Semilattice
All posts

The simulation economy

AI is bringing the cost of knowing how people will react to near zero. This gives rise to what we call the simulation economy: an economic system shaped by simulation at every level.

Joseph Wright
Joseph Wright
Co-Founder, Product & Engineering
A man in a dark suit walking across a zebra crossing, seen from above, with other pedestrians cropped at the top of the frame
Photograph by Gab Pili on Unsplash.

Until now there were two ways to find out how someone would react to something. You could ask them in advance, or you could do it and watch. Both need the person, and that has always made finding out expensive.

The 2020 US census cost about US$14.2 billion, or US$96 for each household. A short survey of the general public costs around a dollar a response on a self-serve panel, so a few hundred dollars for one set of questions. A niche audience, such as finance directors or GPs, costs US$30 to US$300 a response, so thousands for a single study, and each study takes days to field. In total, organisations spent more than US$150 billion in 2024 on research into what people think and do. At those prices, most decisions aren't tested.

But not knowing is expensive too. Projects that depend on how people respond fail at rates that would be unacceptable for a bridge or a medical procedure, a point we made in Humans + Time. When CB Insights read the accounts that 101 failed startups gave of what went wrong, the most common reason, cited by 42%, was that there was no market need for the product.

Now models can predict how specific people will respond to things they have never seen: a price, a sentence, a policy, an offer. The predictions aren't perfect, but the best models now get close to the answers real people give. Each prediction takes minutes and costs almost nothing.

A crowd crossing a city street seen from above, each person casting a long shadow in low sun
Photograph by Levi Meir Clancy on Unsplash.

A prediction doesn't need anyone's time. It runs on an AI model, and the cost of running an AI model of the same ability has dropped about a thousandfold in three years. Epoch AI measures a fall of between 9 and 900 times a year, depending on the level of ability. Ours costs US$139 a month for one person, with no limit on the number of predictions, so the hundredth decision tested in a month costs nothing extra and takes minutes.

When something expensive becomes nearly free, people stop rationing it. You don't stop to decide whether a web search is worth running. The same is about to happen to knowing how people will react, and the effects will reach far beyond the product and marketing research it is used for now.

Below are twelve things that change as a result, with the evidence for how far along each one is today. They fall into four groups: deciding, persuading, being predicted and knowing.

The simulation economy: twelve things that change when knowing how people will react costs almost nothing, in four groups. Deciding: 1, every decision can be made with evidence; 2, one round of testing becomes many; 3, small votes can be forecast, not just elections. Persuading: 4, campaign messages are written for one person; 5, you know what the other side will accept before you meet; 6, bad actors get the same tools. Being predicted: 7, institutions predict you without you knowing; 8, models work out what you never told anyone; 9, predictions change the people they predict. Knowing: 10, predicted opinions become their own kind of data; 11, being unpredictable becomes a strategy; 12, prediction hits a limit, or keeps improving.

Deciding

1. Every decision can be made with evidence.

Count the decisions in a piece of work. When we asked an AI agent to pull every decision out of one section of a feature spec, it found 32. Very few of them would normally have any evidence behind them. Finding out how people will react takes time and money, so the effort goes to the handful of decisions that look big, and the long tail of small ones is decided on a guess.

An office building facade, a grid of windows with people working at desks inside
Photograph by dylan nolte on Unsplash.

As prediction gets cheaper, evidence reaches the whole tail. That is important, because you can't tell how big a decision is by looking at it. One line in that spec said reminder emails would send from the business owner's own email account. It reads as an implementation detail that would never normally be researched, but 75% of simulated owners were very uncomfortable granting the permission that sending from their account requires, and 73% said it was too risky even if they approved every message first. The rest of the feature depends on that permission.

Errors like this one compound. Later work is built on them, and by the time the effect shows up in a product's numbers it is hard to trace back to the line that caused it. Checked at the moment they are made, they get fixed before anything is built on them.

2. One round of testing becomes many.

When it is expensive to find out how people will react to something, you find out once. You pick the version you believe in, test it or something close to it, and ship.

When it costs almost nothing, there is no reason to stop at one round. An agent can generate options, predict how people react to each, keep the best, write variations on those and go again, and each round takes minutes.

The number of things a team could build is far larger than it looks. Just five decisions with between five and eight options each describe 15,360 possible products, and the spec we looked at above held 32 decisions in a single section. One round of testing checks one or two products, so almost every product a team could have built is never looked at. A loop can test every option for each decision and combine the best, reaching products the team never considered.

The spec, as written 1 possible product The spec, as written 1 possible product
Every dot is one of the 15,360 products that five decisions in one spec could describe. The accent dot is the one the spec chose; the other is one of the 15,359 it didn't.

