Mirror Particle is building a model of changing human behavior

Why language models fall short

Over the past year, startups promising to predict human behavior have attracted large investments. Simile raised $200 million at a $2 billion valuation, while Aaru raised $88 million at a $1 billion valuation. Humans&, an AI startup that announced a $480 million seed round at a $4.48 billion valuation in January, launched Persimmon, a product for modeling human behavior.

Mirror Particle co-founder and CEO Abhivyakti Ahuja argues that prompting or fine-tuning a language model to imitate a target audience cannot overcome how those models were trained. They have learned from hundreds of billions of examples, she says, so a small amount of fine-tuning data is like using a water pistol against Niagara Falls. The result remains tied to the past.

Ahuja’s deeper objection is that language models model written language, not the ways people perceive and navigate the world. Human behavior also draws on vision, spatial reasoning and social intelligence. For Mirror Particle, a model that misses those dimensions risks explaining people through things they themselves may not notice.

A model built around change

Mirror Particle is developing its foundation model from scratch. Ahuja describes it as a world model intended to simulate why people act as they do, and how their behavior shifts over time.

The company wants to track not just a static portrait of a person or audience, but what changes, what prompts the change and how much it affects behavior. Even the absence of change could be a useful signal.

To model audience segments, the startup combines customer information about buyers with data on current events, pop culture, social media and other sources. It treats each segment as a system that evolves as people encounter different experiences, with particular attention to observed behavior rather than survey answers.

Mirror Particle has raised angel investment and says it is close to closing its first venture round. Next week, it will take part in the Startup Battlefield 200 competition at TechCrunch Disrupt 2026 in San Francisco, running October 13–15.

From audience research to product decisions

Like competitors, Mirror Particle is initially targeting markets that already spend on this kind of research: market research, brand strategy and product development.

The company says its predictions can help explain why customers might behave a certain way now or in the future, including the motives and constraints behind a recommendation. In one early pilot, a well-known pet food brand wanted to know whether chicken, beef or vegetables would work best on its packaging. Mirror Particle concluded that the image was not the problem: the brand was perceived as mass-market and cheap, and sales would plateau until that perception changed.

A cosmetics brand could also use the system to test a more basic assumption than which ad might appeal to Gen Z: whether that audience wants the product at all. Ahuja’s example is an eyeshadow palette that might be the wrong product for the market, where blush could be a better fit.

I think that move—from choosing among campaign options to challenging the premise behind them—is the more interesting part of the pitch. But the announcement is quiet about how Mirror Particle will validate its predictions against what people later do. It describes a broad mix of data and a model of changing motives, but offers no accuracy results or other measure of predictive performance. That leaves the central distinction unresolved: whether this is a better account of human behavior, or a better way to sell advice that still needs to be tested.

Ahuja’s long-term ambition is to build a “universal layer for predicting human behavior,” moving from large population groups toward conclusions about individuals. The promise depends on getting from an audience-level model to reliable claims about a person—without confusing a plausible explanation with a forecast.