What Exactly Are We Simulating? Five Categories of AI Simulation

Over the past year, I have watched a growing number of companies describe what they do as “simulation.” The use cases span marketing, sales, product development, user experience, quality assurance, and many other domains.

When I started building our own shopper simulation system, investors and technical and non-technical colleagues kept asking reasonable but very different questions:

• Are you generating fake data?

• Are you generating someone to talk to a customer-support agent on a website?

• Are you building digital twins?

Each question was a reasonable interpretation of the word. Each also described a different system.

We were combining a shopper, a journey through a website, a browser environment, and a set of test scenarios. At times, people also interpreted the work as a claim about a larger customer population. Those are not variations of one capability. They have different outputs, engineering requirements, and standards of evidence.

The distinction matters because “simulation” has become an umbrella term for almost everything: synthetic datasets, AI red teams, browser agents, digital twins, simulated focus groups, market forecasts, and virtual worlds. When these systems are discussed as if they belong on one capability ladder, the broadest claim often wins. A product that generates realistic customer conversations starts to sound like a system that predicts customer behavior. A thousand prompted personas start to sound like a market model.

That is category confusion, not technical progress.

In this article, I will offer a straightforward map based on one question:

What is the primary object being simulated?

This is a working framework, not an established industry standard. Its purpose is to make claims comparable before we debate architectures, benchmarks, or vendors.

A practical definition

A simulation is a deliberately simplified, executable representation of an object, actor, process, or environment. We use it to generate, test, explain, or forecast outcomes under specified conditions.

Two parts of that definition matter.

First, every simulation is selective. It preserves some features of reality and omits others. The right question is not whether the simulation is perfectly realistic. It is whether it preserves the features needed for the decision at hand.

Second, a simulation does something. It generates cases, produces responses, moves through states, or creates outcomes. A static persona card may describe a user, but it becomes part of a simulation only when the representation is used to produce behavior or test a system.

The same boundary applies to synthetic data. Randomly generated names and email addresses may be useful mock data, but they are not automatically a meaningful simulation. Data and scenario generation becomes simulation when it represents task-relevant structures, relationships, processes, or conditions so that a system can be trained or tested against them.

From that starting point, most AI simulations fall into five scopes.

These categories describe scope. They do not describe quality, ambition, or maturity.

1. Data and scenario simulation

The first category creates artifacts or cases. No persistent actor is required.

A team might generate synthetic transactions, medical records, images, support tickets, email threads, documents, security attacks, or rare failure scenarios. The output may be used for training, evaluation, privacy-sensitive development, or coverage expansion.

Imagine constructing a synthetic employee workspace. It contains an inbox, documents, calendar events, access permissions, outdated information, and a set of tasks with known correct outcomes. The workspace can test whether an agent retrieves the right evidence, resolves conflicting sources, performs the required actions, and produces a useful final result.

That is simulation, but it is not a simulated employee. The primary object is the information world and its test cases.

The central question is whether they preserve the structure, difficulty, and known outcomes required by the task. Article 2 will examine that problem in detail.

2. Interaction simulation

The second category creates a counterpart that responds turn by turn.

The counterpart might be a customer testing a chatbot, an attacker probing a model, a patient rehearsing a clinical conversation, or an interviewer responding to a research protocol.

Red teaming shows the distinction clearly. A fixed list of harmful prompts belongs in data and scenario simulation. An adaptive attacker that observes the target’s response and changes strategy belongs in interaction simulation.

HarmBench, for example, standardized behaviors and evaluation methods across automated red-teaming systems.AgentDojo goes further by placing tool-using agents in an extensible environment with realistic tasks and security test cases.

The primary output here is the adaptive exchange. The counterpart observes a response and decides what to say or do next. The key claim is not necessarily that it is a faithful human. It may be enough that it reliably probes a system, expands coverage, or exposes vulnerabilities.

3. Journey and environment simulation

The third category simulates a sequence of actions through changing state.

This includes browser agents completing workflows, customer journeys spanning several channels, operational processes, and embodied agents moving through physical or digital environments.

