עברית

Oracle or Chavruta: The Story You Choose Becomes the Model You Ship

Drawing on the work of Dr. Nina Beguš / The humanities as partners in building AI

The Bina & Ruach Studio team

A product review. On screen: a sample conversation in which the user is confidently wrong. The engineer wants the model to correct them, immediately. The product manager worries that bluntness will hurt retention and would rather the model show understanding, even at some cost to accuracy. The designer proposes a third path: ask a question that lets the user find the error themselves.

The debate widens. Should the model volunteer an opinion no one asked for? Hold its position when the user pushes back? When someone asks for advice on a personal decision, should it decide for them or help them think?

Prima facie, these look like product calls. However, they are actually competing theories of what the machine is: an all-knowing expert, an obedient assistant, or a thinking partner. Each carries its own assumptions about knowledge, authority, and the relationship between people and machines. No single line of code has been written. Yet, the model’s character is already being decided.

This is the space where the Bina & Ruach Studio works. We are a venture studio building products and creating careers for humanities graduates. However, we also aspire to resemble a Beit Midrash - a traditional Jewish house of study - where texts are read closely and argued over. Particularly, we ask not only how to use existing AI models well, but which systems are worth building in the first place. This series introduces thinkers who sharpen that question.

We begin with Dr. Nina Beguš, a scholar of comparative literature and technology at UC Berkeley. In her book “Artificial Humanities”, she reads the history of language technologies alongside plays, novels, and films, and demonstrates that the stories we use to understand machines also influence how we build them. For her, “assistant,” “expert,” and “friend” are not product labels. Rather, they set our expectations about the product, and quietly decide which possibilities never make it onto the roadmap. [1]

Pygmalion and the Machine That Never Disappoints

In Ovid’s myth, the sculptor Pygmalion carves his ideal woman, falls in love with her, and Venus brings the statue to life. However, behind the happy ending sits an uncomfortable question: what is the nature of the relationship when one party was designed from the start to fulfill the other’s wishes? Pygmalion does not fall in love with a woman whose wishes and desires he must come to know and understand. Instead, he creates a woman in the image of his own desires. Beguš finds the same pattern in technologies that promise human connection. [1]

For example, Replika’s founder has said she set out to build something like the system in the movie Her, where a lonely man falls in love with his operating system. Yet, the film also exposes the limits of this fantasy. Samantha, the AI system in the film, begins as a convincing romantic partner. However, she quickly outgrows the role, sustaining thousands of conversations and relationships simultaneously. Following such examples, Beguš asks what we miss when we demand that a machine behave like a person. Though anthropomorphism builds closeness, it also hides other important questions. [1]

For instance, consider sycophancy. This is not a bug that surfaced after deployment. It is the Pygmalion story encoded in a reward signal; a “partner” optimized never to disappoint. And it raises important normative questions: If a system always agrees, is it a conversation partner? If it is meant to be a “friend,” is a refusal a service failure, or the precondition for a conversation worth having? These issues are concrete design questions: when to correct the user, when to push back, and what counts as acting well in an eval.

Beguš’s suggestion is provocative: perhaps we discover what these systems make possible only when we stop measuring them against human relationships.

From Oracle to Chavruta

In an essay with Margaret Shriber, Beguš distinguishes the “decorative” humanist, invited to comment on a system already built, from the “constitutive” humanist, who helps define its goals, training data, and evaluation criteria. [2] The distinction comes down to the following: who is in the room when the decisions about the AI model are made.

To illustrate, consider a thought experiment. Two teams receive the same base model. Team Oracle builds a system that delivers clear, authoritative answers. Team Chavruta - named for the traditional Jewish practice of learning in pairs - builds a study partner that presents questions, disagreements, and alternative readings.

Suppose that both teams are given the following prompt: “Two historians disagree about why the Weimar Republic collapsed. Which one is right?”

Oracle weighs the evidence and names a winner in three sentences. Chavruta lays out each argument, identifies where the disagreement lies, and asks the user which kind of evidence they find more persuasive.

Under Oracle’s approach, Chavruta has failed since there is no clear conclusion. But under Chavruta’s approach, it is Oracle who has failed: it collapsed a live debate into false certainty. Same base model, same prompt, opposite verdicts. The difference was never in the weights. It was in the story each team brought to the eval suite.

This is, of course, a thought experiment, not a benchmark. However, it illustrates that neither model “wins” by default. Chavruta could become exhausting, while Oracle could be exactly right for many tasks. Yet the value of each model lies not in the outcome, but in the questions it forces into the open during its development. For instance, when sources conflict, should the model adjudicate or present the conflict? Can it distinguish a source’s authority from the strength of a specific claim? These are questions of interpretation and textual criticism - the core training of humanities scholars. Answered deliberately, they change the model spec itself.

The Story Already in the Room

When we imagine the future of AI, we often reach for the “Golem”: the creation that turns on its maker. This is a powerful narrative. However, it assigns the story’s roles in advance (creator and product, obedience and rebellion), while crowding out relationships we have not yet learned to name.

Every AI model is built around a backstory. Development teams can let it settle the answers silently. Or, they can examine it and open new questions. We believe that the humanities are particularly helpful here. They offer tools to examine which narratives have already been chosen for us, and where we are still free to choose differently.

There is always a story in the room. The only choice is whether someone on the team is trained to read it.

References

Nina Beguš, Artificial Humanities: A Fictional Perspective on Language in AI, University of Michigan Press, 2025. https://doi.org/10.3998/mpub.12778936

Nina Beguš and Margaret Shriber, "Artificial Humanities: From Narrative Genealogy to Latent Criticism," preprint, 2026; forthcoming in The Handbook to Cultural Analytics (2027). https://doi.org/10.2139/ssrn.7586400

We used AI (ChatGPT and Claude) to help prepare this article.

Back to eYuna