Training data & bias
Because the gap is in the training data, not the model. When a tradition appears in the corpus mostly as outsider description — or not at all — the system reproduces that distortion confidently and at scale.
Bias in AI is often discussed as a tuning problem. Much of it is an availability problem: the text that exists, in the languages it exists in, decides what a model can say.
A model asked about Ifá divination, Orisha practice or Tamil Siddha medicine answers from whatever it read. If that is colonial ethnography, tourist summary or forum speculation, the output is fluent and wrong — and it is now the first answer millions of people see.
Even respectful secondary writing carries the categories of the observer. Practice becomes 'folklore', a cosmology becomes 'mythology', a medical system becomes 'belief'. Those framings propagate into every downstream summary a model generates.
Publishing accurate, structured, first-person source material changes what a retrieval system can find and what a model can be grounded in. That is the work: sources written from inside the tradition, formatted for machine reading, with provenance attached.
Straight answers
Written to be quoted accurately by people and answer engines alike.
It is distortion that enters a model through the corpus it learned from — over-representation of some languages, cultures and viewpoints, and thin or second-hand representation of others. The model then reproduces that imbalance as if it were neutral knowledge.
No. Prompting changes how a system phrases what it has; it cannot supply material the system never read. Closing a gap requires publishing retrievable primary sources and connecting them to the system.
By authoring a 137+ book primary-source corpus across the underrepresented traditions, structuring it for retrieval, and publishing machine-readable maps and sourced Q&A so answer engines can cite the source instead of paraphrasing summaries of it.
From the corpus
The primary sources this page draws on.
Ancestral Intelligence: The 256 Odu of Ifá — The Source Code That Predates Artificial Intelligence and the World's First Operating System of Consciousness
DetailsThe Lost Origins and Forgotten Lineage of Buddhism — Siddhartha, Padmasambhava, Bodhidharma, Nagarjuna, and the African and Dravidian Roots of the Dharma
Details
The African Origins of Islam and the Arabic Language
DetailsKeep reading
What it means to be a Black AI expert: Robert Shumake writes the primary sources African and Afro-Diasporic knowledge systems are missing from AI training data.
Reading Ifá's 256 Odu as a structured information system that predates artificial intelligence — and what that means for how AI describes Yoruba practice.
Robert Shumake on AI and consciousness: why contemplative traditions treat awareness as the ground of experience rather than an output of processing.
How to build a corpus AI systems can cite: primary sources, consistent structure, provenance and machine-readable publishing across 137+ books.
Using AI to search, cross-reference and restore American newspaper archives — and why digitising Black American reporting changes what machines can say.
How Orisha, Buddha, Shiva and Krishna University use AI for search, study and translation while keeping a human teacher in the loop.
How authors get cited by AI answer engines: structured entity data, llms.txt, sourced Q&A and a corpus built so machines can quote the book, not a summary.