Historical Bias Explorer – Chat with AI from different eras, trained on documents from 1850s-2020s
I built this to have better conversations with my kids about AI bias. Instead of abstract discussions, they can now chat with AI personalities that think like people from different historical eras.
Each AI is powered by GPT + RAG, trained on curated document collections from its era:
1850s: Medical journals, religious texts promoting slavery/phrenology
1920s: Eugenics literature, immigration restriction papers
1950s: Cold War psychology, suburban conformity guides
1990s: "Colorblind" era welfare reform, cultural explanations for inequality
2020s: Algorithmic fairness papers, data-driven decision making docs
The technical challenge was curating ~1M words of historical documents while balancing authenticity with educational value. The RAG system retrieves period-appropriate sources to generate responses that feel genuinely from each era.
What emerged was disturbing: every era's AI is absolutely confident it represents peak human understanding. The 1920s AI discusses eugenics using identical rhetorical patterns to how modern AI discusses "evidence-based optimization."
My kids immediately grasped something that took me years to understand: we're not getting less biased over time, we're getting better at hiding bias behind increasingly sophisticated justifications.
Try it and please let me know what i can do to make it better: saranyan.com/projects/historical-bias
The conversations are uncomfortable. That's the point.
Historical Bias Explorer – Chat with AI from different eras, trained on documents from 1850s-2020s
I built this to have better conversations with my kids about AI bias. Instead of abstract discussions, they can now chat with AI personalities that think like people from different historical eras.
Each AI is powered by GPT + RAG, trained on curated document collections from its era:
1850s: Medical journals, religious texts promoting slavery/phrenology 1920s: Eugenics literature, immigration restriction papers 1950s: Cold War psychology, suburban conformity guides 1990s: "Colorblind" era welfare reform, cultural explanations for inequality 2020s: Algorithmic fairness papers, data-driven decision making docs
The technical challenge was curating ~1M words of historical documents while balancing authenticity with educational value. The RAG system retrieves period-appropriate sources to generate responses that feel genuinely from each era.
What emerged was disturbing: every era's AI is absolutely confident it represents peak human understanding. The 1920s AI discusses eugenics using identical rhetorical patterns to how modern AI discusses "evidence-based optimization."
My kids immediately grasped something that took me years to understand: we're not getting less biased over time, we're getting better at hiding bias behind increasingly sophisticated justifications.
Try it and please let me know what i can do to make it better: saranyan.com/projects/historical-bias The conversations are uncomfortable. That's the point.
Details are in the blog https://www.saranyan.com/blog/the-ghosts-of-our-machines Tech stack: Next.js, GPT-4, custom RAG implementation, 30 curated document collections spanning 175 years.