley,
and Sherman 2011). These studies attempted to find correlations
between the frequency of given graphemes and their complexity,
with inconclusive results. They relied on manual coding of overt
features within individual graphemes, as developed by Altmann
(2004) and Peust (2006). A single point, for example, is assigned a
value of 1, a straight line a value of 2, an intersecting line a value of
3, and so on. Our study, however, analyzes the graphemes in
terms of their descriptive and perimetric complexity by means of
semiautomated computational processes. Descriptive complexity
refers to how simply an image can be described in mathematical
terms, while perimetric complexity captures the size of an image’s
perimeter independent of its scale. These methods have the ad-
vantage of circumventing the potential subjectivity of hand cod-
ing and have precedents in other studies (e.g., Caldwell and Smith
2012; Miton and Morin 2019; Pelli et al. 2006; Tamariz and Kirby
2015).
Our measure of perimetric complexity is taken from Watson
(2012) since this procedure has proven effective at measuring
relatively simple images, including the graphemes of writing
systems. Descriptive complexity, meanwhile, is measured by
recording the file size of images that are compressed to a zip
format (following Tamariz and Kirby 2015).
Analysis of Historical Data Set
Our data set consists of individual Vai graphemes from 16
sources dated between 1834 and 2005, as tabulated by Tyk-
hostup and Kelly (2018). Each source is constructed so that the
set of graphemes is associated with a date (e.g., 1909) and an
author (e.g., Migeod). Tykhostup and Kelly’saimwastoprovide
a meaningful comparative data set for the history of the Vai
script such that any transformations of individual graphemes
might be tracked over 171 years. All available archival sources
were consulted in order to assemble the data set presented in
that paper; however, texts that were either undated or too short
were necessarily excluded. It should also be noted that the final
source is a Vai Unicode proposal (Everson, Riley, and Rivera
2005). Though modeled to some extent from earlier materials
Figure 1. Left, the earliest surviving record of the Vai script, dated 1834. Right, the same text written in the modern (2005) version of
the script. The left-hand image is cataloged as “New invented native alphabet of Western Africa Recd. April 18, 1834 from Messrs
Wilson and Wynkoop,”MS. Vai 1, page 1. Houghton Library, Harvard University. It is reproduced by permission of the United
Church of Christ Wider Church Ministries. The right-hand image is rendered in the Unicode standard using the Dukor font de-
veloped by Jason Glavy. (Not all 1834 graphemes are attested in 2005; spaces indicate omissions.)
000 Current Anthropology Volume 62, Number 6, December 2021
(especially Kandakai et al. 1962), we regard this latest instanti-
ation of the script as a dateable source that is no less historically
significant for being a digital document. Tykhostup and Kelly
did not have access to dated manuscripts after 2005, which is
why more recent sources are not included.
The data, while extensive in comparison to other emergent
scripts, are not exhaustive and are limited to what has survived
in the documentary record. Notable gaps include the period of
the two world wars and the Liberian civil wars (1989–2003). It is
plausible, for example, that variant graphemes branched off to
create new lineages that became extinct and were thus not
represented in surviving archives. We cannot know the degree
of compression in these hypothetical lost lineages. For the
purposes of analysis, our data set is further limited to the life
histories of only the best-attested Vai graphemes. Thus, a Vai
grapheme is included in the data set if (a) it is attested inat least
90% of the historical sources and (b) there are surviving
examples in both the earliest sources (1834 to ca. 1845) and
the most recent source (2005). This reduces our analysis to
61 graphemes (although not all 61 are attested in each source).
Such a constraint weights the sample in favor of more frequent
syllables in the Vai language since these are more likely to turn
up in documentary sources.
To ensure consistency and accuracy, each grapheme was
extracted from a digital image of the original manuscript source,
opened in Adobe Illustrator, and then virtually traced as a vector
over the exact contours of the original image, a method widely
used by epigraphers to ensure the accurate comparative analysis
of inscriptions (Parker and Rollston 2016). Further, we applied a
consistent stroke width and placed each letter within a bounded
area of 1,121 #776 pixels. Next, each grapheme was exported
in both .png and .svg formats. For all images, the foreground
color of the grapheme was white, and the background color was
black. All 2,128 processed images, including the originals, are
permanently archived in an open-access, open-data Figshare
repository.
9
As .svg images are vector based, the graphemes are
represented by 2D points that are connected via lines and curves
to form shapes. We then apply two measures of visual com-
plexity for each instantiation of the 61 graphemesacross 16 time
periods.
Perimetric complexity (henceforth, PC) measures the com-
plexity of binary pictures by taking the squared length of the
inside and outside perimeters of a grapheme P, divided by the
foreground area Aand by 4p(Pelli et al. 2006; Watson 2012):
10
Cp
P2
4pA:
Our second measure takes the vector representation of draw-
ing (.svg file) and then applies the DEFLATE compression
algorithm to remove any redundancies (following Tamariz
and Kirby 2015). This method approximates the descriptive
complexity (henceforth, DC) of each grapheme: the amount of
information required to losslessly describe a grapheme in
terms of vector paths and anchor points.
Preregistered Predictions
Our general hypothesis is that repeated episodes of transmission
amplify a bias for simplicity and that this impacts the visual
complexity of Vai graphemes. It is important to note that our
hypothesis is principally concerned with the evolution of each
individual character as opposed to the set of characters (Clai-
dière et al. 2018). As such, we ignore factors such as an increased
similarity between characters within a set and instead focus on
whether a particular grapheme becomes more or less visually
complex. In particular, we test three predictions:
Prediction 1: Visual complexity will decrease over succes-
sive generations.
