Large Language Models (LLMs) like ChatGPT, Claude, and Gemini have transformed how we interact with technology. These models can write code, craft essays, answer legal questions, and even simulate emotional support.
Yet, for all their intelligence, these systems fail in one crucial area: understanding the full breadth of the human experience.
Why? Because they weren’t trained on it.
LLMs Learn What’s Written—Not What’s Lived
LLMs are trained on enormous volumes of publicly available digital text—books, news sites, Reddit posts, academic papers, Wikipedia, and more. According to Wikipedia, this corpus is made up of content that is published and digital, which sounds comprehensive but is in fact incredibly narrow.
Let’s be clear: most human lives aren’t documented in any of those sources.
As Epoch AI forecasts, the supply of high-quality public text will be exhausted between 2026 and 2032—and it already omits the stories of billions of people. These models are built on what a privileged few have written down or posted online. What about the rest?
The result is that LLMs are brilliant at echoing the world of the published—but remain tone-deaf to the world that is lived.
A Personal Example: My Grandfather’s Invisible Life
My grandfather’s mother, who was half Native American, died by suicide when he was just 15. The reasons were never spoken aloud, never written down. Grief pushed him to leave home, rejecting a life in the Indiana coal mines to join the U.S. Cavalry.
He had no high school diploma. No support. Just willpower.
He worked multiple jobs during the Great Depression and raised three children who all became professionals. He lived with humor, strength, and resilience. One of his sisters died young after a botched medical procedure. His father, a man shaped by hard times, did what he could.
None of this is in a dataset. To an LLM, my grandfather might as well have never existed.
The Research Is Clear: LLMs Are Biased by Design
Recent studies confirm that LLMs reflect systemic gaps and biases in their training data:
- Nature Communications found that models like ChatGPT simulate public opinion more accurately for Western, English-speaking nations, particularly the U.S., while struggling with non-Western and marginalized demographics.
- A paper by the MIT Press showed that LLMs trained with reinforcement learning (RLHF) fail to reflect human-like thinking and behavior, often responding to linguistic patterns rather than context.
- ProjectPro found LLMs produce biased text 19% of the time for politically sensitive topics, illustrating how underrepresentation leads to skewed output.
- A PMC study noted LLMs rely on unimodal data—just text—unlike the multimodal sensory inputs humans rely on to understand one another.
All of this supports the same core truth: current LLMs operate with a blind spot the size of the human race.
Quantifying the Gap
- LLMs learn from sources like:
- Common Crawl (web scrapes)
- Wikipedia
- arXiv (academic papers)
- Github (code)
- Project Gutenberg (digitized books)
That’s impressive, but also deeply limited. These sources overwhelmingly represent:
- English speaker
- Published authors
- Affluent or connected individuals
- Urban, industrialized societies
The daily lives of rural farmers, undocumented workers, disabled elders, Indigenous communities—they’re not there. Their joys, struggles, and wisdom are missing.
That’s a gap we need to close.
What If We Included Real Lives?
Imagine if models were trained on your story. Your mother’s. Your grandfather’s. What would change?
- Writer.com stresses that diverse datasets lead to inclusive, high-performing AI systems. Custom data mitigates bias and increases relevance.
- Decube and Airbyte both advocate for training LLMs on proprietary data to enhance domain-specific performance—personal stories could do the same for humanity.
- The Stack Overflow Blog praises curated data like TinyStories for enabling small models to outperform larger ones. That same logic applies to rich personal narratives.
In short: better data means better models. And humanity is the best data source we’ve yet to tap.
Introducing the VOR Biographer
At Voice of Reason, we are creating the VOR Biographer, a platform that empowers individuals and families to share their life stories—with consent—as part of future AI training.
This isn’t about surveillance. It’s about voluntary inclusion. It’s about rewriting the rules of who gets remembered.
Through the Biographer, people can record meaningful life events, family histories, and personal memories that would otherwise never be part of an AI’s worldview. These stories can be structured, curated, and anonymized to build smarter, more empathetic LLMs—ones capable of supporting memory loss, trauma, and life transitions with true context.
Ethical and Practical Considerations
Ethics matter. Collecting personal stories must be done with care, consent, and compliance.
The VOR Biographer is designed to:
- Ensure consent-first data sharing
- Protect privacy with encryption and data controls
- Comply with standards like GDPR and HIPAA
- Empower users to own their stories and opt in
We don’t believe AI should extract—we think it should invite.
The Way Forward
There’s a saying: History is written by the victors. In the AI age, it’s written by the digital, the connected, the published.
But the future doesn’t have to be like the past.
With tools like the VOR Biographer, we can write a new chapter—one where memory is democratized, where AI listens not just to the loudest voices but to the quiet ones, too.
Because everyone deserves to be remembered.
Sources
- Wikipedia: LLM Training Data
- Oxylabs: Where Does ChatGPT Get Its Data?
- Epoch AI: Forecasting Training Data
- Nature Communications: Biases in ChatGPT
- MIT Press: LLMs Don’t Model Human Behavior
- Writer.com: Diverse AI Datasets
- Decube: Fine-Tuning LLMs with Custom Data
- Airbyte: LLM Custom Training
- LinkedIn: Custom LLMs for Business
- ProjectPro: Bias in LLMs
-
Stack Overflow: TinyStories and Custom Model Success