Talks

The "What-If Machine": AI Agent Simulation of Human Behavior — Michael Bernstein

52 min read

Michael Bernstein · 2025 · 46:01 · ▶ Watch

Summary

Central question. Leaders, managers, policymakers and product teams decide on “incomplete information about how people are going to react”. Could we build a what-if machine: simulated AI people who react the way our customers, our staff or our community would, so that a decision can be tried before it is launched?

Main thesis. Yes, in a limited and carefully graded way. The talk makes five moves:

  1. Decisions fail because we “can only launch once”. Earlier simulation tools were either too thin (a person as five parameters) or too scripted (the Sims), and “the models have been highly stylized and have had minimal impact”.
  2. Large language models have “read the book” on how people behave, so you can prompt one to play any person. Wrapped in an architecture of memory, reflection and planning, 25 such agents in the town of “Smallville” produce believable days and a self-organising party.
  3. Believable is not accurate. “Disney characters are believable”, so accuracy has to be measured. Agents built from a short demographic or persona description stereotype. Agents built from a two-hour interview with a real person replicate that person’s survey answers about 85% as well as the person replicates themselves two weeks later.
  4. Anyone can build an “agent bank” for their own organisation, but the data must be rich and relevant to the question.
  5. Trust should follow a ladder of risk: possibility, then qualitative attitudes, then quantitative outcomes, then multi-agent simulation. The higher the rung, the less you should trust it.

Close. Two applications beyond market research: “look before you launch” testing for online platforms, and sparring-partner simulations for conflict skills. He mentions the spin-out company Simile, points to an online course, and the recording ends on a QR code.

Stance. A researcher-teacher speaking to a mixed alumni audience of managers and leaders. He is enthusiastic, and also unusually direct about failure: he shows a 13% versus 1.2% error from a company that tried it, says multi-agent simulation is “not ready”, and tells the audience not to “over trust such a method”.

What this talk adds to the other pages on this theme. Several pages in this collection already cover the research. The two Park pages (the 2023 Generative Agents seminar and the 2026 8 Billion Digital Twins interview) are the lead author’s own telling, pitched at researchers and at investors. This page is the advisor’s telling for a general leadership audience:

  • Audience and purpose. The framing is “organizational leadership: team dynamics, org design and predictive decision-making” (the event description), not AI research. The examples are CPG, personal finance, policy and management, and the closing advice is an operating procedure for a manager.
  • The “what-if machine” and “look before you launch”. This is Bernstein’s own coined framing, absent from Park’s 2023 talk (which centres on architecture) and from the 2026 interview (which centres on “social physics” and scale). It makes the simulation a decision tool whose job is to narrow options, not to forecast.
  • A risk ladder and a funnel. Park’s 2023 talk reports errors as limitations; Park in 2026 talks about a “confidence model”. Bernstein gives a four-rung ladder of what to trust, and a concrete workflow: simulate 100 ideas, keep five, A/B test those on real people.
  • Frank counter-evidence. He volunteers a failed company deployment. The other Park pages show the cautionary literature only in their background sections.
  • A how-to for an organisation’s own agent bank, including a “bad / not terrible / better” rule for data.
  • Soft-skills training, with a result (a two-thirds drop in antisocial strategy) that neither Park page covers.
  • Where it matches Park. The architecture and the 1,000-person study are the same; the numbers he quotes are the same normalised figures. Siblings on the methodological side are the Anand page on persona engineering, the Angelopoulos page on prediction-powered inference (a statistical way to correct a cheap predictor with a small human sample, which is the formal version of his “validate on a small subsample”), and the Radius/MilkPEP case-study page.

Map

#Point
1Incomplete information and bad bets; the “can only launch once” problem
2The what-if machine: customers, policy, strategy, failure paths
3Simulation’s history: Schelling’s agent-based models, the Sims, AI teammates, “look before you launch”
4Why old models failed: five-parameter people versus scripts
5LLMs have “read the book” on behaviour; prompting a model to be a person; Smallville; a16z
6Building an agent: a persona, relationships, grounding language into a game world
7Talking to agents, intervening in the world, morning with the Lin family
8The Valentine’s Day party: information diffusion without a “party module”
9A 2025 reuse of Smallville with an epidemic radio report
10Architecture part 1: memory stream and retrieval (recency, importance, relevance)
11Architecture part 2: reflection
12Architecture part 3: planning and reacting
13Believable is not accurate: how do you measure it?
14Demographic agents and persona agents; the stereotyping problem (“rice”)
15The 1,000-person interview study and agent bank design
16Normalising against how people answer themselves two weeks later; the 0.85 result
17Bias results: politics hardest, gender and race small; replication of five experiments; Willer’s study
18How to build your own agent bank: bad, not terrible, better; relevance of data
19Risks: a company’s 13% versus 1.2%
20The ladder of rungs: possibility, qualitative, quantitative, multi-agent
21Mitigations: in-domain data, rough-edged problems, validate on a subsample
22Frontier 1: look before you launch for online platforms
23Frontier 2: conflict-negotiation sparring partners and the experiment
24Business applications, a16z, Simile, the online course

Putting it together

The mental model in one paragraph. Think of the agent bank as a cheap, fast, imperfect rehearsal audience. A language model supplies the general knowledge of how people talk and react. Real data about a specific person (best: a long interview) supplies the specific, so the model stops averaging and starts impersonating. A memory-and-planning architecture keeps one agent consistent over time. None of this makes the audience right; it makes it useful for a particular kind of question. For “what could go wrong?” and “what might people say?”, an 80%-correct answer still teaches you something. For “what percentage will do X?”, an 80%-correct answer can steer a decision wrongly, so it needs a measured error and a real-world check. For “what will a whole market do?”, the error cannot be checked at all, so Bernstein says do not rely on it.

