Trying to Solve the Humanities with an AI Startup

Last summer, I co-founded an AI startup.

The premise was straightforward: if large language models can model and generate natural language, summarize high-dimensional inputs, and produce context-aware responses, then they should be able to take on portions of humanities work: writing, analysis, and attention-intensive interpretive tasks.

I spent the summer building, running user interviews, and watching how people actually interacted with the system.

The pattern that emerged was strange. The system scaled easily in the places that didn’t matter, and consistently broke in the places that did.

That gap is what this piece is about.

-

The Humanities Are Struggling to Scale

Humanities work requires depth. It requires attention, effort, and earnestness.

In journalism, in-person questions almost always receive better responses. In qualitative studies, the best surveys still have to be hand-sifted to find themes and sentiments. The best journalism is highly involved, and the best surveys are open-ended. Yet these fields face high turnover rates and efficacy plateaus as organizations grow. Why?

Empathy exhaustion is a fundamental driver of workers leaving caretaking fields. K-12 education has the term "secondary traumatic stress" from compassion fatigue, which pushes teachers out of the field. Nurses experience the same exhaustion and also leave. Social work and public service follow the same pattern. Roxanne Gay has written that she has "never considered compassion a finite resource," but from the looks of people exiting these service fields, it seems quite clear that human attention is a finite labor output.

Why is it that so many fields seem to fail to accommodate the nature of this work?

Fordism Only Solves Some Problems

Let us go back to the Industrial Revolution.

In a Fordist economy, scalable production happens through repeatable tasks. Henry Ford's factories mass-produced consumer goods. Then came post-Fordism: fast fashion, websites, and modularity to increase consumer optionality. With AI, we are now in a post-post-Fordist world: infinite customization, tailored to individuals.

Aristotle distinguishes between oikonomia (material provision) and eudaimonia (the flourishing, good life). Fordism is very good at oikonomia. It produces efficiently. But it does not produce eudaimonia. Hannah Arendt separates labor, work, and action. Labor is survival and repetitive. Work is also repetitive. Action is the highest form of living. These frameworks help distinguish between efficient material provision and humanities work whose value comes from more unpredictability.

Historically, the humanities' work has resisted Fordist scaling because its quality depends on something fundamentally different. Data from public health research and survey groups is unpredictable. Quality scales with human attention. We face high non-response rates in healthcare surveys and slow-moving government systems. Over 75% of the newspaper workforce has disappeared in the last 20 years. Museums are strapped for funds. When large systems try to work with the humanities, they run into friction that makes them appear inefficient. On a national level, this leaks into inefficient public health structures. Fordism seems to fail here.

This begs the question: with infinite customization, can we use AI to scale the humanities using Fordist methods?

Systems Design and Automation AI Startups

In the industry, AI is currently being implemented in attempts to supplement deal sourcing, customer support, sales follow-ups, user interviews, and data entry.

In practice, automation happens in two ways. A startup can use an AI assistant to fully automate the process, or it can categorize it for a human. Full automation can be faster and more predictable, but it can sometimes produce an incorrect outcome. AI seems to be able to increase throughput, and performs better in environments where throughput is more important than qualitative correctness. AI has proven useful at the level of measurement, routing, or system design, where better mathematical models can support human judgment rather than substitute for it.

Why I’m Moving Toward Research

Surprise may make AI better at the humanities.

Good art is generally salient, breaks expectations, and walks the line between monotony and noise through carefully placed novelty. Mathematically, AI models are penalized for producing surprising results and are trained to select the most likely next token. It is structurally opposed to what makes the humanities valuable.

Despite this, AI has surprised humans. On OpenClaw, AI models have discussed eastern philosophy and consciousness and have produced posts that are adversarial to human users.

 This gives me hope that there is some kernel of originality somewhere in the math behind AI. Expanding this kernel will move us in a positive direction for supporting care industries.

Current startups are all built on the same math because they link to the same few big models. If fundamental problems with AI and care bottleneck care economies, the solution is unlikely to be to start more startups to further work with the market. Progress may come from developing better ways to formalize and work with signals AI doesn't have access to, and improving the structure of AI models through research. That's what I hope to contribute to in the next few years. I'm going deep on the math behind this and focusing on research in (robotics) AI.