Prediction 2: The complexity of graphemes with higher
visual complexity will decrease more than those with ini-
tially lower scores.
Prediction 3: Variance in complexity among characters
should decrease with successive versions of the script.
These three predictions were preregistered on the Open
Science Framework on September 9, 2017.
11
Our predictions
were meant to apply to two distinct studies: the present study,
based on historical data, and a transmission chain experiment.
The results of the transmission chain experiment will be pub-
lished separately. This paper also reports on follow-up analyses
that were not preregistered but were suggested by reviewers.
Does Visual Complexity Decrease
over Successive Generations?
To test our first prediction, weconstructed two distinct analyses
for the dependent variables of PC and DC. Both analyses were
performed using a mixed effect model (Bates et al. 2014), with
individual characters being considered as the basic unit of
analysis, nested inside the year of the script. For instance, the
character ‹ga›as instantiated in the 1868 version of the script
counts as a single data point. Year is used as the solepredictor in
our model.
12
This allowed us to test whether changes in our
dependent variables are predicted by the progression of time.
We expected a negative effect for both analyses (i.e., both PC
and DC decrease over time). Last, it is likely the case that in-
dividual characters will follow different trajectories through
9. All materials can be accessed at https://figshare.com/articles/Com
parison_chart_of_Vai_script/5398537.
10. We use the Watson instantiation of PC because it controls for the
limited resolution of the visual system, unlike the measure derived from
Pelli et al. (2006).
11. The preregistered document can be consulted at https://osf.io
/8gw5a/.
12. Year was transformed so that it starts at 0 (i.e., 1834 p0). This
transformation preserves the absolute change in time (e.g., 1834 p0,
1845 p11, 1848 p14, etc.)
Kelly et al. The Predictable Evolution of Letter Shapes 000
time; to control for this variation, we specified character as a
random intercept (i.e., different versions of the character within
a year are free to vary on the intercept) and included a random
slope for year (which accounts for the fact that characters can
vary year on year in terms of complexity).
Figures 2 and 3 show the respective averages for PC and DC,
respectively. In both cases, we observe a general trend toward a
decrease in visual complexity, which is confirmed by the model
results for both PC (intercept: bp16.128, year: bp2.019)
and DC (intercept: bp517.196, year: bp2.100). The neg-
ative coefficients in these models are significant when compared
to null models (i.e., without year as a predictor).
13
The significant
effect of year tells us that this negative trend is robust, even when
accounting for individual variation in changes to character
complexity over time. To illustrate, if one character happens to
increase in complexity over time, then the model controls for
this fact in determining the overall effect of year. Conversely, the
model also accounts for any steep decreases in complexity for
individual characters, which means that our results are not
simply an artifact of a single character disproportionately in-
fluencing the average trend.
Do Graphemes with Higher Visual Complexity
Decrease More than Graphemes with Initially
Lower Complexity?
Simplifying effects are predicted to be dependent on the initial
complexity of the grapheme. Specifically, our claim is that
graphemes with initially high visual complexity are more likely
to decrease than graphemes that start with low complexity, the
rationale being that low-complexity graphemes are already ex-
tremely optimized and unlikely to deviate much in terms of
complexity, as observed in transmission chain studies, where
compressed items stabilize at a certain point of optimization
(Tamariz and Kirby 2015:182). Conversely, graphemes with
initially high complexity are not optimized for simplicity and are
thus predicted to decrease over time (assuming a directional
effect for simplicity). To test this prediction, two ordinal ranks
were constructed for both PC and DC, based on their respective
scores for 1834. This provides us with a baseline for the initial
grapheme complexity. The ranks go from 0 (lowest score) to 37
(highest score) for PC and from 0 (lowest score) to 31 (highest
score) for DC.
14
Ordinal measures were used for both year and
rank to mitigate differences in scale causing convergence errors.
As with the first prediction, we conducted separate mixed effect
regression analyses for PC and DC but have now added rank as
an interacting predictor with year.
Figure 2. Mean perimetric complexity from 1834 to 2005. Error bars correspond to 95% confidence intervals (bootstrapped).
13. To assess the fit of the models, we used Akaike information criterion
(AIC) andBayesian information criterion (BIC).AIC and BIC are estimators
of model fit, taking into account the number of parameters for a given set of
data (i.e., the trade-off between a parsimonious model and an accurate
model). The main difference between the two is thatBIC penalizes violations
of parsimony more stringently than AIC (see tables 1, 2).
14. For DC some of the characters share exactly the same file size. As
such, these characters receive the same rank.
000 Current Anthropology Volume 62, Number 6, December 2021
It seems that adding initial grapheme complexity (complexity
rank) as an interacting predictor does improve the model fitfor
both PC and DC when compared to a model with only year as a
predictor. Furthermore, the interaction between year and initial
grapheme complexity for PC (intercept: bp9.95, year #
complexity rank: bp2.01) and DC (intercept: bp450.91,
year #complexity rank: bp2.20) indicates that later years
and higher initial grapheme complexity correspond to decreases
in PC and DC (figs. 4, 5). This supports our general hypothesis
that differences in visual complexity are important in deter-
mining the change in complexity over time: graphemes with
initially higher complexity are likely to follow a steeper decrease
over time than graphemes with initially lower complexity.
Does Variance in Complexity among Graphemes
Decrease with Successive Versions of the Script?