Open tensions.

  1. Accuracy is normalised, not absolute. The 0.85 headline is a ratio to human self-consistency, not a hit rate. The raw General Social Survey (GSS) accuracy was 68.85%. A reader hearing “85% accurate” (the shorthand a sceptical listener may hear) is hearing a different thing (idea 5).
  2. The scientific tests and the business tests differ. The 1,000-person study tests individual survey answers with rich private interviews. The company example that failed tested distributions (a percentage) with probably thinner data (idea 7).
  3. A simulator that changes what people do. If many firms test messaging on the same models, the models’ blind spots become the market’s blind spots. Bernstein does not raise this (idea 7).
  4. Leaders asked to “narrow, then A/B test” must have the real test available. Where the real test is impossible (a policy, a reorganisation), the funnel offers no exit (idea 8).

Reading path.

Part 1 — Before you read

Key terms

  • Generative agent. A software character driven by a language model that has a persistent memory, reflects on it and makes plans, so that it stays one consistent “person” over days. Distinct from a chatbot (single turn) and a classical agent in an agent-based model (fixed rules).
  • Agent-based model (ABM). A simulation where many simple, individually-coded agents follow rules and a pattern emerges at the group level. Schelling’s segregation model is the textbook example.
  • Memory stream / reflection / planning. The three components of the Park et al. architecture (ideas 3).
  • RAG (retrieval-augmented generation). Fetch the few most useful stored items and put them into the model’s prompt, instead of giving it everything.
  • Persona agent vs demographic agent vs interview-based agent. Three ways to define a simulated person: a short narrative (persona), a list of attributes (age, job, region: demographic), or the transcript of a long interview (idea 4).
  • Normalised accuracy. The agent’s agreement with the real person, divided by how much that person agrees with their own answers two weeks later. 1.0 means “as consistent as the person is with themselves”.
  • Agent bank. A reusable collection of agents, each tied to a consenting real person, that can be queried in many studies (idea 6).
  • Digital twin. A common name for an agent modelled on one specific real person. Bernstein uses “twinned agents”.
  • The risk ladder. Bernstein’s four rungs: possibility, qualitative, quantitative, multi-agent (idea 7).
  • Rough-edged vs sharp-edged problems. Rough-edged: if the answer is 80% right it still helps. Sharp-edged: the answer is right or wrong and being wrong changes the decision.
  • Warning on “accuracy”. Words mean different things in different traditions: for a psychometrician, accuracy is bounded by reliability; for a marketer, accuracy usually means the gap in a percentage; for an ABM modeller, “validation” often means the model reproduces a pattern, not individuals. The talk moves between these.

People named (a reading list)

PersonWhy they come upStart with
Joon Sung Park”My PhD student” who led the generative agents workThe 2023 paper; see the Park pages
Robert MertonCited for the problem of designing for what a group will doThe unanticipated consequences of purposive social action (1936)†
Thomas SchellingAgent-based models, “Nobel Prize”Micromotives and Macrobehavior (1978)
David GruskySociologist whose American Voices Project supplied the interview scriptThe Stanford Center on Poverty and Inequality†
Robb Willer”My colleague here at Stanford” with a study of 100,000 peopleHewitt, Ashokkumar, Ghezae and Willer, Predicting Results of Social Science Experiments Using Large Language Models
Percy Liang, Lainie YallenNot named in this talk; Simile’s co-founders per the other pagesSee the 2026 Park page
An expert in conflict and negotiation at Stanford Business SchoolCo-built the conflict training toolUnnamed in the talk; see Source notes

The speaker’s own frame

Bernstein is a computer scientist in human-computer interaction (HCI), a field that studies how people use and are changed by systems†. His earlier career was on online platforms: “design of online platforms, social media” where “issues would come up because we couldn’t effectively foresee what was going to go wrong”. So he reads simulation as a design tool for socio-technical systems, an echo of Park’s 2022 Social Simulacra work covered in the 2023 Park page. For him the question is not “can AI replace respondents?” but “can we find the failure before launch?”

The current debate (2025 to 2026)

This is background, not part of the talk. Since the August 2025 webinar the evidence has moved in two directions.

Supportive or ambiguous.