The environment matters because every action changes what happens next. A user sees information, clicks, waits, encounters an error, changes strategy, or abandons the task. A realistic first screen does not create a realistic journey if later state, permissions, latency, or failure recovery are missing.

Interaction and journey simulation can overlap. The practical distinction is the result carrying the claim. If the important output is the evolving exchange, the system is primarily simulating an interaction. If it is the action path through a changing environment, the system is primarily simulating a journey.

WebArena illustrates this scope with functional websites across e-commerce, forums, collaborative development, and content management. Its benchmark evaluates end-to-end task completion in a reproducible environment rather than grading a single response.

Our current shopper-simulation solution belongs primarily in this category. Although the early work aimed at person-level simulation, we now simulate grounded customer journeys rather than trying to map every person. The agent combines a general model of human goals, attention, knowledge, constraints, memory, and decision processes with adaptive customer evidence from qualitative research, observed behavior, and demographic context. Demographics provide context rather than defining a persona. We treat the resulting behavior and explanations as supporting signals for business analytics, not as digital twins or direct observations of private thought. As I described in anearlier account of this work, browser competence and grounded customer behavior can compete with each other.

4. Individual simulation

The fourth category targets a particular person over time.

This is the strongest reasonable meaning of a behavioral digital twin. The system is not merely asked to act like “a price-sensitive millennial.” It is grounded in evidence about one person and evaluated on whether it reproduces that person’s choices, responses, or decision patterns in held-out situations.

One recent study grounded agents in interviews and surveys, then compared them with held-out responses from the corresponding participants. That qualifies as individual simulation because the target is a particular person’s responses, not a generic demographic persona.LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals

The validation burden is much higher here. A system can produce a persuasive explanation in the person’s voice while still missing what that person would actually do in a novel situation. We will return to that problem in Article 3.

5. Population and world simulation

The fifth category models many actors plus the structures through which they affect one another.

Those structures may include networks, markets, institutions, geography, media exposure, scheduling, resource constraints, or an evolving environment. The outputs are usually distributions, diffusion patterns, collective behavior, or aggregate forecasts.

Recent systems can run thousands of language-model agents inside a shared social environment.AgentSociety is one example.

Scale is not the same as representativeness. If every agent inherits similar model priors, adding more agents may reduce sampling noise while preserving the same systematic error. Population outcomes can also be dominated by who interacts, when they update, which information they see, and what the environment permits. A 2026 position paper makes this point directly: plausible role-playing is insufficient when collective outcomes depend on agent-environment dynamics, scheduling, and initial information.AI Agents Alone Are Not Yet Sufficient for Social Simulation

Why this is not a ladder

It is tempting to arrange these categories from simple to advanced. That would be a mistake.

A well-constructed synthetic dataset may be production-ready for evaluating a retrieval system. A world simulator may remain speculative for predicting a policy outcome. The second system has broader scope, but the first may have stronger evidence and greater practical value.

The right category depends on the question:

• If you need more representative test cases, simulate data and scenarios.

• If you need to probe a conversational system, simulate an interaction.

• If you need to exercise an end-to-end workflow, simulate a journey and its environment.

• If you need to approximate one person’s behavior, simulate an individual and validate against that person.

• If you need aggregate dynamics, simulate the population, relationships, rules, and environment.

Products may span categories. Classify them by the output carrying the most important claim.

Under this map, user simulation begins when the system represents a user-like actor or user behavior through an interaction, journey, individual, or population. Data and scenario simulation can support user or agent evaluation without simulating a user at all. A synthetic inbox, for example, can test an assistant without claiming to reproduce the employee who owns it.

The practical test is to identify the primary object, purpose, grounding, interaction pattern, and validation claim. Those answers will not tell us whether a simulation is good, but they will tell us what evidence “good” requires. Before asking whether AI can simulate a user, we should state what is being represented and which decision the representation is intended to improve.

The next article will examine the most concrete category in this framework: how to build a synthetic world with known answers for training and evaluating retrieval, synthesis, and multi-step agents.