Written by Bernard Ramirez
Jack Felix did not set out to mimic people. He wanted to measure how close Artificial Intelligence had already come to achieving this goal. After his success in publishing the indie game GeoZombie, Felix created Posterum Software and launched the Posterum AI app. While giant AI companies optimize for speed and volume, Felix asks a different question. When an AI replies, does it sound like someone who has lived, erred, and felt? The answer is not a slogan. It is a metric. A metric launched by Felix in a white paper that created some industry buzz. Posterum’s Human-AI Variance Score (HAVS) compares real human survey responses with outputs from ChatGPT, Claude, Gemini, and DeepSeek across various topics, including politics, morality, economics, and more.
The app is currently in soft launch on Google Play and the Apple App Store, but the research behind it is drawing attention. A recently published white paper details 16 distinct user profiles, each of which is fed into leading models to observe how their answers shift when they take on personas with different political affiliations and demographic characteristics. The goal is not just personalization. It is alignment. How well does AI reflect the values, tone, and nuance of the person it seems to emulate?
The Measure of Likeness
Most AI evaluations focus on accuracy or fluency. Posterum’s method digs deeper. They gathered survey data from real individuals on topics from economics to moral dilemmas. Then they asked AI models the same questions using digital replicas of those users. The responses were scored not for correctness, but how closely they matched aggregate human answers to the same questions. Early findings suggest that AI can accurately replicate human behavior when taking on specific profiles, although not perfectly, and not uniformly across different categories.
This is not about passing a Turing test. It is about relevance. “We are not trying to fool anyone,” Felix says. “We want AI to respond the way a thoughtful friend would. Not just with data, but with an understanding of the human context.” The index assigns a variance score across different categories, revealing gaps where AI still falls short, even when factually correct. For users, this means fewer robotic replies, more answers that feel grounded in the personality they are trying to match.
Privacy as a Foundation
Personalization usually means surveillance. Not here. Posterum AI stores user profiles locally on the device, not the cloud. No tracking. No data harvesting. When you search, your background shapes the response, but it never leaves your phone. This design choice is not a side note. It is the core. Without it, Felix argues, true personalization is impossible. Trust collapses the moment users suspect they are being watched.
We tested this with early adopters. One user, a teacher, adjusted her profile to reflect her political views and parenting style. Queries about school policy returned answers that acknowledged trade-offs, not just facts. Another, a freelancer, received budgeting advice that took into account irregular income, something that generic tools often overlook. “It is not smarter,” the freelancer noted. “It’s just more suited to my situation.”
The Road Ahead
Felix argues that there are numerous applications for the technology. Yes, using profile input can enhance and personalize search results. But using such detailed profiles can also reduce bias that may be inherent to the artificial intelligence models used.
The Human-AI Variance Score, introduced by Posterum, can also serve as a feedback loop to help train models to be more human-like. Additionally, the score can also measure the evolution of these models over time. Posterum intends to release periodic updates to the score for subsequent releases of artificial intelligence models.
The bigger goal is to shift how we judge AI progress. Not by how fast it answers, but by how well it understands. In a field obsessed with scale, Posterum bets on depth. And in that quiet bet, a new standard may emerge. One where the best AI not only informs, but also resonates.