  • Scientific replication at scale. The Willer study he mentions (Hewitt et al.) has an official headline: GPT-4 predictions of the treatment effects of survey experiments on representative samples correlated with the real effects at r = 0.85, and 0.90 for unpublished studies (Stanford AI for Public Benefit project page; Stanford GSB listing). The GSB entry also notes that predictions overestimated effect sizes and held up less well on a second archive of megastudies. The study count and participant numbers differ between preprint versions, so cite the version you use.
  • The market-research thesis. The a16z article Bernstein cites, “Faster, Smarter, Cheaper: AI Is Reinventing Market Research” (Zach Cohen and Seema Amble, 3 June 2025), opens with generative agents (a16z). A venture firm’s thesis is an argument for investment, not a test.
  • Reuse of Smallville. A June 2025 preprint ran the Smallville setup with news of a swine-flu outbreak; agents reported lower social engagement and lower attendance at a social event (Choi et al., arXiv 2506.13783). It is a variation, not an exact replication of the story Bernstein tells.

Cautionary.

  • Pew Research Center (30 September 2026). Built digital twins of real members of its American Trends Panel and compared them with the people themselves. The twins exaggerated gaps between demographic subgroups, and humans chose “not sure” about four times as often (Pew). The twins there used less than a two-hour interview.
  • Verasight (June 2026). Four experiments found that digital twins got a few high-profile questions right and missed most others; the authors call them predictions, not measurements of public opinion (Verasight).
  • Marketing research. A 2026 study in NIM Marketing Intelligence Review found that synthetic respondents beat chance on individuals and caught broad trends, but overstated positive attitudes (especially to well-known brands) and varied much less than people did (Kaiser et al.).
  • “Digital Twins as Funhouse Mirrors.” In a test with more than 500 data points per person, twins with full records scored only slightly above twins with no individual information, and five systematic distortions appeared (arXiv 2509.19088). That test used prompting, not interviews as memory.

Net effect on this talk. Bernstein’s advice (rungs, relevance, subsample checks, funnel to A/B) already anticipates most of this. What the later work adds is that the subgroup and variance problems appear even with rich input, and that single headline numbers (0.85, r = 0.85) do not transfer to a new domain.

The ideas

Bernstein speaks to executives, so he rarely names the traditions behind his claims. This map includes them anyway. Eight traditions meet in the talk:

 DECISION-MAKING UNDER                 AGENT-BASED MODELLING &
 UNCERTAINTY                           SOCIAL SIMULATION
 Merton · Simon · Tetlock              Schelling · Epstein-Axtell
 scenario planning · pre-mortem        Sims · "highly stylized"
          \                                   /
           \                                 /
            ▼                               ▼
       ┌────────────────────────────────────────────┐
       │  BERNSTEIN: a "what-if machine" for        │
       │  leaders, built from LLMs + memory +       │
       │  real interviews, graded by a risk ladder  │
       └────────────────────────────────────────────┘
            ▲            ▲             ▲           ▲
            │            │             │           │
  COGNITIVE         SURVEY &       REPLICATION &    HCI DESIGN &
  ARCHITECTURES     PSYCHOMETRICS  VALIDATION       TRAINING
  Soar · ACT-R      reliability,   Box · Open        Simulacra,
  Tulving · RAG     silicon        Science · the     prototyping,
                    sampling       "ladder"          rehearsal

Where each tradition is covered:

  • Decision-making under uncertainty (scenario planning, pre-mortem): idea 1
  • Agent-based modelling and the “stylized” critique: idea 2
  • Cognitive architectures and memory: idea 3
  • Stereotyping, personas and silicon sampling: idea 4
  • Survey methodology and psychometrics (reliability ceiling): idea 5
  • Data design for an agent bank: idea 6
  • Model validation and the epistemology of simulation: idea 7
  • HCI prototyping and simulated rehearsal: idea 8

1. The what-if machine: deciding when you can only launch once

In short. Bernstein starts from a manager’s problem, not a technology. We make decisions with “incomplete information about how people are going to react”, we cannot test them on the real world more than once, and so we bet. A simulator that could say “if we do this, here is how they might react” would improve many decisions. The pitch is a rehearsal, not an oracle.

🎙 In the talk.

  • “We make these decisions based on the best data we can but that data is often incomplete. And this kind of incomplete information means that we often make bad bets.”
  • “It’s not that you’re not smart. It’s just really hard.”
  • “We can only launch once.”
  • The problem is old: he cites Merton on how hard it is to design for what a group of people will do. “We all want to get out of the city so that we can avoid the crowds, but then everyone wants to get out of the city and go on vacation at the same time.”
  • “Imagine that you could say if my organization took this path, how might our customers react? Or if you could say, if we were to fail or if something were to go wrong, what’s the pathway through which that would maybe happen?”
  • “If you had this kind of what if machine, my claim is that we could make better decisions much more often.”
  • His own history with platforms: “we couldn’t effectively foresee what was going to go wrong” and the aim is rules set “before I launch the system, so that it doesn’t all become a dumpster fire”.