A loop optimises exactly what it is told to measure, and nothing else. Told to maximise sign-ups, it will find the version that gets the most sign-ups, even if that version over-promises. Choosing what to measure, and what must not be traded away for it, becomes the main job left for the people running the loop.

3. Small votes can be forecast, not just elections.

Some outcomes are decided by many people together: a vote, a strike ballot, whether a policy is taken up, how a market splits between two products. Forecasting them has meant polling, and polling is only worth paying for when the outcome is big. National elections get forecasts, but a union branch's ballot, a town's consultation on a new road or a school's parents voting on a change of term dates do not.

Three people crossing a patch of sunlight in a dark plaza, seen from above
Photograph by jin cl on Unsplash.

When prediction costs almost nothing, the small ones can be forecast too. A simulated population can be asked about an outcome that affects a few hundred people as easily as one that affects a country, and it can be asked again as the situation changes.

GPT-4's predictions of the results of 70 social science experiments correlated with the real results at 0.85, about as well as human forecasters, though it overestimates the size of effects. Close elections may still be beyond it. In one public attempt, the day before the 2024 US presidential election, an AI polling company published win probabilities from simulated voters for the seven swing states. It favoured Harris in four of them, and Trump won all seven.

Persuading

4. Campaign messages are written for one person.

Personalisation today means choosing which of a few existing messages to show to which person. With a model of the individual, the message can be written for that one person and checked against the model before it is sent.

Several people walking separately across an open city square, seen from above, each with a long shadow
Photograph by Ethan on Unsplash.

Writing for one person is at an early stage. The best models predict a specific person's exact answer to a survey question about two times in three, which is not yet good enough to write to one person with confidence. What has been tried so far is cruder: giving a language model a few facts about the reader and having it tailor the message. In a study of 8,587 people, political messages that GPT-4 tailored to the reader's attributes persuaded no better than its generic ones. That study used no model of the recipient, only a list of their attributes.

5. You know what the other side will accept before you meet.

Much of the cost of a negotiation, a dispute or a public consultation is discovering what the other party will agree to. With a model of the other party, you can find that out before the first meeting. Even a little knowledge of the other side already makes a measurable difference. In one study, 900 people debated either a human or GPT-4. When GPT-4 was given six facts about its opponent (gender, age, ethnicity, education, employment and politics), the odds that the opponent ended up agreeing with it were 81% higher than against a human debater. Without those facts it did no better than the humans.

The shadows of two people walking side by side, cast across stone paving
Photograph by Anne Nygård on Unsplash.

We haven't found a published case of anyone negotiating against a model of a specific counterparty, but nothing stands in the way. A union could rehearse a pay round against a model of the employer's board, or a buyer could rehearse against a model of one supplier, and neither would have to tell the other side. There is no norm yet for saying that you modelled someone before you met them.

6. Bad actors get the same tools.

Anything that helps an honest organisation find the message a person will respond to helps a scammer, a propagandist or an extremist recruiter do the same. A university experiment has already shown how well machine-written persuasion works. From November 2024 to March 2025, researchers at the University of Zurich ran AI accounts on a Reddit debate forum without telling anyone. For some replies, a second model first inferred the poster's gender, age, ethnicity, location and politics from their posting history. The accounts posed as, among others, a rape victim and a trauma counsellor. Their replies changed the poster's mind about six times as often as the human average. The experiment ran for four months before the forum's moderators made it public.

The same has been measured for scams. In a study of 101 people, generic phishing emails were clicked 12% of the time. Emails that human experts tailored to each recipient were clicked 54% of the time, and so were emails that AI researched and wrote on its own. A well-targeted scam used to need an expert's time for each victim, and here the AI matched the expert. Models also attack each other: in a competition in which language model agents negotiated with each other more than 180,000 times, one of the top agents used prompt injection, text written to hijack another model, to extract its counterpart's private information.

Being predicted

7. Institutions predict you without you knowing.

The closest precedent is the credit score. FICO, the US credit scoring company, introduced its score in 1989, and for more than a decade lenders used a number about you that you had no right to see. California gave consumers that right in 2000, and federal law followed in December 2003, for a fee.

Two people walking past the dark colonnade and closed doorway of a large stone building
Photograph by Fumiaki Hayashi on Unsplash.

For predictions about people, one such right already exists. California's privacy law counts "inferences drawn" about a consumer's preferences, psychological trends and behaviour as personal information, and the state's attorney general confirmed in 2022 that a business must hand them over on request. Whether a simulated reaction to an offer you were never shown counts as an inference about you has not been tested.

8. Models work out what you never told anyone.

In 2013, researchers showed that a person's Facebook Likes were enough to tell gay men from straight men 88% of the time and Democrats from Republicans 85% of the time. None of those people had stated either fact.

