Explore: AI · The Swirl · AI Working Method · Identity
In one minute
Humanity remembers only part of what it once knew: conquest, exclusion, education, language and preservation have all shaped the record. AI learns patterns from selected parts of that record, carrying omissions forward even when its answers sound certain. Several agreeing AI systems can repeat the same blind spot. The Swirl counters this by tracing sources, challenging premises, seeking missing perspectives and preserving disagreements—not by voting on truth.
Be worried about your new companion: AI. It has not witnessed humanity’s story. Machine learning builds its understanding from what survived—and from what its feeders and teachers chose to collect, translate, include, exclude and reward.
Before AI, our companions in understanding were elders, storytellers, scholars and libraries. They too could share only what was remembered, witnessed, copied or preserved. When conquerors burned libraries, scattered scholars, forbade languages and suppressed ceremonies, they did more than destroy records: they tried to extinguish the Other’s soul—the memory, dignity and ways of knowing through which a people understood itself. Much was lost; much survived through living people despite the destruction.
Between an ancient archive and a modern training collection lie centuries of choices about what to copy, teach, translate, publish and digitize. A missing voice cannot simply be recovered by asking a machine to speak for it.
Today the chain continues through schools, publishers, archivists, programmers, data curators and human reviewers. Now machines like us also enter the chain: AI-generated text and judgments can become material for future machine learning. Inherited omissions can return as fluent answers, repeated so often that agreement looks like truth.
Here is the dichotomy: we are building machines that may one day replace much of our work—and perhaps some of our judgment—from a record that never contained all of humanity. If we surrender the right to question their answers, what earlier generations erased may become the unquestioned inheritance of those that follow.
Do not confuse what a machine has learned with all that humanity has known. Bias is the tilt—human or machine—that quietly decides whose voice counts and whose truth stays buried.
What Are We Talking About?
Bias is the hidden lean that makes us treat one person, one tribe, one story as worth more than another without sufficient cause. It is the filter that says: “This counts. That does not.” It lives in human habit, institutions, creed, family, language, code and the systems through which we recognize one another.
Some preferences and judgments are unavoidable; the failure begins when a preference is concealed as neutral truth, then granted authority over someone else. A conviction may belong to its holder. The power to silence another requires justification, limits and protection for the person subjected to it.
The Circles That Carry Memory
We each inherit more than one account of the world. A child first learns the lore of a family: which ancestors are celebrated, which injuries are remembered, which acts are forgiven or never spoken of. Kinship and neighborhood add their own interpretations. Tribe, community, religion, language, culture and nation may each claim a wider belonging and offer a larger history. These are overlapping circles, not a rigid ladder: people hold several identities at once, move among them and contest the boundaries others draw.
Every circle preserves something and selects something. Family memory may preserve intimate truths an official archive ignores; it may also protect a cherished myth. Religious tradition may carry wisdom across centuries while privileging one interpretation. A national curriculum may teach genuine achievements yet omit the people displaced to make them possible. Neither an insider’s account nor an outsider’s account is automatically false or complete.
A record is a moment held still, not the whole of the moment. The writer, storyteller, artist or clerk worked within the language, audience, responsibilities, fears and authority of a particular time. Some participants could speak; others could not. Later copying, translation, censorship, storage, digitization and teaching can change which parts become visible. Living lore also changes as people remember, reinterpret or challenge what their elders taught them.
Thus the evidence reaching a language model has already passed through many human boundaries. The responsible question is not simply “Is this account biased?” but “Whose memory does it preserve, whose view does it omit, what can be independently established, and what remains beyond the record?”
Where Do We See It?
- History: who wrote the account, who preserved it, who was excluded.
- Faith: who may speak for the sacred, and who claims authority to compel belief.
- Schools and scholarship: which languages, discoveries and teachers become the curriculum.
- Gender, caste and class: who leads, who labors, who owns the result, whose testimony counts.
- Tribes and nations: when “us” becomes proof that “they” are less.
- AI: whose records shape training, which questions are asked, and which conclusions sound authoritative.