📚 Deeper theory

  • The vacation example is a coordination problem. Each person acts reasonably; the sum is bad. Economists call such a gap between individual intention and group outcome an externality or a fallacy of composition. It is exactly what a simulation of many interacting agents is designed to show, because no single actor’s plan reveals it.†
  • The “can only launch once” claim is about missing counterfactuals. To know if a policy worked you would need the same world, same moment, without the policy. Statisticians call this the fundamental problem of causal inference (Holland, 1986)†. Simulation is a way to produce plausible counterfactuals when real ones cannot exist.
  • Two older uses of “what-if” already exist in management. Scenario planning (famously at Shell in the 1970s) and the pre-mortem (Gary Klein, 2007: imagine the project has failed and write why) both rehearse failure before it happens. Bernstein’s “pathway through which that would happen” is a pre-mortem done by a machine.†
  • Why the framing matters. A “what-if” machine is judged by whether it surfaces options you had not thought of, not by whether it predicts. That choice protects it from the most common critique (it is not a forecaster) but moves the burden to the user: someone must still judge which of the machine’s paths are plausible. Idea 7 returns to this.

Where it is weak. The pitch moves quickly from “decisions are hard” to “a model of people can help”. But a plan that fails because the organisation could not execute it, or because competitors reacted, is not something customer simulations will show. And a simulator tells you what a model of people would do, which may be what the decision-maker already believes. This is confirmation by simulation.

📚 Roots: Decision-making under uncertainty

The problem. Humans make decisions without knowing the consequences, and a decision’s quality and its outcome are different things.

The point. Bernstein’s “bad bets” are decisions made with the usual information and still wrong. The talk offers a tool to widen the set of consequences considered before the bet.

The chain of thinkers.

  • Robert Merton (1936), “unanticipated consequences of purposive social action”: even good intentions yield unintended outcomes, from ignorance, error, short-term interest or self-defeating predictions. Bernstein dates his Merton reference to 1906, which cannot be right (see Source notes), but the idea is Merton’s.†
  • Herbert Simon (1950s), bounded rationality and satisficing: people cannot compute the optimal choice, so they find a good-enough one. A what-if machine extends the options a bounded decision-maker can explore.†
  • Kahneman and Tversky (1970s), planning fallacy and overconfidence: planners systematically underestimate time and cost. Rehearsal tools aim to break this.†
  • Pierre Wack and Shell (1970s), scenario planning: build several coherent futures to prepare for, instead of one forecast. Bernstein’s “multiverse” language (in Park’s Social Simulacra) shares this logic.†
  • Philip Tetlock (2005 and 2015), forecasting: experts do little better than chance on long-range political predictions, but trained “superforecasters” do much better with feedback and calibration. Bernstein’s insistence on measuring accuracy against ground truth is Tetlock’s standard.†
  • Gary Klein (2007), the pre-mortem.†

What they leave out. Most of this literature assumes a human reasoner who imagines the consequences. A simulator changes who does the imagining, and whose imagination (the model’s training data) is at work.

Check yourself.

  1. Why does “we can only launch once” make simulation more valuable than more data about past customers?
  2. What is the difference between a forecast and a pre-mortem, and which does Bernstein want?
  3. Name one decision in your own work where the main risk is not customer reaction. Would this machine help?

2. Why earlier simulations failed, and what changed

In short. Simulation of people is old. Its two classical forms are mathematically elegant but thin, or rich but hand-scripted. In the research literature that meant “highly stylized and minimal impact”. Bernstein’s turning point is that language models, having read vast amounts of human writing, can be prompted to be a particular person, and many different prompts give a crowd.

🎙 In the talk.

  • “It goes all the way back to 1978. Here’s Thomas Schelling won a Nobel Prize in economics for generating this idea of what are called agent-based models.” (see Source notes: the Nobel citation was not for this)
  • “I bet a bunch of you have played the Sims.”
  • Rigid options: “you say a human is five parameters” or “you write scripts and say, if I get punched, then I fall down and I get mad.”
  • The conclusion in the literature: “the models have been highly stylized and have had minimal impact.”
  • The change: LLMs “have trained on a variety of human behaviors, which is to say that they’ve literally read the book on all the research about how people behave”.
  • The recipe: “a prompt that roughly equates to be this person, here’s a name, a description, and in this situation, how would they react? And by multiplexing those descriptions we can create the entire crowd of different people.”

📚 Deeper theory

  • Why scripted agents hit a wall. A hand-written rule base can only contain behaviours that its author foresaw. AI researchers met this in the 1980s as the knowledge-engineering bottleneck of expert systems. Bernstein describes the same wall in different words: “it’s always going to be incomplete”.†
  • Why “five parameters” is not a straw man. The Big Five personality model describes people on five dimensions and predicts many life outcomes. It is rich as psychology and thin as a simulation: five numbers cannot generate a plausible conversation.†
  • The LLM as a “compressed culture”. A model trained on web text and books carries a statistical imprint of how many kinds of people write, argue and decide. That is why a prompt can shift it into one voice. It is also why it carries the average of its sources, with stereotypes and gaps (idea 4).†
  • From Bernstein’s earlier work. The same idea, applied to online communities, is Social Simulacra (2022), discussed in the 2023 Park page. Bernstein’s point there is the same: you do not need a perfect predictor, you need plausible members to test rules on.
  • Key shift in what a “model of a person” is. Old models parameterised a person with variables the modeller chose. A prompted LLM lets the person’s own story be the parameter. This is the idea behind interviews (idea 5).

