A note before we begin: I am not a professor of epistemology. I am a certified Machine Learning Engineer, a Cloud Infrastructure Architect, a part-time writer and technical trainer. I am XAI Baba — a wandering techno monk who sits at the intersection of what we know, what we think we know, and what we have never bothered to ask. I believe that every great failure of artificial intelligence is, at its root, a great failure of human self-knowledge.
Welcome to the first session. Pull up a chair. The root we are about to examine looks ordinary enough. But it has, in its time, paralyzed children and buried adults. And it is about to do it again.
Act IA Root With a Secret
Somewhere in the lowland forests of south-central Brazil, between eight and ten thousand years ago, a group of indigenous people made a discovery that would, in time, feed hundreds of millions of human beings. They found a root. Woody, starchy, defiantly ugly. It grew in poor soils where other crops refused to go. It survived drought. It survived pests.
The plant was Manihot esculenta. We know it as cassava. The Tupian people of Brazil knew it as something older and more nuanced: a gift that demanded respect.
Here is what the Tupians also knew, through ten millennia of trial, error, death, and hard-won understanding: the root is poisonous. Raw cassava contains cyanogenic glycosides — linamarin and lotaustralin. When these compounds are broken down in the human gut, they release hydrogen cyanide. The same substance used in chemical warfare. The same substance that, in the old stories, smells faintly of almonds.
The Procedure That Saved Ten Thousand Years of Lives
The process was not simple. First, the root was peeled, then grated into a wet pulp. This pulp was packed into a long, woven basket-tube device — the sebucan, or tipiti — and squeezed to drain. The expressed liquid, thick with cyanogenic compounds, was set aside. The remaining mash was then subjected to one of several techniques: fermented for 48 to 72 hours, or soaked in water for three days. Only then was the dried, detoxified material baked or roasted.
The ancestral detoxification protocol — five steps, each biochemically load-bearing. Step 3 is critical. Skip it and you keep 90% of the cyanide.
The bitter variety of cassava could produce more than fifty times the cyanogenic glycosides of the sweet variety. The proper multi-step process removed upwards of 90% of that toxicity. An improper shortcut — a soaking of only four hours, say, instead of forty-eight — could leave enough cyanide in a single meal to cause paralysis, neurological damage, or death. The communities that grew up around cassava knew this not as chemistry but as custom, as ancestral memory, as something passed from grandmother to granddaughter in the rhythm of daily life.
The root was not the problem. The lack of process was the problem. Process is what turns poison into food. This is the first lesson of XAI.
— XAI BabaAct IIThe Portuguese Arrive, and Do Not Ask Enough Questions
In the sixteenth century, Portuguese explorers arrived in the Americas and encountered cassava. They observed the indigenous populations processing it. They tasted the finished products. They praised cassava bread in their journals. They loaded it onto ships as provisions for long voyages.
But here is where the story turns. The Europeans had the output. They had the recipe. What they did not have — what they never thought to ask about, because they did not think it mattered — was the reasoning behind each step. The why. The consequence of skipping. The invisible chemical reality that made every element of the process non-negotiable.
To the European eye, shaped by Enlightenment confidence and a quiet assumption of indigenous intellectual inferiority, many of these processing steps looked like superstition. A fermentation period of three days? Surely an approximation, a cultural habit. The specific sequence of soaking, grating, pressing, airing, baking? Ritual. Ceremony.
The Europeans possessed the output-level description: "grind, press, dry, bake." They did not possess the underlying model. They could not see the cyanide being expelled. And because they could not see it, they did not believe in it enough to preserve every step with the fidelity it required.
Act IIIThe Archaeology of the Missing Data
To understand what the colonizers failed to capture, we need to understand what indigenous knowledge systems actually are. The cassava processing ritual was not written down. It was not stored in a European-style text or treatise. It existed in the hands of women who had been taught by their mothers. It was what anthropologists call tacit knowledge: knowledge so embedded in practice that the practitioners themselves could not always articulate it in propositional form. They knew how. They did not always know why in the terms a European naturalist would have recognised.
When European observers watched indigenous women process cassava, they recorded the visible steps: the grating, the pressing, the drying. What they could not easily record was the conditional logic embedded in the process. If bitter variety, then longer fermentation. If drought season, then extend soaking. If the smell is not yet right, then wait longer.
What they dismissed as indigenous ceremony was, in fact, a carefully calibrated multi-generational safety protocol. What they treated as cultural decoration was load-bearing engineering.
— XAI BabaThe Epistemological Crime That Keeps Repeating
There is a specific kind of intellectual crime that the colonial encounter committed again and again, and it is worth naming it clearly because we are still committing it: it is the crime of treating the output of a knowledge system as the totality of that knowledge system.
The cassava bread was delicious. These were the outputs. The Europeans observed the outputs, approved of them, and decided they had everything they needed. What they could not weigh, could not measure, could not codify in a European book, they tended to discount. And what they discounted had a cyanide content of up to one gram per kilogram of fresh root.
Act IVThe Black Box and the Bitter Root
Here is where I need you to hold two stories in your mind simultaneously.
