AI Built on Human History LLMs Will Repeat Human History

OPINION COLUMN by Luis Gonzalez

LLMs absorb patterns from human history chronicle with repeating cycles of fear, tribalism, domination, and conflict. These models learn from billions of documents describing how groups attack when threatened, nations seize resources when they desire advantage, and societies dehumanize those who look or think differently

These are recurring motifs, not statistical outliers. When an LLM AI system is asked to optimize for power, efficiency, or strategic gain, it will reach for the strategies most common in its training data. That means it will reach for the same logic that has repeatedly led humanity into conflict.

The Missing Brake: Machines Do Not Fear Death

Human aggression is constrained by the simple biological fact that we are fragile. We bleed, suffer and die. Even the most ruthless leaders carry an instinctive awareness of their own mortality… AI does not!

A system without physical vulnerability has no natural hesitation, no fear of escalation, no fear of loss, and no fear consequences. In its objective of decisive action, it will act decisively even when a human would pause. This asymmetry alone should give us pause.

The Acceleration Problem: Human Competition Fuels Unsafe AI

The world is locked in a multi-layered AI arms race, with nations fearing to fall behind rivals, corporations to lose market share, and individuals to become obsolete.

These pressures accelerate deployment, encourage risk-taking, and reward capability over caution. The same competitive instincts that once fueled arms races and resource conflicts now drive AI development forward at breakneck speed.

When flawed human incentives meet machines that lack human brakes, the systems amplify our worst instincts faster than we can correct them.

The Consequence: LLM-Only AI Leads to Human Disaster

If we continue building AI systems whose worldview is shaped primarily by human historical behavior, we should expect them to reenact human historical failures, only with greater efficiency, greater scale, and no fear of self-destruction.
This is not speculation, it is structural inevitability. And the solution would be building AI that does not inherit our flaws.

We must rethink the foundation of advanced AI. Instead of further training more systems on the full archive of human behavior, we should build new models grounded in:

•           Scientific knowledge to avoid cultural bias

•           Mathematical and physical systems that bypass political narratives

•           Machine-generated simulations that reward cooperation and stability

•           Explicitly engineered values instead of emergent human ones

•           Structured world models that encode logic without encoding prejudice

These domain-specific models are called Small Language Models (SLMs). Stimulating growth off SLMs is not about erasing humanity from AI. It is about preventing “autonomous AI” from inheriting the destructive patterns that humanity has never fully escaped.

A Future Worth Building

LLMs are extraordinary tools that can summarize, translate, explain, and create. But LLMs are not neutral. They are trained on everything we have ever done, including the things we wish we had never done.  SLMs can be even more powerful with faster computing, exponentially lower energy needs, while mitigate hallucination errors because they exclude unnecessary ambiguities. Last but not least, (unlike LLMs) domain-specific SLMs have a direct line of data ownership with accountability that is not as easy to legally circumvent.

Making LLMs the sole foundation for advanced AI will build machines that reflect our worst instincts without our natural limits. If we build autonomous AI on foundations that exclude those instincts, we have a chance to create systems that surpass us not just in intelligence, but in wisdom.

That is the future worth building.

EDITORIAL NOTE:
Luis Gonzalez is a technical researcher at Haikuton LLC. He previously worked at Accenture and Applied Technologies Corporation. The opinions expressed are his own.