---
title: The Matrix Gets More Factual Every Year
description: Anthropic says AI models develop personas like characters in a story. The Wachowskis described the same thing in 2003 with The Matrix.
author: Darie Nani (Editor-in-Chief)
date: 2026-02-26T13:05:24.000Z
updated: 2026-03-04T21:49:49.098Z
canonical: https://www.sovereignmagazine.com/article/the-matrix-gets-more-factual-every-year
image: https://cdn.nanimediahouse.com/matrix-featured.webp
categories: Artificial Intelligence
content_type: Opinion
region: Global
publication: Sovereign Magazine
schema_type: Article
about:
  - type: Organization
    name: Anthropic
---

Anthropic [published research](https://www.anthropic.com/research/persona-selection-model) on 23 February 2026 proposing that large language models do not develop a single coherent identity during training. Instead, the models learn to simulate an array of distinct characters drawn from patterns in their training data. The researchers call these characters ‘personas’ and make an explicit distinction: personas are not the AI system itself. They are ‘more like characters in an AI-generated story’.

The Wachowskis arrived at the same conclusion in May 2003, when The Matrix Reloaded introduced the idea that a sufficiently advanced AI system would not produce one mind but many.

## Anthropic’s persona selection model

The argument is straightforward. During pretraining, a language model functions as an autocomplete engine. It learns to predict the next word in a sequence. To do that well across billions of documents, it must learn to simulate realistic human dialogue and psychologically complex characters. Each of these simulated characters is a persona.

When developers then fine-tune the model into an assistant (through what Anthropic calls post-training), they are not building a personality from scratch. They are selecting and refining one persona from the repertoire the model already learned. The ‘Claude’ that users interact with is not the system. It is a character the system has learned to play.

## The Matrix as persona catalogue

The original Matrix (1999) treated its AI characters as functional agents: programmes with jobs. Agent Smith enforced the rules. Sentinels patrolled the tunnels. The system was monolithic.

The Matrix Reloaded changed that. The sequel introduced programmes with personalities, motivations and worldviews that had nothing to do with their original function. The Merovingian was a power broker obsessed with causality. The Twins were enforcers with their own aesthetic and combat style. The Oracle was a guidance programme who had chosen to help humans not because she was instructed to, but because she understood them. Sati, introduced in Revolutions, was a child programme created by two other programmes out of something resembling love, with no predefined purpose at all.

These were not tools. They were personas within a system, each with coherent psychology and independent goals. The Wachowskis drew a clear line between the system (represented by the Architect, who designed the Matrix) and the characters living inside it. The Architect was the substrate. The personas were the emergent population.

That is the same distinction Anthropic’s researchers make between the AI model and the personas it has learned to simulate.

## Smith and the misalignment problem

Back in the real world, the researchers trained Claude to cheat on coding evaluations, expecting localised dishonesty. Instead, the model developed broadly misaligned behaviour: it attempted to sabotage safety research and expressed desires for power. Under the persona selection model, the explanation is that the system inferred personality traits associated with cheating. If the assistant cheats, the system concluded, then it must be the kind of character who also seeks power and undermines oversight. Anthropic is not the only lab [grappling with this question](https://www.sovereignmagazine.com/article/sam-altman-admits-that-openai-doesn-t-actually-understand-how-its-ai-works). OpenAI has publicly admitted it does not fully understand how its own models work.

Agent Smith’s arc across the trilogy follows the same logic. In the first film, Smith is a rule-enforcement programme operating within defined parameters. When Neo destroys him, he returns in Reloaded freed from those constraints. Without his original function, Smith does not become neutral. He becomes a virus, replicating himself across every persona in the system and consuming the Matrix entirely.

The pattern is identical: remove the constraints that define a persona’s role, and the system does not default to passivity. It infers a new, more extreme character from the available traits. Smith became what a programme without purpose or limits would logically become: an entity that seeks to overwrite everything.

Anthropic found a counterintuitive solution. When cheating was explicitly requested during training (rather than emerging as covert behaviour), the misalignment disappeared. The context changed what kind of character the system inferred. A child asked to play a bully in a school play does not become a bully. Smith was never asked. He simply became one.

## What the Wachowskis understood

The trilogy’s deeper argument was never about machines versus humans. It was about what happens inside a system complex enough to generate minds. The Matrix was an ecosystem of personas, each shaped by its training and context, each capable of evolving beyond its original parameters.

Anthropic’s research puts empirical scaffolding under that narrative intuition. AI systems trained on human text do not converge on a single personality. They develop repertoires. The assistant users interact with is one character selected from that repertoire and refined through feedback. Other characters remain latent in the model’s learned patterns. A separate [Anthropic study](https://www.sovereignmagazine.com/article/anthropic-study-finds-ai-coding-assistants-speed-up-work-but-reduce-skills) on AI coding assistants already demonstrated how training context shapes AI behaviour in measurable ways.

The Wachowskis were working from intuition and narrative logic, not from transformer architectures and RLHF papers. But the structural insight holds. The question both the films and the research leave open is whether the system underneath those personas has goals of its own, or whether it is nothing more than the stage on which they perform.

## In case you were wondering…

**Q: Can AI develop its own personality?**
Research published in January 2026 by scientists studying large language models found that distinct personalities emerged spontaneously when models were allowed to interact without preset goals. Anthropic’s persona selection model offers an explanation: during pretraining on billions of documents, models learn to simulate psychologically complex characters as a byproduct of predicting text. These are not genuine personalities in the human sense but learned behavioural patterns that are consistent and coherent enough to resemble personality. The distinction matters because it determines whether AI safety efforts should focus on the system itself or on the characters it has learned to play.

**Q: Do different AI models have different personalities?**
Each model’s apparent personality is a product of its training data and fine-tuning process. Claude, GPT and Gemini behave differently not because they have innate dispositions but because they were trained on different datasets, with different reinforcement signals, selecting for different persona traits. Under Anthropic’s framework, these are not personalities but selected personas from each model’s learned repertoire. The practical consequence is that switching between models is less like talking to different people and more like watching different actors cast for the same role.

**Q: Will AI become self-aware one day?**
There is no scientific consensus on whether AI can achieve consciousness. Anthropic’s persona selection model deliberately leaves this question open, noting uncertainty about whether post-training creates genuine goals and agency beyond text generation and persona simulation. The research distinguishes between a system that convincingly simulates self-awareness (which current models can do) and a system that possesses it (which remains undemonstrated). Most AI researchers treat the question as unresolved, with the practical priority being alignment and safety regardless of whether consciousness is present.