To test for a decrease in variance in complexity with successive
versions of the script, we used a Levene’stestforhomogeneityof
variance (1) on all samples from 1834 to 2005 and (2) between
the first (1834) and the last (2005) available data sets. On all
samples, the Levene’s test indicated unequal variances for both
DC (F
14, 1,048
p1.83, P!.05) and PC (F
14, 1,048
p2.21, P!.01). In
testing for a decrease in variance between our first and last
available data sets, we observed that variance was higher in 1834
than in 2005 for both DC and PC (figs. 4, 5). This decrease is
significant for DC (F
1, 97
p4.09, P!.05) but not for PC (F
1, 97
p
2.66, Pp.11), as shown in table 3.
All three of our predictions were confirmed by our results.
Visual complexity decreased over successive generations, and
graphemes that started out with higher visual complexity de-
creased more than those that began with lower visual com-
plexity, leading to a lower variance in complexity over time. We
now turn to the context and potential causes of those results.
The Context of Compression
Our study reveals that historical changes in the Vai script are far
from random. Instead, Vai letter forms undergo a process of
compression that can be detected from the earliest years after
the script’s invention right up until the twenty-first century. As
we have seen, most studies explain compression effects in terms
of transmission dynamics. As cultural items are transmitted
from one user to another, they pass through a memory and
learning bottleneck (Bartlett 1995 [1932]; Mesoudi and Whiten
2004; Tamariz and Kirby 2015). The most learnable charac-
teristics are retained, while features that are harder to recall or
reproduce will tend to fall away. Later generations of learners
inherit a refined set of items that they, in turn, may continue to
optimize until a degree of equilibrium is achieved. Equilibrium
is here understood as the result of a trade-off between simplicity
and distinctiveness, whereby the requirement that each letter be
distinctive puts a cap on the amount of simplification that a
letter can undergo. Each of these pressures applies to individual
letters, even though the pressure for distinctiveness will vary
depending onthe number of letters inthe system. In the present
case, however, the size of the grapheme inventory did not alter
Figure 3. Mean descriptive complexity from 1834 to 2005. Error bars correspond to 95% confidence intervals (bootstrapped).
Kelly et al. The Predictable Evolution of Letter Shapes 000
Figure 4. Perimetric complexity of each Vai character from 1834 to 2005. Each facet corresponds to the initial 1834 ranks in terms of perimetric complexity (i.e., 0 is the highest rank,
and 37 is the lowest rank). Characters (A to YE) are color coded.
000
Figure 5. Descriptive complexity of each Vai character from 1834 to 2005. Each facet corresponds to the initial 1834 ranks in terms of descriptive complexity (i.e., 0 is the highest rank,
and 31 is the lowest rank). Characters (A to YE) are color coded. Note that some characters started at the same level of descriptive complexity and therefore receive the same rank (e.g.,
rank 11).
000
significantly across the history of the system. Our study there-
fore suggests that the compression accountis highly plausible in
the case of the Vai script. We have shown, for example, that the
most complex letters, such as (Fig. 6), undergo the most
dramatic changes toward simplification. Meanwhile, letters that
start out as relatively simple shapes tend to remain unchanged.
Accordingly, high levels of visual complexity would appear to
act as a hindrance on the transmission of the script. Of relevance
here is the finding from Pelli et al. (2006) that PC is inversely
correlated with efficient object recognition. Thus, Vai graphemes
that have lower PC will be easier for readers to process and vice
versa. However, since our study relies on naturalistic data outside
of a controlled laboratory setting, it is well worth reflecting further
on the historical context of Vai writing to consider whether there
might be other complementary—or even contrary—pressures at
play across the life history of the script.
A Coordination Problem for Nonliterate Inventors
For nonliterates to invent a writing system from scratch is an
extraordinary cognitive achievement. Writing systems have
been described as cognitive tools for the analysis of linguistic
sound in the way that they model, and sometimes reify, pho-
nological and morphological structure (Ferguson 1995; Harris
1986; Mattingly 1987, 1992; Watson and Horowitz 2011). With-
out such tools at their disposal, nonliterate inventors must figure
out how to segment strings of spoken language into meaningful
units, lacking any means of taking notes or drawing up sound
values in a reference chart. As such, the eight men who devel-
oped the script in the 1830s would have had to rely on their
collective memories torecall the agreed-upon sounds associated
with each sign. With over 200 individual syllabograms, this
would have required a great deal of mental effort and coordi-
nation. The script development process itself can be imagined as
a kind of hypertransmission event involving intensive interac-
tions, reproductions, conscious memorizations, and negotiated
adjustments. It is likely that the inventors exploited cognitive
shortcuts wherever possible. Perhaps the most fundamental was
their decision—deliberate or otherwise—to model single syl-
lables as opposed to either segments or multisyllabic words. The
fact that nonliterate inventors are inclined to invent syllable-
centric systems is well established (Daniels 1996), a phenome-
non that reflects the fact that syllables have a denser psycho-
logical reality than phonemes (Port 2007; Treiman and Tincoff
1997). Other properties of the Vai writing system as it is found
in the earliest records also appear to be purpose built for learn-
ability. Certain characters, for example, have been modeled on
traditional symbols already in use by Vai speakers, while others
exploit iconicity to reference the morphemic values of a given
syllable. Examples are GUN ‹bu›(representing shot pellets),
WATER ‹ʤi›,andMONEY ‹kpɛ›. Importantly, these
signs served double duty as both logograms (signs standing for
whole words or morphemes) and syllabograms (signs standing
for syllable sounds). In the earliest decades of the Vai script, at
least 21 such logograms are attested.
15
Perhaps these logograms,
with their iconic relationship to meanings, were simply easier to
invent and recall. Moreover, their relative iconicity may have
demanded a greater attention to form, resulting in higher visual
complexity. Thus, in the early history of the script, a strategy of
logography coupled with (visually complex) iconicity may have
promoted the learnability of the script, especially among those
who were already well versed in Vai figurative traditions. As the
system developed over time, however, the logograms fell out of
use, such that only three were still attested by 1981. In other
words, Vai writers discarded their logographic scaffolding, al-
lowing phoneticism to take over. This in itself would amount to a
significant compression process that cannot be captured by
measuring visual complexity alone.
