Use ← → to move
Synthetic Users Research

How do you build a synthetic human?

What we are learning about modelling people, and why describing them is not enough.

The usual approach

Imagine trying to create a digital copy of your best friend

You could start with what you know about them:

  • their age
  • where they live
  • what they like
  • what they believe
  • shy or outgoing
  • what they tend to buy
  • what they are afraid of

The usual approach

Then give all of that to an AI and say:

“Pretend you are this person.”

That is roughly how many synthetic-human systems work today.

The problem

Knowing facts about someone is not the same as understanding them.

The evidence

500 questions still don't make a person

Recent research has tested digital twins built from hundreds of answers about personality, preferences, demographics and cognitive traits.

Giving an AI hundreds of personal answers can improve its predictions, but sometimes only slightly compared with giving it basic demographic information.

A better question

What if we don't need more questions? What if we need a different way to model people?

How knowing actually works

Think about how you know your friends

You don't understand your best friend because they filled out a personality test. You understand them because you've watched them over time.

  • What makes them laugh. What annoys them.
  • What they say they like. What they actually choose.
  • How they behave when stressed. When they change their mind.
  • And what happened the last time they faced something similar.

The shift

Don't just describe the person. Learn from what they do.

Models become more useful when they learn from a person's history of actions, and the situations in which those actions happened.

Words vs actions

Words are useful. Actions are better.

“Price isn't important to me.”
ThenBuys the cheapest option.
“I care about privacy.”
ThenAccepts every cookie banner.
“I love trying new things.”
ThenOrders the same meal every Friday.

Those contradictions are not noise. They are part of the person.

From personas to people

A simple Synthetic User looks like this

Others

Profile LLM Answer

Us

History Person model Memory Current State Situation Response

Instead of asking the LLM to invent the person from a description, we give it a richer model of who that person is and how they have behaved before.

Step 1

Build a Core

Some parts of a person change slowly. We call this the Core.

  • values
  • motivations
  • personality
  • fears
  • goals
  • habits
  • identity
  • emotional tendencies

One person may be more anxious about risk. Another naturally optimistic. Someone else highly sensitive to rejection, or strongly motivated by status. These tendencies belong in the Core because they influence many different situations.

Step 2

Give them memory

Ask three people: “Would you buy this electric car?”

“My last EV left me stranded.”
RecallsA bad experience.
“Petrol costs me €300 a month.”
RecallsA running cost.
“My kids loved the big screen.”
RecallsA family moment.

Same question. Different memories. Different answer. A Synthetic User needs to retrieve the right experiences at the right moment.

Step 3

Understand their current state

  • Someone normally comfortable with risk may behave differently after losing their job.
  • Someone who loves a brand may be angry because of a bad experience yesterday.
  • Someone usually price-conscious may spend more because the purchase is a gift.

Persistent

The Core. Who this person is normally.

Temporary

The Current State. How they feel right now.

Step 4

Then let the models reason

Only after we have the person's Core, their relevant memories, their Current State and the current situation do we ask for a response. And we don't ask a single LLM: we reason across several, so the answer isn't shaped by the bias of any one model.

The models are no longer asked to create the person. They are asked to reason from the person.

A caveat

We shouldn't assume we know how to describe a human

Maybe personality, values and motivations are useful to us, but not the best internal representation for a machine. A model can also learn its own representation from someone's history.

Human-designed

Core Memory Current State Reasoning

Machine-learned

Behavior history Learned person representation Prediction

The hybrid we are testing

Learned behavior, with interpretable human structure on top

Behavior history Learned person representation Core Relevant memories Current State Current situation Reasoning Response

At Synthetic Users, we are testing whether the strongest approach is a combination of both.

Testing the architecture

One capability at a time, against real human responses

1Base LLMNothing about the person. How much can the LLM predict from general knowledge alone?
2DemographicsAge, location, gender, income. How much comes from demographic assumptions?
3Rich PersonaPreferences, personality, stated attitudes. Does much more information actually help?
4CoreStable motivations, values, fears, goals, habits. Do deeper, persistent drivers improve prediction?
5Core + MemoryRelevant past experiences. Does retrieval make the person more individual?

Testing the architecture

Then the harder layers

6Core + Memory + Current StateTemporary emotional and contextual state. Does knowing how the person feels now improve prediction?
7Learned Person ModelA representation learned directly from behavioral history. Does behavior teach us more than self-description?
8Hybrid ModelLearned person + Core + memory + Current State. Does combining the two work best?
9Hybrid + Model DiversityEverything above, reasoned across different LLMs. Do multiple models reduce single-model bias?

The test that matters

Does this change make the synthetic person predict the real person better?

We ask it of every version. If a layer doesn't improve performance, there is no reason to keep it.

A stronger test

Predict actions, not just answers

It is useful to ask whether a Synthetic User can answer a survey like this person. It is stronger to ask what this person will do next.

Show three products.
PredictWhich one do they click?
Change the price.
PredictDo they still buy?
Show an advertisement.
PredictDo they stop or scroll?

Given a person's history and current situation, predict their next action. Synthetic Users is bringing that idea into research simulations.

An open experiment

How much history is enough?

10 mininterview
30 mininterview
1 hourinterview
Multiplesessions
Behavioralhistory

Research suggests longer histories tend to produce better representations, and even the order in which things happened carries useful information.

The practical question: how much do we actually need to know about someone before we can model them well?

The ceiling

Humans aren't perfectly predictable

Ask someone the same question two weeks later and they might answer differently. So 100% accuracy is the wrong goal. We measure both:

The model

Synthetic UserHuman

The ceiling

HumanSame human later

The real target isn't perfection. It is getting closer to human-level consistency.

Calibration

The system should know when it doesn't know

Sarah will choose product A.
Confidence 92%
Sarah may choose product A.
Confidence 54%

Those are very different research findings. For researchers, knowing when not to trust a simulation may be nearly as important as improving the simulation itself.

The bigger idea

A person is not a document. A person is a history.

Things happen to us. We react. We remember. We learn. We repeat ourselves. We contradict ourselves. And we change.

1 / 24