Where it is weak. The talk says “the vast majority of people who I talk to have not [used ABMs]”, which is an anecdotal sample of a business audience. ABM is widely used in epidemiology, ecology and transport†, and ABM modellers would add that thin, transparent models are valued because you can inspect why they behave as they do. An LLM agent is a black box.

📚 Roots: Agent-based modelling and social simulation

The problem. How do group-level patterns (a segregated city, a pandemic curve, a market crash) arise from individual choices that nobody intended to produce them?

The point. Bernstein’s what-if machine is an ABM whose agents are far more complex than anything in the ABM tradition. He takes the tradition’s goal (explore emergence by running a world) and swaps the agent’s insides from a rule to a language model.

The chain of thinkers.

  • Thomas Schelling, segregation model (1971; book 1978). Agents with only a mild preference for same-group neighbours, moving when unhappy, produce strongly segregated maps. It was first built on a chessboard with coins, not a computer (Wikipedia on Schelling’s model). The principle: macro-patterns need not match micro-intentions. Schelling’s 2005 Nobel was for work on conflict and cooperation using game theory, which is related but different.
  • John von Neumann and cellular automata (1940s) and Conway’s Game of Life (1970). Rules on a grid creating complex patterns.†
  • Joshua Epstein and Robert Axtell, Growing Artificial Societies (1996). “Sugarscape”: agents harvesting a resource landscape, producing trade, inequality and migration. The slogan is generative social science: “if you didn’t grow it, you didn’t explain it”.†
  • Robert Axelrod (1984), tournaments of repeated prisoner’s dilemma strategies. Simple strategies like tit-for-tat winning.†
  • Epidemic ABMs (2000s onward): large agent models run on supercomputers to test interventions, the type Bernstein shows on a slide.†
  • Joshua Epstein, “Why Model?” (2008): lists at least sixteen reasons to build a model other than prediction, including to illuminate core dynamics, to suggest new questions, to discover new questions, and to expose the unexpected. This is the best theory for Bernstein’s “possibility rung” (idea 7).†
  • Cognitive-science side: Allen Newell and Herbert Simon, who wanted programs that think like people (1950s–70s); and game AI, such as The Sims (2000), which uses scripted needs and moods.†

What they leave out. The ABM tradition’s strongest norm is parsimony: a model with many free parameters can fit anything†. A generative agent has billions of hidden parameters. The older tradition would ask what claims such a model can test.

Check yourself.

  1. What does Schelling’s model show that no individual agent “intended”?
  2. Why is a transparent but thin model sometimes better than a rich black box?
  3. Which of Epstein’s reasons for modelling (other than prediction) does Bernstein’s “possibility” use case match?

3. The architecture: memory, reflection, planning, and what a party tells us

In short. To keep an LLM-based character coherent across a day, Bernstein’s team gave it a memory stream with a smart retrieval step, reflection that turns events into beliefs, and planning that is revised as things happen. The result is not scripted: a single seed (“plan a Valentine’s Day party”) produced invitations, help, attendance and even a crush. He is clear that the same architecture has no “party planning module”.

🎙 In the talk.

  • The setup: 25 generative agents in the town of Smallville, “all completely autonomous AIs that are each playing the part of a different person.”
  • “We have to tell that agent everything that it needs to know at the beginning” or the agent “will wake up in a bed next to Mei Lynn” and not know her.
  • Grounding: agents “mostly just act in the sense of they use natural language”, which is translated into movement and an emoji.
  • Interventions: “I can take the toaster, set it on fire, and the agent will recognize what’s going on and run over to put out the fire”.
  • The party: “nothing about this AI agent architecture that we’re talking about today is like a party planning module.”
  • Outcome: “of the 25 agents in the town, 12 of them heard about the Valentine’s Day party. Of those 12, five came to the party.”
  • Honesty about validity: “It reads to me as broadly plausible, but it’s not necessarily easy to tell.”
  • Memory: “a blow by blow record of everything that the agent observes”. Retrieval chooses by recency, importance and relevance: “almost like running a little bit of a Google search over all of these memories”. Reason: models “can get distracted with very, very long contexts”.
  • Reflection: “shower thoughts, figuratively speaking”. Memories are grouped and summarised into higher-level conclusions that are written back into memory.
  • Planning: “plan out their full day. Then we would drill in… hour by hour… minute by minute”, and on seeing something the agent decides “if it should react” and replans.

📚 Deeper theory

  • Why the party shows diffusion. The result is a pattern known from network science: information spreads along social ties. A rumor network like this is the same object epidemiologists model (hence the 2025 reuse with an epidemic radio report)†. In the paper the numbers are 12 agents who heard and five attended; secondary sources agree on 12 and 5 (Park et al.; news summaries). Bernstein’s “three said busy, four interested but did not show” matches the secondary reports of the same run.
  • The epidemic variant is a variation, not a copy. A 2025 preprint adds news of an outbreak and finds lower social engagement (arXiv 2506.13783). I could not find the specific “only Klaus showed up” detail in the sources checked.
  • Retrieval by recency, importance, relevance is a design for a limited context. Bernstein says the whole stream could fit but the model gets “distracted”. That claim is an engineering one about long-context behaviour in 2023-era models; how well it holds for later models is an open question.†
  • Detail already treated in depth in the 2023 Park page (memory, reflection and planning each get their own idea there). This page does not repeat the scoring study (TrueSkill, ablations).