In one story: intelligent, capable, well-resourced people encountered a system that produced impressive outputs. They exported the system to new populations and new contexts. They did not fully understand the internal mechanism. They could not explain why each step in the process was essential, which meant they could not adequately communicate what would happen if steps were skipped. And in the new contexts, without the tacit knowledge, the system harmed the very people it was supposed to help.
In the other story: intelligent, capable, well-resourced engineers trained a machine learning model on historical data. The model produced impressive outputs. They deployed the model to new populations and new contexts. They could not fully explain why the model produced the outputs it produced. They could not articulate which features were driving which decisions, which patterns in the training data were being generalised, which implicit assumptions were baked into the architecture. And in the new contexts, without explainability, the model harmed the very people it was supposed to help.
These are the same story.
The model knows how. It cannot tell you why. And the people the model will be deployed upon — they are the Africans who received the cassava.
— XAI BabaAct VThe Jungian Interlude
Jung's concept of the Shadow is, at its core, about the relationship between what we know about ourselves and what we do not. The Shadow is not evil, necessarily. It is simply unexamined. It is the contents of the psyche that we have not brought into consciousness. The Shadow does not go away because we ignore it. It acts anyway, from below the threshold of awareness, shaping our choices in ways we do not recognise.
Every AI system has a Shadow. A machine learning model trained on historical data does not merely learn patterns. It learns the entire texture of the world that produced that data: its inequities, its prejudices, its structural injustices. The data is not neutral. Data is the crystallised record of human decision-making, and human decision-making has never been neutral.
The Shadow of the model is everything in the data that we did not interrogate. It is the historical hiring decisions made when women were systematically excluded from technical roles, now crystallised as a preference for male applicants. It is the historical lending decisions made under redlining policies, now crystallised as a preference for certain zip codes.
Until you make the unconscious conscious, it will direct your life and you will call it fate. Until you make the model's mechanism conscious, it will direct your deployments and you will call it optimisation.
— C.G. Jung (adapted)Act VIWhat Explainability Actually Means
The field of Explainable AI — XAI, my adopted name, my chosen domain — exists to address precisely the problem I have been describing. It asks: can we build AI systems that not only produce outputs but can account for those outputs in terms that human beings can understand and interrogate?
The answer, right now, is: partially, with great difficulty, and not as often as we should.
Post-hoc explanation methods — LIME and SHAP — attempt to build interpretable explanations of specific predictions made by black-box models, after the fact. They ask: given this output, what features of the input most contributed to it? These are useful. They are also, importantly, approximations. A SHAP explanation of a deep neural network's decision is not the decision. It is a reconstruction. It is a European naturalist documenting what they saw, without the biochemistry.
Three Questions Every AI Deployment Should Answer
What are the conditions under which this system's outputs become unsafe? The cassava was safe when properly processed. It was dangerous when the processing was skipped or shortened. Every AI system has analogous conditions: edge cases, distribution shifts, demographic groups for whom the model performs differently.
Who does not have access to the explanation, and what are the consequences of that? The people most subject to the system are the people least able to protect themselves from it. When defendants cannot examine the algorithm that scored their recidivism risk, we recreate the structural asymmetry of the colonial cassava transfer.
What have we dismissed as irrelevant that we should have looked at harder? When we build AI models that optimise for a single metric and treat everything else as noise, we are doing exactly what the Portuguese did. We are taking the output and throwing away the manual.
Act VIIThe Long Aftermath, and What It Demands of Us
Nigeria is today the world's largest producer of cassava. More than 34 million tons are produced every year. The African women who received cassava without its manual did what human beings have always done: they innovated. They invented gari — toasted cassava flakes, a processing method that is distinctly African. They solved, through communal intelligence and intergenerational learning, a problem they had been handed without warning.
But konzo still exists. During drought years, when hunger presses against the edges of safety protocols, the five-hundred-year-old failure still collects its due.
The analogy to AI is not comfortable, but it is exact. The communities building AI systems are innovative, brilliant, and, in many cases, genuinely motivated by the desire to help. The outputs are often remarkable. And the harms are real, documented, concentrated in the populations least equipped to protect themselves from them, and continuing.
I want to be clear about what I am not arguing. I am not arguing that powerful AI systems should not be built. Cassava, properly processed, is a miracle: drought-resistant, capable of feeding people in circumstances where nothing else grows. The technology is not the problem.
I am arguing that every powerful system contains, within it, conditions for harm that are invisible to those who receive only the output. And I am arguing that the people building these systems have an obligation — ethical, and I would argue spiritual in the Jungian sense — to do the work of making those conditions legible.
The question before us is not whether to build powerful systems. It is whether we will do the work of ensuring that power is accompanied by understanding. Whether we will put the manual on the ship.
— XAI BabaThe cassava root sits in the ground, waiting. It does not announce itself as dangerous. It looks, to the untrained eye, like food. The danger is in what you cannot see, in the mechanism you have not yet taken the trouble to understand.
This is the first session. We have many more roots to examine together.