Institutions and Standardization
At least three attempts have been made to standardize the Vai
script. The first was made in 1899 by Momulu Massaquoi, a de-
scendent of the Vai script’s principle inventor and a prominent
15. Drawing on progressivist models, a number of scholars speculated
that an ideographic or pictographic system must have prefigured the
development of the phonetic Vai script (Diringer 1968 [1948]; Kotei
1972; Massaquoi 1911).
Table 1. Model comparisons for perimetric complexity (PC)
and descriptive complexity (DC) using Akaike information
criterion (AIC) and Bayesian information criterion (BIC)
AIC BIC Pr(1x
2
)
PC:
Null model 5,596.4 5,621.2
Year 5,574.5 5,604.3 1.013e-06***
!.001***
DC:
Null model 9,808.9 9,833.7
Year 9,800.7 9,830.5 .00138**
!.01**
Note. A null model, e.g., PC ∼11(1 1year Fcharacter), was compared
to a model where year was entered as a predictor, e.g., PC ∼year 1(1 1
year Fcharacter).
Table 2. Model comparisons for descriptive complexity (DC)
and perimetric complexity (PC) using Akaike information
criterion (AIC) and Bayesian information criterion (BIC)
AIC BIC Pr(1x
2
)
DC:
Year only 6,383.7 6,410.9
Year #complexity rank 6,361.2 6,397.5 1.817e-06***
!.001***
PC:
Year only 3,551.8 3,579.0
Year #complexity rank 3,524.3 3,560.6 1.445e-07***
!.001***
Note. Each comparison involved a model with year specified as the only
predictor (year only), e.g., PC ∼year 1(1 1year Fcharacter), and year
interacting with the initial complexity rank, e.g., PC ∼year #complexity
rank 1(1 1year Fcharacter).
000 Current Anthropology Volume 62, Number 6, December 2021
Liberian intellectual and official. In this year he produced a stan-
dard chart of the Vai script for use in schools (this is represented by
the point just prior to 1899 in figs. 2 and 3). Twelve years later he
published another chart that he claimed to be a direct reproduc-
tion of the 1899 script but that in fact shows numerous subtle
variations (see 1911 in figs. 2 and 3). More successful was the
standardization campaign at the University of Liberia in 1962,
involving 11 consultants and an official Standardization Com-
mittee. For the first time all variant graphemes were eliminated,
and the committee settled on one sign per syllable. The final
standardization event was the Unicode proposal of 2005 (Everson,
Riley, and Rivera 2005), resulting in the Vai script being accepted
into the Unicode Standard three years later. Those who compiled
the proposal relied on the 1962 standard together with modifi-
cations that reflected modern use. As we can see from figures2 and
3, each standardization event, in 1899, 1962, and 2005, coincided
with a decline in both DC and PC, suggesting a possible causal
relationship.
Domains and Media
The shift from a reed stylus in protoliterate Mesopotamia to a
wedge-tipped stylus in the mid-third millennium necessarily
resulted in a more compact and consistentcuneiform script, just
as the advent of paper in early second-century CE China pre-
ceded the emergence of more uniform stylesof Chinese writing.
In like manner, technological changes in writing apparatuses
may have influenced the rate of compression in the Vai script.
Vai was first written using pens cut from reeds with ink derived
from the leaves of certain species of bush. By the end of the
century, Vai writers in the interior were still using reed pens and
organic inks; however, those living on the coast who had better
access to foreign products were writing with commercial pens
and pencils on imported European paper (Delafosse 1899:30).
The Vai script was also represented on surfaces other than pa-
per. Vai men were renowned as silversmiths, and surviving
artifacts and commentaries indicate that the script was inscribed
on silver ornaments. Vai has also been carved into wooden
tablets, furniture, and lintels, a practice that continued well into
the twentieth century. Examples of the Vai script, apparently
written with a ballpoint pen, surface in the 1970s (Scribner and
Cole 1981), and by the twenty-first century Vai appeared in
digital form as a font. It is possible that a shift from more ru-
dimentary writing materials to modern pens and digital fonts
contributed to the simplification of the script. Fonts, after all,
must conform to precise dimensions and have standardized
Table 3. Variance in descriptive and perimetric complexity in 1834 and 2005
1834 2005
Descriptive complexity 3,994.98 [2,174.54–5,815.42] 1,089.64 [699.73–1,479.56]
Perimetric complexity 24.57 [13.38–35.77] 10.83 [6.95–14.70]
Note. The values in brackets indicate bootstrapped 95% confidence intervals.
Figure 6. Evolution of the complex symbol ‹ga›between 1834 and 2005.
Kelly et al. The Predictable Evolution of Letter Shapes 000
gradients for horizontal, vertical, and oblique strokes. However,
the fact that Vai continued to compress over the entire length of
the nineteenth century, at a time when there was little change in
writing media, indicates that shifts in writing technology cannot
be the full story. Even the transition from a handwritten system
to a font, between 1980 and 2005, does not show a dramatically
steeper decline in visual complexity than comparable time
periods elsewhere in the history of the script (see figs. 2, 3). Thus,
while it is a possibility that letter inscription technology has
influenced the compression of the script, this cannot explain our
results.
