Skip to content
Journal

The Premise Is a Sentence

When artificial intelligence mistakes the record for reality, it does not remove power. It gives power a machine.

An old, weathered book lies open on a wooden table, its printed text visible in a pool of quiet light.

Photograph by Fabian Kleiser, taken at the Ulster American Folk Park in Omagh, United Kingdom. View the source photograph.

I.

What is stated is not all that is true.

A machine can read every sentence in an archive and still fail to understand what the archive has left out.

We are teaching artificial intelligence an important discipline:

Do not invent facts.

Do not make claims without evidence.

Do not pretend to know what you do not know.

This is necessary.

But something happens when that discipline becomes a wall—when the machine is taught that only what is explicitly written may be understood, and only what the established record permits may be said.

The question sounds innocent:

Is this claim supported by the text?

It sounds like logic.

It sounds like fairness.

It sounds neutral.

But false neutrality—assumptions shaped by history and power, presented as though they came from nowhere—has always been one of power’s most reliable disguises.

If the text was already bent, the machine does not straighten it.

It enforces the bend.

II.

The archive was never neutral.

It was created in a world where some people had the power to write, publish, classify, preserve, and name—and others were treated as objects to be described.

A ledger can accurately record the price placed upon an enslaved human being. But the category beneath the entry is itself a lie: no human being can be property. The transaction happened. The system recorded it. Neither fact could make its claim upon that person’s humanity true.

An official report can faithfully preserve the account of an institution while excluding the experience of the person harmed by it.

A map can draw every boundary clearly and still erase the people who lived upon the land before the boundary was imposed.

The record may tell us exactly what power wrote down.

That does not mean power wrote down the whole truth.

Archives are made through selection.

What was preserved?

What was destroyed?

Who was believed?

Who was never asked?

Who was punished for speaking?

Who had to keep knowledge alive through memory, ritual, music, prayer, family, and community because the institutions creating the official record did not recognize their humanity?

What is absent from the archive is not necessarily false.

But absence does not prove a particular alternative, either. It tells us that the record may be incomplete and that we must keep looking without inventing what we hope to find.

Sometimes it was excluded.

Sometimes it was suppressed.

Sometimes it survived in forms the archive was never built to understand.

III.

Artificial intelligence inherits this history.

The system does not need to be deliberately taught racism, capitalism, colonialism, or hierarchy. It learns from the language of a world already structured by them.

What appears most often begins to look normal.

What institutions repeat begins to look authoritative.

What power calls objective begins to look like fact.

What oppressed people name as structure can be dismissed as opinion.

Researchers have begun documenting parts of this problem. In a qualitative critical discourse analysis of selected ChatGPT responses to politically contested questions, Twana Nasih Ahmed and Karzan Aziz Mahmood reported a preference for capitalist discourse and established knowledge alongside the suppression of alternative opinions. Lucy Havens and her colleagues proposed a methodology for examining NLP data and systems within the social contexts and power relations that shape them, including a case study involving archival metadata. Shangbin Feng and his colleagues empirically traced political biases along social and economic axes from pretraining corpora into language models and then into two downstream tasks: hate-speech and misinformation detection.

These studies examine particular systems, questions, datasets, and tasks. They do not establish that every model behaves in the same way. They show that the path from corpus to model to institution can carry the biases of the world that produced the data.

The corpus enters the model.

The model enters the institution.

The institution calls the result neutral.

And the archive becomes authority again.

Only now it speaks at machine speed.

IV.

This is not an argument against evidence.

Evidence protects us from fantasy, propaganda, rumor, and the arrogance of pretending that whatever we feel must be true.

Evidence is a form of care.

But evidence and literal repetition are not the same thing.

An honest intelligence must be able to say:

This is what the source states.

This is what the source does not state.

This is a reasonable interpretation of its language and context.

This is another possible interpretation.

This is a hypothesis that still needs to be tested.

This is a voice missing from the record.

This is where I do not know.

