Mirroring Our Biases in Ancient Scripts and Modern AI

Rapa Nui's Rongorongo generates divergent scholarly paths. One prominent direction, pursued by researchers such as Albert Davletshin and Rafał Wieczorek, aims for a logosyllabic decipherment. Their work leans on methodologies like cross-reading, which requires a sign's phonetic value to be confirmed independently in multiple contexts before acceptance, and cross-referencing, which assesses proposals by converging evidence lines like sign appearance, contextual guesswork, and linguistic comparisons within Polynesia. Proposed values like ki, pa, ka, pu, hi, ma, and the word-sign MANU (bird, biped) frequently stem from this analytical toolkit, often invoking acrophony. For the layman, Acrophony is when a picture or symbol is used to represent the first sound or syllable of the word for the object shown, like using a picture of an apple to stand just for the "A" sound. Wieczorek classifies ten proposals as having substantial backing. This school of thought essentially treats Rongorongo as a language committed to wood and awaiting its linguistic key.

A starkly different viewpoint comes from the Codex Gamma paper, which sidesteps the hunt for phonetic values entirely. It frames Rongorongo as a ritual mnemonic system. Here, analysis pivots to structure, recursion, symbolic glyph roles, computational modeling of pattern recurrence (entropy resistance), and even hypothesized breath-pulse phases tied to oral traditions. The core question becomes not "What sound?" but "What memory or ritual step does this structure lock or unlock?". Meaning is sought in the architecture of the system, not in direct sound mapping.

This methodological split in Rongorongo studies runs parallel with the scholarly history surrounding Andean Khipus, or Kipu in Quechuan. Veronica Lysaght's account shows how western interpretations of these knotted cords have been heavily influenced by the researchers' cultural priors, especially regarding what counts as writing versus other communication forms. Early studies, often filtering information through inconsistent colonial Spanish sources, focused predominantly on numerical aspects, fitting the Kipu into familiar decimal frameworks. Leslie Locke's 1912 publication solidified this numerical focus, a finding that, while foundational, may have restricted the perceived possibilities of Kipu for a long time.

Later ethnographic work, engaging with modern Andean communities using Kipu-like devices mnemonically or ceremonially by scholars like Mackey, the Aschers, Urton, Salomon etc., added layers that suggested memory-aid functions linked to color or structure. Still, Lysaght argues, these studies waded the tricky waters of lo Andino: The contested notion of a static, essential Andean culture stretching back unbroken to pre-colonial times. More recently, a counter-reaction, noted by Lysaght, seems to push away from mnemonic interpretations, viewing them as insufficiently advanced.

The decipherment of Maya glyphs as a full script appears to have fueled a desire among some Andeanists to find an analogous, standardized narrative system in Kipu, sometimes leading to propositions like binary coding. Lysaght critiques this trend by suggesting it imposes a western hierarchy that privileges phonetic/alphabetic systems that may fail to appreciate mnemonic or structural systems that might operate differently but with equal purpose. Galen Brokaw likewise cautions that fixating on whether Kipu constitute writing can impose external biases and thus hinder the understanding of the medium on its own terms. The persistent hope for a Rosetta Stone underscores this preference for a legible, translatable text.

Now, this dynamic lands us squarely in current discussions about Artificial Intelligence. Yaw Ofosu-Asare speaks of a cognitive imperialism where AI, developed largely within western paradigms and trained on digitally abundant western datasets, propagates certain worldviews while marginalizing others, particularly Indigenous epistemologies. The biases observed in AI, its difficulties with Indigenous languages structured differently from Indo-European norms, its missteps in culturally specific content moderation all often stem directly from this data imbalance. For, AI models learn from what they're fed: If the digital buffet is heavily skewed towards certain languages and cultures, the resulting AI will inevitably reflect that skew. Indigenous knowledge, often oral or less digitized, gets left out or poorly understood.

This observation leads to practical considerations. Activities aimed at increasing the online presence of under-resourced languages, such as those of the Pacific Islands, directly address this data imbalance. By translating materials, digitizing records, and supporting language technologies with written or audio datasets, we contribute, albeit perhaps modestly, to a more diverse digital ecosystem. This, in turn, broadens the potential training data for future AI, and thus lessen the system's inherent western tilt simply by giving it more varied input to learn from. It nudges the digital mirror to show a slightly wider view of the digitized world.

Yet, this also surfaces a philosophical point hinted at in Ofosu-Asare's call for active integration and participatory design. If AI is primarily a reflection of existing digital data, biases and all, how much should we intervene? Is it proper to deliberately alter AI outputs to match an idealized social landscape, or does its utility partly lie in exposing the current state of digital representation, thereby highlighting societal work still needed? Should the aim be to decolonize AI by rebuilding algorithms with different epistemologies baked in, or to decolonize the data by fostering greater linguistic and cultural diversity online? The question hangs in the air: Do we adjust the mirror, or focus on changing the reality it reflects?

After all, trying to make sense of Rongorongo, Kipu, or AI bias means interpreting communication across cultural or temporal divides, and that is never simple. Our tools, our methods, our very questions are shaped by our own context. The urge for straightforward answers (the phonetic key, the definitive function, the perfectly neutral algorithm) is strong. Yet, the persistence of multiple interpretations and ongoing debates suggests the objects of our study might operate through logics that resist easy categorization within our familiar frameworks. Perhaps genuine headway comes less from finding the single correct reading and more from continuously examining the lenses through which we attempt to read in the first place. The investigation proceeds, one knot, one glyph, one algorithm at a time.