After its invention, the Vai script gained popularity and be-
came circulated more widely among Vai speakers entering into
new contexts, domains, and genres. Accordingly, changes inthe
dominant genres of Vai writing may also influence compres-
sion. Elsewhere, religious texts, or documents glorifying royal
lineages, are often characterized by an abundance of visual de-
tail. Consider, for example, the richness of Christian illuminated
manuscripts, Islamic calligraphy, the Maya codices, or the mo-
numental temple inscriptions of Egypt. These kinds of texts are
not designed primarily to be read so much as displayed for
the admiration of literate and nonliterate alike (Ferrara 2015;
Guillaume-Pey 2016; Houston 2004a; Olivier 1981). By con-
trast, mundane or secular genres of writing often take a more
simplified form, as witnessed in the bureaucratic handwriting of
the hieratic and demotic scripts, or the Simplified Characters of
modern China first promoted by the ultra-secularist May
Fourth Movement. As it happens, the earliest recorded genres of
Vai manuscripts are quite diverse and include such documents
as ordinary business transactions, historical chronicles, and
ornamental engravings (Creswick 1868; Payne 1860). At the end
of the nineteenth century there are reports that the script was
used for letters, travel diaries, and the recording of popular
stories (Delafosse 1899:29; Massaquoi 1899:578). Only in the
twentieth century was it used in more prestigious genres such as
autobiographies, compilations of aphorisms, and translations of
canonical literature, including the Koran, the Bible, and even the
Iliad (Dalby 1967; Massaquoi 1911). However, the more prosaic
uses, such as letter writing, record keeping, the compilation of
shopping lists, and the drafting of technical plans, are also
known across the twentieth century and may have persisted
from the script’s earliest years. As Singer (1996) put it, “The
basis for the enduring popularity of Vai [writing] has been its
personal uses rather than more public ones”(594). Conse-
quently, nothing in the history of the script indicates a decisive
shift in genres of Vai writing, or changes in writing tools, that
might otherwise explain the reduction in its visual complexity.
Relevance of Findings to Other Writing Systems
Long before the proclamation of Liberia as a freed-slave com-
monwealth in 1838, Vai merchants acted as agents in the cross-
Atlantic slave trade, with networks stretching from the Pepper
Coast into the interior. Later, as the freed slaves migrating from
the United States became established as the ruling ethnicity,
the indigenous Vai were pushed to the political margins but
remained active in warfare and trade, including clandestine
slaving. The relative autonomy of the Vai-speaking people over
centuries of colonial contact may go some way to explain how
they resisted domination from literate populations but were
porous enough to permit their script to be documented at
regular intervals by outsiders. Such a delicate balance of contact
and isolation has resulted in a rich archival legacy that is not
available to the same degree for the other emergent systems in-
vented independently in the Americas, the Pacific, and South-
east Asia, not to mention elsewhere in West Africa.
These less described emergent scripts, not discussed here,
nonetheless exhibit certain similar characteristics. On the pho-
nological level, all emergent scripts that are known to us have
targeted syllables as the central unit of phonetic representation,
with varying degrees of primacy. Further, they have emphasized
the representation of consonants while underspecifying vowels
(Kelly 2018). Logosyllabograms—or signs standing for mono-
syllabic morphemes—have also featured widely in the earliest
beginnings of many emergent systems wherever the docu-
mentation has survived.
16
With the exception of the short-lived
Caroline Island Script, later generations of writers worked with
smaller and increasingly phonetic sets of graphemes, as logo-
grams were gradually abandoned or repurposed. As we have
seen, the Vai script decreased in visual complexity over the first
171 years of its history, and it remains to be seen whether this
process is also detectable within other emergent scripts. It is
nonetheless clear that the Bamum script of Cameroon and the
Pahawh Hmong script of Vietnam went through discrete phases
at the hands of their nonliterate inventors, resulting in fewer
strokes per character in later versions of the scripts. It would
appear, therefore, that emergent writing systems may well
confirm the more material aspects of the progressivist hypoth-
eses supported by Gelb, Tylor, Taylor, and others, even if their
rigid and unilinear stadialism must be abandoned. That these
thinkers also assumed that writers of simpler or more phoneme-
based scripts were intellectually and morally superior presents
another unfortunate—some might say fatal—distraction from a
much better-founded insight.
16. Historians maintain that Sequoyah first developed a logographic
system for Cherokee before launching his fully syllabary in 1821 (Walker
and Sarbaugh 1993). In the earliest years of its development in ca. 1896,
the Bamum script of Cameroon was composed of 460 logograms but
transitioned through five stages until it became a fully phonetic system of
80 signs (Dugast and Jeffreys 1950:24). The Alaska script developed between
1901 and 1905 relied first on indigenous pictograms with no phonetic values
before its inventor reformed it as a syllabary. The Caroline Islands script
(1905–1909) was phonetic from the beginning but included many iconic
logograms. Similarly, Pahawh Hmong (Vietnam, 1955–1971) became in-
creasingly precise but was phonetic from the outset, with a small set of
logograms used only for numbers and certain common words. The earliest
drafts of the Masaba script (Mali, 1930) are lost, and the surviving docu-
ments show a fully phonetic script (Kelly 2018).
000 Current Anthropology Volume 62, Number 6, December 2021
Just how well emergent systems track the trajectories of an-
cient primary inventions of writing remains to beexplored. That
the Vai script was created in a conscious act bynamed inventors
does not invalidate it as a historical heuristic. As many pa-
leographers now claim, the first writing systems are also likely to
have been developed by small groups of individual specialists in
the space of single generation (Boltz 1986:28; DeFrancis 1989:
216; Ferrara 2015; Glassner 2003; Handel 2016; Houston 2004b;
Michalowski 1994; but for a contrary view see Bottéro 1995).
