I Trained an AI on My 10 Years of Emails – It Learned to Predict My Life
For years, I wondered: what patterns hide in my own digital footprint?
So, I ran an experiment. I took 10 years of my emails, carefully cleaned the data, and trained a local AI model to analyze them. The results were surprising—sometimes funny, sometimes uncomfortably accurate. Here’s what I discovered.
️ Data Preparation & Privacy Considerations
The first challenge was safely preparing my email dataset.
Steps I took:
-
Exported 10 years of Gmail archives.
-
Removed sensitive data (passwords, financial details).
-
Anonymized names where possible.
-
Stored everything on an offline local server—no cloud uploads.
✅ Example Tip:
If you’re trying this yourself, always strip sensitive attachments and never use commercial cloud services. Local-first ensures privacy.
⚙️ Training Process with Local Models
Instead of sending my personal data to a big tech company, I used LLaMA-based local models.
-
Vectorized all emails using embeddings.
-
Trained on tone, structure, and recurring themes.
-
Built a simple interface where I could ask: “What did my past emails predict about my future?”
✅ Example Prompt:
“Based on my email history, predict the kind of job I’d likely accept in the next 2 years.”
Surprising Discoveries About Personal Patterns
The AI surfaced patterns I never consciously noticed:
-
I tend to change projects every ~18 months.
-
Mondays are my most productive writing days.
-
I consistently respond fastest to emails sent before 10am.
It was like holding up a mirror to my digital behavior.
Life Predictions & Their Accuracy
When asked to predict my future, the AI said:
-
I’d likely start a new major project within the next year.
-
I’d move cities in the next 3–5 years (based on recurring “moving” conversations).
-
My communication tone shifts before big decisions—more short, blunt messages.
Some predictions felt eerily true. Others? More like educated guesses.
✍️ Communication Style Analysis
The AI also critiqued my writing:
-
70% of my emails were overly long.
-
My “urgent” tone often triggered faster replies.
-
My most persuasive emails followed a simple structure: Context → Value → Action.
✅ Example Insight:
Instead of writing “We should discuss this project sometime soon,” my successful emails said: “Let’s set up a 20-minute call this Thursday at 3pm.”
⏱️ Productivity Insights
By analyzing timestamps, the AI found:
-
My deep-focus emails usually go out between 7–9am.
-
I tend to procrastinate replies that require attachments.
-
I’m most likely to ignore emails with subject lines longer than 7 words.
These insights helped me redesign my workflow for higher efficiency.
⚖️ Ethical Considerations & Data Ownership
Training AI on personal data raises tough questions:
-
Who owns the insights—me or the AI?
-
Would I ever risk sharing this model? (No.)
-
Should employers ever use this on staff emails? (Probably not—it crosses ethical lines.)
The privacy-first approach was non-negotiable. No cloud uploads. No external servers.
️ Tutorial: How You Can Try This Safely
-
Export your email data (e.g., Gmail Takeout).
-
Clean and anonymize sensitive info.
-
Use local tools like LLaMA, GPT4All, or private embeddings.
-
Run simple analysis: productivity patterns, sentiment shifts, communication style.
-
Keep all files offline for security.
✅ Example Starter Prompt:
“Analyze my last 1,000 emails. Identify common communication habits that help or hurt my productivity.”
Final Thoughts
Training AI on my own emails didn’t just predict parts of my future—it revealed truths about my past.
It taught me:
-
My habits leave data trails.
-
Small behavior patterns can add up to big life insights.
-
AI can be both a productivity coach and a mirror—if used responsibly.
For anyone curious: this experiment showed me that personal AI isn’t about replacing humans—it’s about learning more about ourselves.