Where it is weak. A party that “emerges” tells us the system is coherent, not correct. The same page-level caveat Bernstein gives himself: it reads as plausible but “it’s not necessarily easy to tell” if real life would match. He also leaves out that agents tend to be too polite and agreeable (the 2023 Park talk lists this).

📚 Roots: Cognitive architectures and memory

The problem. What does an agent need, beyond raw intelligence, to behave as one continuous person?

The point. Memory, reflection and planning are not new inventions in the talk. They map onto decades of cognitive science, which says an agent needs separate stores for events and for generalisations, plus goals.

The chain of thinkers.

  • Endel Tulving (1972): episodic memory (events: “I ate lunch at noon”) versus semantic memory (facts and generalisations: “I like soup”). Bernstein says explicitly that the raw stream is “purely episodic” and that reflection adds the semantic layer.†
  • Allen Newell, Unified Theories of Cognition (1990), and Soar; John Anderson and ACT-R. Architectures where memory, goals and rules interact, designed to model human problem-solving. Smallville’s loop (perceive, retrieve, plan, act) is a descendant in spirit.†
  • Hierarchical planning (STRIPS, 1971; HTN planning). Plan coarse then refine, as the “full day, then hour, then minute” scheme does.†
  • Ebbinghaus (1885) and decay curves: recent items are retrieved more easily. The “recency” score echoes this. Retrieval by importance has an echo in the flashbulb memory literature on emotional events. Relevance is plain information retrieval (cosine similarity of embeddings).†
  • Retrieval-augmented generation (Lewis et al., 2020). The neural technique Bernstein names as RAG.†
  • Dennett and narrative selfhood (1991): a self is the story told about oneself. Reflection is a machine version, which produces a “I read a lot” summary and stores it as a trait.†

What they leave out. Emotion, body and drives, which the Sims tradition makes central. Also, forgetting: Smallville agents accumulate; real memory prunes and distorts, and the distortion matters for how people actually decide.

Check yourself.

  1. What is the difference between “Klaus is reading about gentrification” and “Klaus spends a lot of time reading”? Where is each stored?
  2. Why can’t you just paste the whole memory stream into the prompt?
  3. If the party had no one who remembered the invitation, which component of the architecture would have failed?

4. Believable is not accurate: stereotypes and the three kinds of agent

In short. The Smallville agents are convincing, but “Disney characters are believable, cartoons are believable, but they may not be accurate”. To use them for decisions, you need to know whether a simulated person says what the real person would say. The cheap ways to define an agent (a few demographics, or a short narrative) produce stereotyped behaviour.

🎙 In the talk.

  • “Everything I’ve showed so far shows believability, but Disney characters are believable, cartoons are believable, but they may not be accurate.”
  • “How do you actually know?“. Replication papers “show a different result” when others try them.
  • A demographic agent is “my age, where I live, my job” and a persona agent is “more narrative descriptions of the person”.
  • Both “can produce very simplified and stereotyped behaviors”.
  • The example: from his former student, from South Korea. “What is he going to have for lunch? The model says rice. Not great.”
  • The alternative: “rich qualitative information is an interesting direction”.

📚 Deeper theory

  • A model fills the gaps with the average. Given little information, a language model completes the person from the most common pattern for that label: that is stereotyping as a statistical act, not malice. The fewer the facts, the more the label does the work. Bernstein’s “rice” is the classic case; the same dynamic produces the “far-right conservative” result in idea 5.
  • Stereotypes come in two flavours. (1) Wrong on average: the model’s picture of the group is off. (2) Right on average, wrong for the individual: a Korean person eats rice often, but this one had a sandwich. Variance collapse is the second kind: every simulated member behaves like the group mean. This is the central finding of the later critiques (see Part 1: Pew, NIM).
  • “Silicon sampling” came first. In 2023 Argyle et al. showed that GPT-3 conditioned on demographic backstories could reproduce voting patterns across groups (“Out of One, Many”, arXiv 2209.06899, cited in the 2023 Park page). Bernstein’s point is that this works at the group level and not at the individual level.
  • Personas are a design tradition. In product design, a “persona” is a fictional composite user (Alan Cooper, 1999) used to keep real users in mind. The generative-agent version turns a design artifact into an executable one.†

Where it is weak. Bernstein shows the contrast (demographic vs interview) but does not show what a good persona could achieve; the Anand page in this collection looks at persona engineering in more depth. And the “rice” story is an anecdote about one prompt. How much of this survives a stronger model or a better prompt is a testable matter.

📚 Roots: Stereotyping and the limits of group-level inference

The problem. From a label (age, nationality, party), how much can you say about one person? The honest answer is “some, with a wide margin”, and models tend to forget the margin.

The point. The interview fixes this because it replaces the label with evidence about the individual, which is the old distinction between nomothetic (group laws) and idiographic (the individual case) knowledge.

The chain of thinkers.