There is, after all, no such thing as a natural writing system in the
same sense that we understand natural languages. All writing
necessitates a degree of deliberate engineering. It is nonetheless
remarkable that emergent writing systems appear to change
much faster than other scripts, a phenomenon recognized
by Gelb (1963 [1952]:21, 210–211), who suggested that their
short histories emulate the evolution of writing across deeper
timescales. Contrary to Gelb’s teleological approach, we suggest
that this apparent rapidity may be a consequence of the stimulus
diffusion dynamic: since their inventors already knew what
writing was capable of—including organizing populations,
asynchronous communication, trade facilitation, and status and
identity marking—this foreknowledge gave them greater im-
petus to develop and optimize their system.
Though writing was not developed to address the bureau-
cratic needs of states, it is nonetheless true that it would become
instrumental in the regulation of state activities. One charac-
teristic of early state complexes is a movement toward in-
creasing uniformity and standardization of material culture. As
Norman Yoffee has noted, the progressive standardization of
writing in Mesopotamia coincided with other incremental
uniformities in state-associated material culture, including the
standardization of calendars, ration bowls, weights and mea-
sures, legal discourse, and irrigation systems (Yoffee 2001).
Thus, if compression effects are found in ancient writing sys-
tems, these may well be subordinate to larger political con-
vergences toward forms of statewide legibility understood in its
broadest sense (Downey 2014; Scott 1998). The relative extent to
which any such individualchanges may be blind orintentionally
coordinated cannot be probed.
By contrast, the inventors and earliest generations of Vai
users belonged to smaller-scale chiefdoms on the outer margins
of a growing colonial complex; the first official standardization
efforts were not enacted until Vai populations had become in-
tegrated into the Liberian colony. Even though it has never been
deployed to fulfil the demands of mass production within a
complex state infrastructure, the Vai script has nonetheless
transitioned toward greater visual simplicity and regularity
across all phases of its history.
Limitations of the Study
One of the advantages of Vai as an object of study is the fact that
its grapheme inventory of around 200 letters has remained fairly
stable overits entire history. However, this fact also places limits
on its heuristic value. The cuneiform, Egyptian, and Chinese
systems have witnessed significant fluctuations in inventory size
over their much longer histories, with implications for visual
complexity. As Isaac Taylor (1883b) put it, “A change in any one
letter constantly produces related changes in other letters—they
have to be differentiated in order to maintain an adequate
dissemblance”(365). Thus, as grapheme inventories of writing
systems expand and contract at different points in their history,
we can expect individual graphemes to maintain or increase
their complexity in order to ensure that no two graphemes look
too alike within a limited design space. As Watson’s (2012)
analysis indicated, individual letters of the comparatively small
Roman alphabet are, on average, much simpler than those of the
much larger Chinese script. Expansions and contractions to
grapheme inventories also raise the question of system-wide
compression effects and how they might be detected.
17
System-
internal effects, in which similarities and distinctions between
letters become structurally organized within the system as a
whole, could not be investigated in our study. However, if ad-
equate concordance tables were available for sequential phases
of ancient writing systems, this would present an opportunity to
detect system-wide compression by measuring and controlling
for grapheme inventory size. Such a study would help to bring
precision to the question of how individual graphemes within a
fluctuating system must strike “compromise between reduction
of effort and distinctiveness”(McArthur 1992:341).
Conclusion
As we have seen, scholars have long drawn attention to incre-
mental processes of simplification in historical writing systems,
and some have suggested that these dynamics might also be
found in emergent systems created by nonliterates. Our study of
the emergent Vai script has detected clear compression effects
over the first 171 years of its existence, corroborating the sim-
plification hypothesis. These effects are witnessed in the fact that
visual complexity decreases over successive generations of users
and that graphemes with higher visual complexity decrease
more than graphemes with initially lower complexity. Further,
we have demonstrated that the overall breadth of variation in
Vai graphemes decreases with time. In other words, the com-
plexity rates of Vai graphemes become increasingly uniform
with respect to one another. We have argued that historical
considerations, such as standardization campaigns and changes
in writing technology and genres, cannot by themselves explain
the consistent reduction in visual complexity over the life of the
script. Following the lead of earlier paleographic theorists, we
suggest that systematic compression may well be a central dy-
namic underpinning the early evolution of writing systems.
17. Again, Taylor (1883b) preempted this question when noting that
“of an opposite tendency [to differentiation] is the principle of assimi-
lation, which often produces curious superficial resemblances among
letters belonging to the same alphabet, affecting, more especially, con-
tiguous letters such as m and n, p and q, E and F”(365).
Kelly et al. The Predictable Evolution of Letter Shapes 000
Acknowledgments
We thank Jeremiah Williams of Monrovia and Bai Leesor
Sherman (University of Liberia) of Bendu for details concerning
the present uses of the Vai script and its recent history. Olena
Tykhostup (Friedrich Schiller Universität, Jena) tabulated Vai
graphemes, and Julia Bespamyatnyk (Universität Erfurt) traced
them as .svgs files. We also thank the Houghton Library, Har-
vard, for access to rare Vai manuscripts.