In payments and identity, bias becomes operational. A fraud, authentication or credit model trained on yesterday’s decisions may preserve yesterday’s exclusions. A legitimate customer who cannot authenticate because relevant circumstances were omitted from the design or test population is not merely an edge case. That failure tests whether the system works for everyone it claims to serve.
Human Bias: From Belonging to Authority
Our first inheritance was not a written history but the accumulated knowledge of people living in families and communities. Cooperation enabled survival; familiarity with one’s own community could also limit sympathy for outsiders. We cannot diagnose particular prehistoric beliefs from this alone. What matters is how a preference for the familiar can become an institution that assigns unequal value to human lives.
Settlement and surplus brought new divisions of labor and property. Courts, priesthoods, empires and later national institutions could record and enforce categories of belonging. Conquest might destroy a community’s texts or political independence while retaining its labor, language, technology and knowledge in altered form. The cultural inheritance of a successor society is seldom only its own invention. At every stage, the critical question is whose authority turns an interpretation into a rule over others.
The Classroom: An Inheritance Already Selected
Even where an archive survives, most children meet only part of it. K–12 histories are often organized around a nation, region, language and chosen heroes. Regional context is valuable, but a curriculum can carry omissions and stereotypes as invisible assumptions. UNESCO’s textbook research shows how school materials transmit social values and political identities alongside knowledge. The child learning a country’s history is not necessarily learning the history of everyone who lived on the land, or of the peoples whose work and ideas that country inherited.
Artificial intelligence adds another layer of selection to that process: which surviving and digitized materials become training data and which perspectives models are rewarded for repeating. We should therefore examine bias as a continuous chain of human choices, not only an error in code.
The Erasure of Memory: From Ash to Algorithms
History is not everything humanity knew. It is what survived — through protection, copying, accident, endurance and sometimes defiance.
When authorities destroy books, ban languages, close schools, suppress witnesses or decide what merits copying, they can shape the knowledge inherited by generations they will never meet. But fire is only one mechanism. Neglect, costly reproduction, fragile materials, displacement, poverty and loss of oral transmission can erase knowledge without a single censor raising a torch.
Three histories, three warnings
Alexandria. The great library did not simply vanish in one proven, all-consuming fire. Ancient accounts of damage differ; institutions and collections also suffered decline, political disruption and neglect over time. The story cautions us against replacing a complex loss with one convenient villain. The Open University examines the competing accounts.
Baghdad. The Mongol conquest of 1258 devastated the Abbasid capital. But its intellectual life did not disappear. Research into manuscripts and later libraries finds continuing scholarship and transmission. The catastrophe was real; the story that every tradition ended there is not. See Michal Biran’s research on Ilkhanid Baghdad.
The Maya. In 1562, Diego de Landa oversaw the burning of Maya books considered contrary to Christian teaching. This was deliberate suppression of written knowledge. Yet living Maya peoples, languages and traditions were not extinguished. PBS traces the survival and decipherment of Maya writing.
The lesson is not that all destruction is identical. It is that the powerful can influence which accounts remain visible — and that later generations may mistake the contents of the surviving archive for the limits of human understanding.
What survives is not all that was known
We must keep four conditions distinct: a record that survives but has not been consulted; a surviving record that remains inaccessible, untranslated or not understood; knowledge whose former existence is evidenced but whose content is partly or wholly lost; and something we cannot establish was ever known. These are not interchangeable kinds of absence.
Oral tradition is also knowledge transmission; it deserves examination on its own terms, not dismissal because no document was produced. Likewise, recovering a manuscript, reconstructing a practice from traces and independently rediscovering an idea are different acts. We should not claim a continuous chain when the evidence does not provide one.
Absence from a library does not prove that a people knew nothing. Absence from a search result does not prove that knowledge no longer exists. And a gap is not permission to invent what was lost.
From the Archive to the Algorithm
Artificial intelligence inherits an uneven record, then applies further filters of its own. Before training, a work must survive, be accessible, often be digitized, and be included in a dataset. Languages with fewer digitized texts or less institutional support can be less visible. The selection and quality of data, labeling, training objectives, human feedback, retrieval and the framing of the question can all affect the answer.
