AI Writes Because Humans Wrote First. So Do We.

By Sebastian Martinez Torregrosa – May 20 th, 2026

{{IMAGE:2}}

The core claim

Large language models (LLMs) do not conjure ideas from nothing. They compress the accumulated output of human culture—language, logic, stories, metaphors, formulas—into a format that can be queried in seconds. This process mirrors how each generation learns from the work of its predecessors, only the compression happens at "machine time" rather than over centuries.

Human learning as inheritance

From the moment we pick up a pencil we are copying:

• Letters are traced from the alphabet we saw as children.
• Logic is inherited from philosophers whose texts we never cite directly.
• Mathematics builds on millennia of theorems, often without explicit attribution.
• Artistic taste is shaped by paintings, songs, jokes, and the countless micro‑nuances we absorb.

The same mechanism underlies AI training. An LLM ingests billions of tokens, identifies statistical patterns, and produces a model that can generate text that feels like the sum of those patterns. It is not a single author; it is a distilled cultural artifact.

The “culture guards”

Historically, priestly or academic classes have claimed exclusive authority over interpretation. When a new tool threatens that monopoly, the response is often framed as protection of “expertise.” In the AI debate, this manifests as accusations of plagiarism and calls for heavy regulation.

The argument is less about factual errors—both humans and models err—and more about who controls the gate. If anyone can summon the weight of civilization‑scale knowledge with a $20‑a‑month subscription, the traditional prestige tax erodes.

Why the outrage feels moral rather than technical

• Scale vs. individual capability: A single brain cannot hold a million styles simultaneously. AI can, and that efficiency is uncomfortable for those whose value derives from being the filter.
• Economic displacement: Creators who once earned millions by remixing cultural tropes now face competition from cheaper, faster AI‑generated equivalents.
• Perceived loss of originality: The myth that human art is a solitary spark ignores the reality that every work is already a recombination of prior influences.

AI’s fallibility mirrors human limits

Sebastian stresses that AI is not perfect. It produces plausible but sometimes wrong answers, reflecting the consensus patterns in its training data. Humans are equally prone to bias, misinterpretation, and institutional echo chambers. The solution is not to discard the technology but to develop guardrails—human review, provenance tracking, and diverse data sources—just as we critique books or news articles.

Model collapse and the danger of self‑reinforcement

When AI systems are trained predominantly on AI‑generated text, a feedback loop can emerge, reducing the diversity of ideas—model collapse. This mirrors how academic departments can become insular, repeatedly citing the same canon. The antidote is human‑augmented training: injecting fresh, dispersed knowledge, minority perspectives, and real‑world experience into the data pipeline.

Democratized enlightenment

The most compelling vision is a world where AI acts as a public infrastructure for cultural knowledge:

1. Rapid validation – A researcher can test a hypothesis against the entire body of scientific literature in minutes.
2. Idea amplification – An independent creator can combine niche insights from disparate fields without needing a university’s library.
3. Feedback loops – Humans correct AI outputs, feeding those corrections back into future models, creating a virtuous cycle of improvement.

In this model, expertise remains valuable, but its role shifts from gatekeeping to curation and mentorship. The barrier to entry lowers, allowing more voices to contribute to the collective conversation.

Balancing adoption and caution

Sebastian advocates a middle path:

• Adopt AI tools for their ability to surface and synthesize knowledge quickly.
• Guard against over‑reliance by maintaining human oversight, especially in high‑stakes domains.
• Regulate specific misuse cases (e.g., deepfakes, unauthorized voice cloning) without stifling the broader cultural benefits.

Conclusion

AI does not invent language, logic, or beauty; it repackages what humanity has already built. The controversy is less about theft and more about who gets to control the keys to that repository. By treating AI as a shared infrastructure rather than a proprietary weapon, we can preserve the spirit of cultural inheritance while accelerating its evolution.

---

Related reading:

• If AI Trains Mostly On AI Text, Where Does New Knowledge Come From?
• Model Collapse Explained: Why Diversity in Training Data Matters

Follow the author on Twitter for more thoughts on AI ethics and open‑source tooling.