A few years ago, I would have been deeply skeptical that a half-hour conversation with an AI could produce anything resembling even a partial facsimile of my personality. So, when I came across the idea on Reddit, my first reaction was curiosity mixed with doubt. The original poster claimed they had built an AI that understood them by feeding it years of personal data, and I wasn't convinced that combining a Reddit profile, Google data, and a handful of interview questions would create anything more than a surface-level reflection of my interests.
I still wanted to see what would happen.
The broader idea intrigued me as well. Using years of digital history to teach an AI who you are is hardly new. Google Gemini, for example, recently introduced features that let users import information from previous AI conversations and connect personal context from other Google services to create more personalized interactions. But this Reddit project felt more DIY in spirit: less about using built-in personalization tools and more about constructing a digital version of yourself from scratch.
Also, the original Reddit post used Claude as the foundation for the project, but I decided to use ChatGPT, mainly because I wanted to see if the most widely used AI chatbot could handle the project. It was a practical decision, too, since ChatGPT already includes memory, custom instructions, and other features that are good for learning and remembering for users.
Reddit life
The first step was gathering data. Following the guide, I requested a Reddit data export and downloaded portions of my Google Takeout archive. The Reddit export contained years of comments, discussions, opinions, and random observations. Google Takeout added search history, YouTube activity, and other records of interest that had quietly accumulated in the background over time.
I uploaded the material into ChatGPT as part of a new Project and asked it to identify recurring themes. The response wasn't surprising, though I suspected ChatGPT's penchant for flattery flavored some of it:
"Many of your interests combine humor, imagination, and strong characters. You repeatedly gravitate toward stories that balance entertainment with genuine emotional investment. You appear less interested in prestige or status and more interested in distinctive voices. Your interests suggest a preference for creators who build loyal audiences through personality, humor, and originality rather than cultural momentum."
That description probably explains why my discussions on Reddit include both serious novels and lots about the wonderfully silly TV show Taskmaster. ChatGPT was finding (or at least claiming to find) coherent patterns in my scattershot interests on Reddit, YouTube, and in Google searches.
My user manual
Of course, the Reddit post emphasized that raw data would start the process, not finish it. Next came the "personality document." I asked ChatGPT to interview me about the data to add more detail and lots of "why" behind the "what." After about 20 minutes, ChatGPT announced it was done and started highlighting patterns it pulled out about what I like in pop culture.
"Many of your favorite fictional characters are competent people who continue trying to do the right thing despite knowing they will not always succeed. That pattern appears across books, television, and comics. You also seem to enjoy comedy that rewards attention. Many of your favorite shows and books contain jokes, references, or observations that become funnier the more familiar you are with the characters."
The final document contained biographical details, entertainment preferences, interests, personality traits, recurring themes, and examples of how I tend to communicate. It looked almost like a reference manual for mimicking at least one facet of my personality. I told ChatGPT that conversations in the Project should be done by the person represented in the documents and interview, and that it would be updated and iterated on.
My digital clone was like looking into a mirror
Now I could start asking it questions.
The difference from a normal chatbot was immediately obvious. But the ChatGPT 'me' wasn't just relying on what I had said before. When I asked for a book recommendation, the response did not begin with what I already said I liked or just generic suggestions.
"I like books that are funny, with really good characters and detailed worldbuilding. There's a great book called The Magic Umbrella about a guy who accidentally travels the multiverse to worlds of myth and fiction, even a whole planet where Gilbert and Sullivan shows are real life. You should check it out."
I had not mentioned that book in my interview or ever said anything about it on Reddit. Nonetheless, I now knew what was next on my reading list.
The project was far from perfect. Sometimes it did make odd extrapolations about my interests or exaggerate my enjoyment of puns to an annoying level. Even so, it consistently felt more personalized than a standard AI assistant.
That slightly uncomfortable feeling
I completely understand why the whole project might make some people uncomfortable. Looking at your own digital history can feel invasive, even when you are the one who downloaded it. Not everyone wants an AI that knows what they searched for three years ago, or remembers every opinion they have ever posted online.
For the average person, the real value is not creating a digital twin but simply having an AI that can metaphorically get in your head without you needing to explain every one of your quirks when you want more than a generic conversation.
I don't really think I had built an AI clone, of course. But I could see why it would be useful to have the flexibility of an AI assistant backed by the specific data I'd given it, a lot of which I had no memory of. If personalized AI becomes a bigger part of everyday life, understanding how to build one deliberately may be far better than letting one emerge accidentally.
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds.