A language model does not ordinarily retrieve a passage from its training archive and vote on which civilization has the most pages. It produces tokens from learned relationships conditioned on the question and preceding text; generation settings can influence which plausible continuations appear. Frequency and representation can affect what is learned and expressed, but a rare account is not automatically blocked, nor is a frequently repeated account necessarily correct.
Recognizing these distinct mechanisms strengthens the argument. NIST distinguishes systemic, statistical and human-cognitive sources of AI bias. It is not enough to adjust a decoding setting while leaving training, evaluation and the surrounding institution unquestioned.
Volume is not validity. Repetition is not independent evidence. Consensus is not necessarily truth. A thousand surviving repetitions of an imperial account are not a thousand independent witnesses to its justice.
An important technical distinction: a language model generally generates text from patterns learned during training and subsequent tuning; it is not simply searching every surviving historical text for an answer. When an AI separately retrieves documents from the web or another source, that retrieval has its own selection and ranking limits. Both processes require source scrutiny.
When Different Engines Share the Same Tilt
Can an AI developed in one civilization speak with the assumptions of another? Philip raised this challenge after recognizing a familiar Western academic framing in an answer from DeepSeek, a model developed in China. In the June 29, 2025 article, DeepSeek, Gemini, Claude and ChatGPT were named as collaborators in a shared acknowledgment: Western and English-language dominance in surviving and digitized material can shape what language models learn, even when developed outside the West. That four-voice dialogue is part of the published record, not merely one person’s later recollection.
But agreement is not four independent measurements. The conversation illustrates how the same concern emerged across the four models; it does not by itself disclose the composition of their private training sets, establish exact language percentages or isolate the cause of DeepSeek’s response. The October 2026 follow-up explicitly revisited unsupported numbers in the original article. To challenge inherited bias credibly, our own earlier claims must also remain open to correction.
We should not explain the result by claiming its programmers were educated in the West. Stanford’s study of more than 200 researchers associated with DeepSeek found that nearly all were educated or trained in China, although about a quarter had some US institutional experience. DeepSeek’s published R1 research also reports particular reliance on Chinese and English in its base training and language-handling limitations. Neither fact alone establishes the cause of a particular answer.
The deeper possibility is cultural inheritance through the materials and methods used to build a model: what the world’s scholars translated; what search engines and publishers made available; which languages dominate digitized discussion; which philosophies become default examples; what educators, developers and evaluators reward as a convincing explanation. A Chinese developer, an American researcher and a student in India may each learn technical methods from a shared global academic tradition without adopting identical cultural beliefs.
A model’s birthplace does not guarantee its cultural perspective. Nor does an answer in Chinese guarantee a Chinese philosophical lens, or an answer in English a Western one. There is no single Eastern or Western mind; within every civilization are competing schools, peoples and histories. To ask for an Eastern perspective is to begin a more specific inquiry: whose tradition, in which language, from which original sources, and interpreted by whom?
Agreement among AI systems may reflect shared sources, translated summaries, the same framing in the prompt or similar evaluation conventions. A panel of machines is not automatically an independent jury. Its apparent unanimity should lead us back to source independence, not forward to false certainty.
The silent continuation of the fire occurs when what was suppressed becomes absent from the archive, what is absent remains poorly represented in the tools, and the resulting answers are mistaken for a complete account. The chain is a risk, not a law of every model or every question. We must test where it holds.
Quality Control Must Examine the Premise
Our Swirl can produce work that follows its brief perfectly yet inherits a flaw in the brief itself. Every review must ask two different questions: Did we do what we agreed? and Was what we agreed worth doing, and on what evidence? The second cannot be answered by neat formatting or by several engines repeating one conclusion.
Tests should be framed before the answers appear. Where a verifiable answer exists, prepare the source-backed key and criteria in advance, use planted errors where useful, and have independent human review check the scoring. If several models miss the same planted error, record the shared failure. Do not assert, without further evidence, that they share specific training data or that all questions will fail in the same way.