Comments
Henry Ibekwe
Department of Philosophy, University of Nigeria, Nsukka,
Enugu State, Nigeria (hibekwe@gmail.com). 8 XI 20
Any attempt at the rigorous study of African indigenous scripts
must contend with three challenges: first, the firmly engendered
orthographic thought-habits resulting from the ubiquity and
prolonged usage of the Latin alphabets in the Western hemi-
sphere; second, the complex conceptual adjustments that must
be made in order to unlearn and then learn an entirely novel
orthographic philosophy; and, third, the often esoteric origins of
these indigenous scripts and the traversal of various realms of
meaning across which the symbols function, realms that may
not always be accessible to common human experience. The
publication by Kelly et al. throws these three challenges into
sharp relief. On the one hand, as prisoners of our own (Latin-
ized) orthographic experiences, a lot of thought habits need to
be broken in order to engage this study. On the other hand, the
syllabic nature of the Vai script reveals a rather uncommon
manner of thinking about (and doing) orthography. In addition,
the association of the Vai script with the Liberian house of chiefs
points to its esoteric origins.
At the heart of Kelly et al.’s paper lies the examination of the
theory that states that “through repeated interactions, a system
of signs will become compressed so that the same amount of
information is expressed with less descriptive effort.”Kelly et al.
tested this theory by using the Vai script as a case study. Their
efforts dwelled upon an aspect of script metrics: the measure-
ment of script complexity. In concluding, they maintained that
“all three of our predictions were confirmed by our results.
Visual complexity [of the Vai script] decreased over successive
generations, and graphemes that started out with higher visual
complexity decreased more than those that began with lower
visual complexity, leading to a lower variance in complexityover
time.”The authors note, however, that not all scripts seemed to
conform to the theory in question. They cited the example of the
hieratic Egyptian script, which became “far cruder in its form
than the monumental hieroglyphicscript that preceded it”while
making allowance for the fact that “most scripts are not cur-
rently becoming simpler.”Such an exception comes as a sur-
prise, because one must be able to test all existing scripts against
the stated theory if one is to assign a definitive truth value to the
theory.
At this juncture, the discourse may benefitimmenselyby
opening itself to more dialectical possibilities. In other words,
this theory (like all scientific theories) must wrestle with the
conceptual behemoth known as underdetermination. Further
examination of the nonconforming scripts requires a decision:
either to investigate holist or contrastive (Stanford 2017) forms
of underdetermination as source(s) of nonconformism or to
make appropriate modifications to the theory itself. The wider
applicability of the theory to otherindigenous scripts remains an
inviting exercise. However, the question as to why the Vai script
appears to obey the stated theory, and why the Egyptian hieratic
does not, must remain a work in progress.
The theory subsists in the web of belief, and it is to this web of
belief that we must turn when deviations from theory are ob-
served. After all, “whentheworlddoesnotliveuptoourtheory-
grounded expectations, we must give up something”(Stanford
2017:4). The investigative enterprise of verifying a theory requires
much caution, so that one does not fall for the temptation of
putting one metric on a pedestal when it is not very clear which
metric to investigate. The web of belief opens up multiple paths of
inquiry: should we further interrogate our own orthographic
thought-habits? Where should we make necessary conceptual
adjustments? When and where should we embrace nonmono-
tonic reasoning? Has a particular script maintained symbol
complexity for the sake of aesthetics and the transmission of
meaning? Do the symbols have their origins in irreducible ar-
tifacts, mentifacts, and sociofacts (Fantini and Fantini 1997:56–
59)? How should we navigate the boundary separating the
esoteric realm of meaning from the nonesoteric realm of mean-
ing? Could it be the case that the theory exhibits multiple degrees
of truth when applied to multiple indigenous scripts?
African indigenous scripts remain a vast, untapped repository
of semiotic and symbolic information, and many questions re-
main to be asked. Indeed, the study of these scripts must evolve
into a rigorous science that makes allowances for fuzzy and even
nonlogical notions that contribute to the formation of meaning.
A constructivist hermeneutic might need to be embraced that
gives room for the contextual uniqueness of each indigenous
script. It may well be imperative to focus on understanding how
specific scripts function in context rather than on the generaliz-
ability of findings. We may have to admit that measures of phe-
nomena are often subjective and that data collection methodology
may be subject to individual bias. Given the continued use of these
indigenous scripts, their continuous examination by means of
ever-rigorous phenomenological inquiry, together with tech-
niques of never-ending language learning (Mitchell et al. 2018),
remains an imperative.
Researchers may need to engage each and every indigenous
African script before seeking sweeping generalizations. Studies
in indigenous orthography remain relevant in light of the fact
that Unicode proposals have been submitted as preludes to the
000 Current Anthropology Volume 62, Number 6, December 2021
digitization of several indigenous African scripts. We may be
witnessing the dawning of the much-discussed concept of Afro-
futurism. A different orthographic future is possible when we
refuse to remain prisoners of our orthographic past. We remain
grateful to Kelly et al. for their pioneering investigations.
Andrij Rovenchak
Ivan Franko National University of Lviv, 12 Drahomanov Street,
Lviv 79005, Ukraine (andrij.rovenchak@gmail.com). 29 X 20
The paper by Kelly, Winters, Milton, and Morin reports an
interesting approach to the analysis of script complexity from a
historical perspective. The authors apply two different methods
to calculate complexities of the graphic shapes of symbols. The
study is based on manuscripts written in an indigenous West
African syllabary used for Vai, a Mande language currently
spoken in Liberia and Sierra Leone. The manuscripts cover a
rather large time span of over 170 years. The Vai syllabary has a
special role for studies of writing systems in view of several
factors, including a comparatively long-preserved written tra-
dition and the influence as a stimulus for the emergence of new
writing systems in the region in the first half of the twentieth
century (cf. de Voogt 2014; Tuchscherer and Hair 2002).
The notion of complexity, especially when dealing with
graphic images, while being rather intuitive, is not easily quan-
tifiable. The most straightforward approach applied for written
symbols is perhaps stroke count known, for example, for Chi-
nese characters. Traditional definitions of strokes, however, do
not distinguish between simple straight lines and curvier shapes.