An empty office at night, seen from the street through half-open blinds
Photograph by jatinder nagra on Unsplash.

The law covers part of this. In 2022 the EU's highest court ruled that data from which sensitive information can be deduced is itself sensitive data. But the rule on profiling in the GDPR, the EU's data protection law, is about decisions and not inferences. Article 22 gives you rights when a system decides something about you with no person involved, such as refusing a loan or rejecting a job application. An inference is a conclusion a system draws about you, such as that you are about to change jobs or that a certain sentence will persuade you. A model that predicts your reaction in order to choose the wording you are shown draws inferences about you constantly, but Article 22 doesn't cover it, because choosing which wording you see isn't a decision like refusing a loan.

9. Predictions change the people they predict.

The effect of a prediction on the people it describes is old. Economists know it as Goodhart's law: a pattern in people's behaviour stops holding once someone starts acting on it. In machine learning it is called performative prediction.

The long shadow of a person standing at the top of a flight of stairs, stretching down the steps
Photograph by Buddy AN on Unsplash.

The effect has been measured. In a field experiment in Kenya, 1,557 people were told the rule by which their phone use would be scored. They changed their phone use, and the standard scoring models lost 23% of their predictive power.

What is new is how often people are predicted. Until now, a prediction about a person was made occasionally, by a credit bureau, an insurer or a pollster. When prediction costs almost nothing, one is made before every message, price and offer a person sees. People are then responding to predictions of themselves constantly, and each new model is trained on behaviour that the last model already shaped.

Knowing

10. Predicted opinions become their own kind of data.

A predicted opinion isn't a fact, because nobody said it, and it isn't a fabrication, because it is derived from what real people have said and done. We expect it to become a recognised third kind of data: used widely, labelled as predicted, and held to standards of its own.

The labelling has started. In March 2026 a news outlet reported a poll in which 88% of respondents saw maternal mortality as a serious problem. The figure came from an AI simulation with no human respondents, and the outlet later added a note saying so. Three months later the American Association for Public Opinion Research amended its code: AI-generated responses must be identified as AI-generated, and the words "poll" and "survey" imply that the answers came from people. The international research industry's code of conduct now defines a "synthetic persona" and requires that clients are told when one is used.

The standards have not been written yet. The rules so far say what a predicted opinion must not be called. They don't say how accurate it has to be before it is reported. We think it should be published with its accuracy against real answers beside it, the way a poll is published with its margin of error.

11. Being unpredictable becomes a strategy.

Two firms setting prices, or two governments setting tariffs, each decide on the basis of what the other will do. Until now each has had to guess, and people are not good at it. In experiments on games where the best move depends on what the other players do, the average player thinks about one and a half steps ahead.

Two people on side-by-side escalators, seen from directly above
Photograph by Shane Rounce on Unsplash.

A model of the other side removes that limit. Each side can simulate the other's response, then the other's response to that response, the way a chess player does, and revise all of it as new signals arrive. Each side also knows that the other is doing the same.

Ted Chiang's short story "Understand" pictures a contest like this. Two men made superintelligent by an experimental drug each work out that the other is a threat. Most of the fight happens in their heads, with each one modelling what the other will do, given that the other is modelling him.

Poker suggests one way a contest like this could end. Pluribus, the program that beat professionals at six-player poker in 2019, learned by playing copies of itself, with no data on human players, and one of the professionals said its main strength was its use of mixed strategies, which means varying its play at random. If that carries over, the best answer to an opponent who can predict you is to be deliberately unpredictable.

12. Prediction hits a limit, or keeps improving.

A prediction about a person may have a ceiling, set by how much of what they do is random. If someone would give one answer to a question 90% of the time and a different answer the other 10%, purely at random, no model can be right about which one they give more than 90% of the time. And people do vary: asked the same survey questions two weeks apart, they give the same answer 81% of the time.

One person sitting alone on a long bank of identical escalators
Photograph by Alejandro G. on Unsplash.

What isn't known is how much of that variation is random. Some of it may be mood, what the person read that morning or how the question was put, which a model with more information could predict. If most of it is like that, the ceiling is higher than it looks, and prediction keeps improving until it is a science of human behaviour. If most of it is random, prediction stays a tool that is right most of the time.


A man sitting on the wide stone steps of a building with tall columns
Photograph by Flow Clark on Unsplash.

We build user models for product and marketing research, covering the first two items on this list. On item 10 we already do what we argue for: every model we build ships with its own accuracy report against real answers, and our guides label every simulated figure as a prediction.

The simulation economy is already arriving. How people will react to it is, for now, hard to predict.