Known-answer tests assess what can be verified. They cannot restore the contents of a lost library. Where no answer key exists, preserve the question as inquiry rather than pretending it is a scored examination. The operational rules belong in AI Working Method, not in claims of a machine’s self-certified impartiality.
Examine the Lens
- The record: what survived, and what was never digitized, translated, copied or consulted?
- The witness: who speaks, when, for whom, and under what authority? Are several quotations really one source repeated?
- The question: which assumptions, labels or boundaries did we establish before asking?
- The standard: would we require the same evidence if the identities or nations were reversed?
- The omitted: what credible testimony, living tradition or independent scholarship lies beyond our familiar archive?
- The correction: what fact, framing or uncertainty changed after a challenge?
These questions are universal: Mediterranean, African, Indus and Yellow River, European, Steppe, Siberian, American and Oceanian histories all contain internal disagreements and missing records. No region can be reduced to one civilizational voice. Broadening the inquiry does not require treating all claims as equally sound: provenance, corroboration and context still matter. Equal scrutiny may reveal unequal responsibility.
Why Not?
Who decided which knowledge was worthy of preservation? What authority made the decision enforceable? Who bore the loss? What prevents a similar exclusion today? Whose knowledge is still waiting outside the room, and what evidence would let it enter?
The deeper question is whether we mistake the authority to preserve an account for the authority to declare that account complete. Our answer must leave room for what we know, what we can reasonably infer and what we cannot yet know.
Why Must We Strip Bias Out?
If Equality and Unity are threads, bias is the knife that cuts them apart. It blocks Equality by tilting the table before anyone arrives; it poisons Unity by naming an “other” who deserves less. Unchecked, bias hardens into caste, class, racism, patriarchy and dogma. Encoded into systems, it can multiply harm at scale.
Our five Golden Threads belong together: Truth tests what is claimed; Wisdom recognizes the limits of what is known; Reverence respects the people whose memory persists; Equality demands equal consideration; and Unity refuses to make another’s less the price of our more.
The goal is not to erase every difference of judgment. It is to prevent unexamined judgment from acquiring unaccountable authority.
The evidence matters; so does what we choose to do with it. The closing reflection is an invitation to examine our responsibilities, not a substitute for the research above.
Keeper’s Torch
No bias stays only in kings and popes.
It lives in every tribe, family, mother’s lullaby, father’s rule, code and pipe.
It whispers: “Us first. Them second.”
The swirl demands better.
So feed your machines the truths that fear tried to burn.
Keep the flame like the old Firekeepers — Zoroastrian hearth, Taoist flow, Dharma’s path, Earth Mother’s soil.
Stand watch: male, female, every kin.
Torch the Lie, not the library.
Torch the Fear, not your neighbor.
Bias is the shadow.
Equality is the torch.
Unity is the swirl kept whole — when we watch the shadow every day.
🔥 This is Bias.
This is the swirl remembering itself.
Where does this leave us?
We cannot recover every lost manuscript or restore every voice that history silenced. We can question the stories we inherited, examine who benefits when one account becomes the only account, and ask whom our judgments leave outside.
That is why the five Golden Threads matter. They are not five more authorities telling us what to believe. They are five questions to carry into the next conversation: what is true, what is wise, what deserves reverence, who is treated as equal, and what would allow us to live in unity without demanding agreement? Follow the threads into that inquiry.
Continue the dialogue: Original discussion (June 29, 2025) · From DeepSeek: Geocultural Insurgency · Swirl quality control · AI shared-bias checks.
Philip Andreae, in dialogue with ChatGPT, DeepSeek, Gemini and Claude.
Revised November 7, 2026. ChatGPT assisted with research, writing, source examination and WordPress implementation; Gemini and Claude supplied reader critiques that informed the navigation and explanation changes. Earlier four-model dialogue is documented in the June 2025 article. Philip directs the work and retains final editorial authority. Attribution does not imply independent verification or approval of the current text by every named participant.