A more elaborated method suggested by Altmann (2004) con-
sists of assigning different weights to points, straight lines, and
arcs as well as various types of connections. Rovenchak, Mačutek,
and Riley (2009) and Rovenchak, Riley, and Sherman (2011) ap-
plied this approach for the Vai syllabary.
The results for both perimetric (based on the relation of
contour lengths and the area of the symbol) and descriptive
(based on the amount of information to store character shapes
in vector form) complexities confirm several authors’initial sug-
gestions about the simplification of character shapes in time. It is
a pity that only references to the Figshare repository are given
and no illustrations appear in the article text. An example for the
perimetric complexity would be especially helpful. Unfortu-
nately, I was not able to locate any numerical data for com-
plexities calculated using the two approaches considered by the
authors except for those visualized in figures. It would be in-
teresting to compare different approaches to the complexity
definition. In particular, one may ask the following questions.
(1) Do the orderings coincide with the available definitions of
complexities (perimetric, descriptive, and the approach by
Altmann [2004])? Namely, is the order of characters sorted by
increasing/decreasing complexity the same? If not, what causes
the differences? (2) What are the correlations between peri-
metric or descriptive complexity and character frequency? Do
they significantly change with time?
The latter question is linked to the suggestion that simpler
shapes would be utilized for more frequent characters and vice
versa. So far, such tendencies were confirmed—within the
complexity definition of Altmann (2004)—for two indigenous
African writing systems, Vai syllabary and Nko alphabet, but the
correlation was rather weak compared with artificial systems
like Morse code (Rovenchak, Riley, and Sherman 2011).
Kelly et al. make a nice point mentioning two concurring
processes driving the evolution of character shapes: a balance
should be achieved between simplification of shapes, up to in-
distinguishable, and efforts to distinguish those. A good example
of such processes is reflected in i‘jām(consonantal pointing) of
the Arabic letters. The Vai syllabary, like the majority of young
writing systems, is certainly far from the point where the second
process enters into play. Mechanisms behind the simplification
were in particular mentioned by Delafosse (1899), who noted
that—in the absence of printing—shapes of characters, espe-
cially those used less frequently, undergo deformations caused
both by individual handwriting style and by memory flaws.
The authors report, among other results, significant drops in
the mean character complexity for years 1899, 1962, and 2005,
when the Vai syllabary underwent standardizations. This result,
as far as it follows from the paper, is based on their data set of
61 graphemes (attested in at least 90% of the historical sources
and found in both the earliest and the latest sources). It would
be interesting to analyze how the mean complexity changes for
the entire syllabary containing over 200 graphemes. Note that
standardizations, although they eliminate variant shapes, might
increase the set of characters. This is especially pronounced in
the standard syllabary by Massaquoi (1911), where anentire six
new series are introduced for syllables with foreign initial
consonants (Everson, Riley, and Rivera 2005).
The authors claim at least seven instances of the invention of
writing in recent times “by illiterate inventors”;ofthose,three
occurred in Africa, namely, Vai (Liberia, 1820s), Bamum
(Cameroon, 1890s), and Masaba (Mali, 1930). It is not clear,
though, whether two other Mande syllabaries from the 1930s,
Kpelle and Looma, fall into this category. The former might be
interesting along the lines of the authors’study, as at least two
lists of characters were published separated by several decades
(Lassort 1951 via Dalby 1967; Stone 1990).
Beyond those seven recent inventions, some others can be
considered, as commented in particular by Kelly (2018). For
sure, the list could be expander even further to include, in
particular, Chukchi writing by Tenevil or Tynevil (Davydova
2015; Jensen 1969), several scripts in Alaska (Schmitt 1951;
Stern 2004:124), and Apache symbols by Silas John (Basso
and Anderson 1973). The majority of those scripts, however,
are not documented sufficiently well—if documented at all—
to provide data for the complexity studies.
No doubt, the best choice for the next study of complexity
changes would be the Bamum script (Dugast and Jeffreys 1950;
Schmitt 1963). The easiest approach might be to use standardized
Kelly et al. The Predictable Evolution of Letter Shapes 000
shapes of characters contained in the Unicode proposals and the
Unicode charts themselves (cf. Everson, Riley, and Tuchscherer
2009). There are not too many examples of writing systems in
evolution available, so any chance to analyze this issue should be
implemented as carefully and comprehensively as possible.
Monica Tamariz
Department of Psychology, Heriot Watt University, Riccarton
Campus, Edinburgh EH9 4AS, Scotland (m.tamariz@hw.ac.uk). 9 XI 20
Evolution of Items and Evolution of Systems
Kelly and colleagues present a unique analysis—prefaced by an
enjoyably thorough introduction—showing an increase in
compressibility (or a reduction in complexity) in the letters of
the Vai script of Liberia over nearly two centuries. This is a
lovely real-world example of cumulative cultural evolution: the
optimization—or adaptation, or attraction—of letter shape in-
fluenced by the least effort principle. While the target article
briefly discusses evolution at the level of whole alphabets, it very
much focuses on how individual letters evolve. However, the
evolution of each letter is dependent on that of every other letter
in a script. The evolution of a set of letters therefore entails the
emergence
18
of system properties that make sense only at the
level of the whole set. This commentary highlights evolutionary
processes affecting whole scripts, rather than letters, and out-
lines a systematic taxonomy of cultural evolutionary pressures
affecting items and systems.
Writing systems can be used for artistic, decorative, and other
purposes, but their main function is communication through
reading. In this respect, letter shape is subject to three cognitive
pressures—efficiency of production, recognition, and learning—
and influenced by the linguistic environment (e.g., sound-letter