  • Wilhelm Windelband (1894): nomothetic vs idiographic sciences.†
  • Gordon Allport, The Nature of Prejudice (1954): stereotypes as over-generalised categories, and his later argument for studying the individual (personal dispositions).†
  • Walter Mischel (1968): situation changes behaviour more than trait labels predict; “personality coefficient” around 0.3. Relevant: predicting a person from a short description has a ceiling.†
  • The ecological fallacy (Robinson, 1950): group-level correlations need not hold for individuals; the mirror, the atomistic fallacy, is the reverse.†
  • Bias in language models (Bolukbasi et al., 2016; Bender et al., 2021): embeddings and models inherit societal biases from training text.†
  • Alan Cooper (1999), personas as design tools.†

What they leave out. Who counts as an “unusual” individual, and who is missing from the training data altogether. Bernstein’s solution depends on being able to interview people; groups that are hard to reach stay stereotyped.

Check yourself.

  1. Why does “rice” feel right and still fail?
  2. What is the difference between being wrong about a group and being wrong about a person in it?
  3. Which would you trust more for a new product: 5 demographic variables or 20 minutes of conversation? Why?

5. The 1,000-person study and the “two weeks later” yardstick

In short. The key experiment gives each of about 1,000 real Americans a two-hour interview, builds an AI twin whose memory is that interview, and then compares twin and person on surveys and experiments. Because people do not agree even with themselves over two weeks, the score is normalised by that self-agreement. On the General Social Survey, interview-based agents score about 0.85, against about 0.70 for agents given only demographics or a persona.

🎙 In the talk.

  • “We went out and ran two-hour interviews with a representative sample of 1,000 Americans.”
  • The script came from the American Voices Project and “literally starts with the question tell me the story of your life”.
  • “We use this two-hour interview to construct a digital twin of that same person… with that interview as the memory of the agent.”
  • The tests: the GSS (“a 170-ish question survey”), “the big five personality index, some behavioral economic games, some replications of experiments”.
  • The prompt is simple: put the interview first and say “based on this interview transcript, I’d like you to predict how this person would respond”.
  • The yardstick: “every result I’m going to show you here is going to be a ratio where 1.0 means that agents replicate people’s responses as accurately as people replicate themselves two weeks later.”
  • Scores: random guessing “about a third”; persona or demographic agents “about 70%”; “if you use these full interviews… 85% as well as they replicate themselves two weeks later on the General Social Survey”.
  • Other tasks: “these normalized coefficients are all generally in this similar range”.
  • Bias: “politics is the hardest to model”, “the difference between the best performing group and the worst performing group is around 8% normalized accuracy”, with “far right conservatives” worst; gender “sub 1% difference”.
  • Replication: “Of those five studies, our agents replicated four of the five.” The fifth also failed to replicate in the 1,000 real people.
  • Willer’s work: “100,000 people” and “a correlation between 0.85 and 0.9” on effect sizes.

📚 Deeper theory

  • Verified figures (from arXiv 2411.10109): 1,052 participants; on the GSS the interview agents reached a normalised accuracy of 0.85 (raw accuracy 68.85% divided by participants’ own two-week repeat accuracy of 81.25%), against 0.71 for demographic-only agents and 0.70 for persona agents. On the Big Five the normalised correlation was 0.80, and on economic games 0.66.
  • Check against the talk. He gives “about 70%” for both persona and demographic (matches 0.70 and 0.71) and “85%” (matches). He gives no separate numbers for the Big Five or the games, only that they are “in a similar range”; the paper’s 0.80 and 0.66 show the economic-game result is clearly lower than the GSS, so “similar” is generous. He says “1,000”, the paper 1,052. His “about a third” for random guessing is not checked here.
  • Why normalise. A test-retest ceiling is standard in measurement: no predictor can be more consistent with a noisy instrument than the instrument is with itself†. The 81.25% self-consistency is that ceiling. Hence 68.85% is “85% of the ceiling”, and a layperson may mishear it as “85% accurate”, which it is not.
  • The 80%-deleted result. He deleted a random 80% of an interview and the ratio fell “from 85%… to 0.79”. It suggests redundancy in a long interview, not that content does not matter (idea 6).
  • Politics as the hardest axis and “value aligned” models being “unhappy” with far-right answers is the speaker’s interpretation; the paper reports the gap. The reason (model alignment, interviewees being “coy”) is a hypothesis, not a result.
  • The Willer study is a separate paper by other authors, and tests a different thing: whether a model predicts the average treatment effect of an experiment, not an individual’s answer. Published headline: r = 0.85, 0.90 for unpublished studies (project page). It matches Bernstein’s “0.85 and 0.9”. Effects were overestimated, and a correlation can be high while still misjudging size.
  • The replication note is a good story and a small sample. Five studies, four replicated, and the fifth missed by people as well. Five is too few to say much about the method; it is an illustration.

Where it is weak.

  • Scope. US adults, two-hour private interviews, survey-type outcomes. Nothing here covers behaviour with money or consequences, other countries, or small groups.
  • Information in the interview. It is not clear how much of the agent’s skill is the model and how much is that the interview already contains answers to the survey items, since some GSS topics are discussed in the interview. Bernstein’s own remark about “relevant” content (idea 6) is the check.
  • Averages hide who is wrong. The pattern in later work: small differences by group in average accuracy can sit alongside a big loss of variance.