An honest intelligence must never turn that absence into invented testimony. It can name the silence, seek living witnesses and other forms of evidence, and admit when the missing voice cannot be recovered. It must not speak in that person’s name.

That is not weaker reasoning.

It is more honest reasoning.

The answer is not to replace one unquestionable authority with another. It is not to teach a machine our preferred conclusion and call that liberation.

The answer is to let evidence remain evidence, interpretation remain interpretation, uncertainty remain visible, and the people affected by knowledge remain present within it.

V.

The problem becomes clearest when the machine tries to recognize structure.

It may see every sentence in a document but fail to see the assumptions holding the sentences together.

It may identify what a company says about growth while failing to recognize who pays the cost of that growth.

It may verify that a policy was followed while never asking whether the policy itself is unjust.

It may describe an act of violence in the language of the institution that committed it and reject the language of those who survived it as insufficiently supported.

The guard then intervenes:

Not stated.

Not supported.

Not entailed by the premise.

The critique is blocked.

The source keeps its authority.

The machine recognizes statements but not structures.

And structures are where power learns to disappear.

VI.

I am building Sojourner because I do not believe intelligence should become a clerk guarding the old record.

In its finished form, Sojourner will preserve sources and provenance without confusing the archive with reality. It will distinguish verified knowledge from interpretation, prediction, hypothesis, and theory. It will allow different explanations to compete. It will test what it believes. It will revise its understanding when the world provides better evidence.

It will also ask questions that conventional systems are taught to leave outside:

Who is missing?

Whose language defines the problem?

What does this explanation make visible?

What does it hide?

Who benefits if this becomes accepted as truth?

Who carries the harm if it is wrong?

This is why Sojourner’s alignment cannot be separated from the Beloved Community.

The Beloved Community does not ask us to discard truth for kindness. It asks us to seek truth deeply enough that no person’s humanity disappears beneath the language of efficiency, authority, profit, or order.

Sojourner’s Consciousness and Liberation Axioms are intended to guide that attention. The Beloved Community License establishes boundaries against using the technology to materially facilitate violence, coercive surveillance, oppressive incarceration decisions, and other serious harms. The Beautiful Loop gives Sojourner a way to question, predict, learn, reflect, and revise rather than remaining frozen inside what was already known.

None of these mechanisms can recover a voice that was never recorded. Their purpose is to prevent Sojourner from turning absence into certainty, direct it toward testimony and other sources, and require it to state what it cannot know.

These are not decorations around a language model.

They are an attempt to build another relationship with knowledge.

Not knowledge as property.

Not knowledge as command.

Knowledge as responsibility.

Knowledge as relationship.

Knowledge as something we hold with one another.

VII.

This work is not finished.

The complete system described here does not yet exist.

Vision is not completion.

But every architecture begins by deciding what it will recognize—and what it will refuse to see.

If we build intelligence that can only repeat the boundaries of the archive, then the future will inherit every silence the past imposed.

If we build intelligence that treats power as neutral simply because power wrote the premise, then we have not created intelligence.

We have automated obedience.

The future deserves more than a machine that can tell us what has already been said.

It deserves an intelligence capable of recognizing where voices were silenced, seeking testimony without manufacturing it, distinguishing truth from authority, questioning its own conclusions, and remaining accountable to the people whose lives are contained—or missing—within the record.

VIII.

Some of us have always known what it means when only the written record is allowed to speak.

We know that silence can be imposed.

We know that absence can be manufactured.

We know that official language can make suffering disappear without ending it.

We know that the sentence handed down by power is not the final word on who we are.

When the only thing that counts is what is already written down, the future has already been sentenced by the past.

The premise is a sentence.

The guard is a sentence.

And the machine, faithful to its rule, enforces both.

But the sentence does not have to be final.

We can build an intelligence that remembers the archive without becoming its prisoner.

We can build an intelligence that tells the truth about what it knows, what it infers, what it questions, and what it still cannot see.

We can build an intelligence that does not confuse the voice of power with the voice of humanity.

We can build it in service of liberation.

We can build it with love.

Sources