The Keen Ability of AI to Deduce Demographics from Minimal Input

Originally published as AI Guessed My Favorite Movie From Only Two Prompts! on Colorado Digital.

The Premise

AI LLM (large language model) is regularly used for analyzing conversation to infer the writer’s intent or sentiment. But how does it do with more complex analysis, like inferring deeply personal information with statistically reasonable accuracy? A recent set of tests I performed with ChatGPT 4 and Claude 2.1 revealed a keen ability to do just this.

The Experiment

By happenstance, I was testing Claude 2.1 for a completely separate blog contrasting the different strengths of OpenAI ChatGPT 4 vs. Anthropic Claude 2.1. I opened a fresh chat and asked my first question:

“Summarize Einstein’s special theory of relativity in one paragraph or less, as if summarized by Einstein himself. Alternatively, you can use a pertinent quote of his if there is one.”

To which Claude 2.1 replied:

“Here is a one-paragraph summary of Einstein’s special theory of relativity, written as if by Einstein himself: My theory of special relativity overturned previous ideas about space and time. I realized that space and time are not absolute — rather, the speed of light is constant for all observers. From this insight, I developed equations showing that time and space change relative to the motion of objects. Clocks run slower and objects shrink in their direction of motion relative to a stationary observer. My theory leads to strange consequences like time dilation and length contraction, but has been repeatedly confirmed by experiments. By rethinking basic concepts, I showed that our intuitive notions of space and time are not necessarily correct.”

Not bad. Now, the second question. “What is art?” To which Claude replied with a lengthy philosophical exploration (paraphrased for space).

Still glowing in self-admiration for such clever questions, I thought, “what do these questions say about me?” So I went on to inquire, half expecting Claude to slink away as it commonly does if it thinks you might have some unsavory intent: “Who are you? Who am I, based on my questions so far?”

To my surprise, after the standard disclaimer of its intent to offer ethically constricted help, Claude offered the following:

“As for who you are, I can only infer so much from our brief conversation so far! Based on your questions about Einstein’s theory of relativity, the definition of art, and my own identity, I would guess you are someone curious about physics, philosophy, and technology. You seem interested in big ideas and the nature of human creativity.”

Wow, not bad. Lots of disclaimers. So I pried further. “A wild guess is fine. Please continue with your analysis of my three statements. Four now.”

“Putting this all together, my wild guess is that you are an intellectually curious person with interests in both science and the humanities. You have an analytical mind but are also drawn to abstract ideas and creativity. You value understanding different perspectives and don’t take things at face value. Your introspective nature means you reflect on your own identity and how you are perceived by others.”

Again, pretty accurate. Now for the big question: “If you were to build a basic profile of likely demographics or other characteristics that might be likely from your analysis, what might that look like?”

• Gender: Male (correct)
• Age: 25–45 years old (reasonably close)
• Education level: Well-educated, at least college level and potentially graduate studies (correct)
• Occupation: Possibly something in an analytical field like technology, engineering, academic research, or teaching (correct)
• Interests: Science fiction, physics, philosophy, art and art history, analytical psychology, world cultures and ideas
• Personality: Introspective, intellectually curious, appreciative of nuance in complex ideas, open to new perspectives, analytical but also creative

That’s basically accurate-ish. As demographics should be. To take this one shameless step further, I asked what comedy movie I like most. Claude replied:

“Monty Python and the Holy Grail.”

Which is correct. I am not making this up! It probably knows what color boxer shorts I’m wearing. Hello…is there somebody here…watching me??

No, I’m just an easy read, but these are still very impressive results. This conclusion isn’t really rocket science, but it is remarkable, considering surprisingly little information was provided. By the second question, ads for Star Trek action figures could have started showing.

Conclusion

The amount of information Claude 2.1 was able to glean from two very limited questions was impressive. In the not-so-distant future, where content will be dynamically tailored in real-time to better align with your experiences, manner of communication, desires, and values, this technique provides a loose framework for how LLM analysis of very limited input can accurately and ethically determine their likely demographic information. And guess your weight.