📚 Roots: Survey methodology and psychometrics

The problem. People answer the same question differently on different days, and asking a question changes the answer. What does it mean for a simulation to be “accurate” when its target is noisy?

The point. The “two weeks later” ratio is straight from measurement theory: a score cannot exceed the reliability of what it predicts. It is the reason a number like 0.85 can be both “good” and “not 85% right”.

The chain of thinkers.

  • Charles Spearman (1904): correction for attenuation. Observed correlations between two noisy measures understate the true relation; divide by the square root of their reliabilities. The normalisation resembles this.†
  • Classical test theory (Lord and Novick, 1968): observed score = true score + error; reliability is the share of variance that is true.†
  • Test-retest reliability (Cronbach and others, 1950s): correlate scores over time.†
  • Survey response theory (Tourangeau, Rips, Rasinski, 2000): answering involves comprehension, retrieval, judgement and response mapping, and each adds noise and context effects. A model that reads an interview and answers is another respondent, with its own response process.†
  • Open Science Collaboration (2015): only a minority of 100 psychology studies replicated with the original effect size. That is the background to Bernstein’s remark about the fifth, “bad science” study.†
  • Argyle et al. (2023), silicon sampling (checked above).†
  • Prediction-powered inference (Angelopoulos et al., 2023): combining many cheap model predictions with a few real labels to get honest confidence intervals. This is the formal bridge to the “validate on a small subsample” advice in idea 7; see the Angelopoulos page in this collection.†

What they leave out. Validity beyond reliability: a measurement can be reproducible and still not measure what you want. Predicting a survey answer is not predicting what a person does.

Check yourself.

  1. If a person’s own two-week repeat accuracy were 60%, what would “0.85” mean in raw terms?
  2. Why do the persona and demographic agents sit at about 0.70?
  3. What would you want to see before trusting that a result on the GSS tells you about customer purchases?

6. Building your own agent bank: data you need

In short. The study yields a practical rule for an organisation: the better the data you hold about each person, the better the agent, and the data has to be relevant to the question. A single label is bad; five or six demographics is “not terrible”; rich, relevant, first-person material is best. Interviews can be shorter than two hours with a modest loss.

🎙 In the talk.

  • “We have this agent bank that we can now use to test maybe depolarization or climate change kinds of questions”.
  • The design question: “who ought to be in that agent bank that you would build for questions that you care about? Should they be representative of your customers? Should they be representative of people who are not currently your customers, but could be?”
  • Existing assets: “Maybe your company has internal marketing or user research data… You just have a data set sitting there that you could turn into an agent bank.”
  • “A bad way to do this would be to define agents with a single demographic variable, just be a conservative, be a Republican. This produces all sorts of stereotyping and underestimates the variance.”
  • “Not terrible would be five or six demographic variables” at about 70%. “Better was data as rich as you can gather.”
  • The short-interview test: deleting a random 80% took the ratio “to 0.79. It didn’t really decrease it that much, which surprised me to be honest.”
  • The limit: “The remaining parts of the interview have to be relevant… if I’m only interviewing you about sports teams and then I’m trying to predict your retirement planning… a little bit too disjoint.”
  • “You have to make sure that the data you gather is relevant.”

📚 Deeper theory

  • This is the section most directly usable by a manager. It turns the research into a checklist: (1) choose who is in the bank; (2) gather first-person, in-domain material; (3) avoid label-only agents; (4) test.
  • Consent and privacy are a design constraint, not a footnote. He notes the interview data “remains private for human subjects reasons” and the bank holds “consented participants”. An organisation using customer data would need equivalent consent and data-protection arrangements. (†: the specific legal regimes)
  • Representative of whom? The study stratified a US sample. A company bank drawn from existing customers inherits that bias: it can say little about non-customers, which Bernstein flags (“not currently your customers, but could be”).
  • Redundancy explains the 0.79. People mention the same traits, values and constraints in many places, so a random fifth of an interview keeps most of the structure. Relevance explains the limit: what you did not talk about cannot be recovered.
  • A different design is possible. The 2026 Park page describes a company approach that adds behavioural and experimental data and updates the model itself; Bernstein’s advice here uses only prompt-level data.

Where it is weak. “Rich” and “relevant” are not defined. In practice the manager must guess in advance which questions the bank will be asked; the bank is tied to its original design questions.

Check yourself.

  1. A bank is built from customer-satisfaction surveys. What questions can it answer, and what can’t it?
  2. Why might a shorter interview lose little accuracy but a mis-targeted one lose a lot?
  3. Whom would you add to a bank to learn about non-customers?

7. Trust: the risk ladder, the 13% versus 1.2% error, and the funnel

In short. Bernstein is most careful here. He shows an error from a company that tried this, then sorts uses into four rungs by how much they ask of the model, and says how much evidence each needs. The sensible workflow is to use simulation to cut 100 ideas to five, then A/B test those five with real people.

🎙 In the talk.

  • “We cannot proceed without talking about the risks of this. There’s literature. Not every study replicates correctly.”
  • The case: a company’s simulation of how familiar people are with their retirement-plan fees. For the youngest group, “13% of people said they were very familiar… instead of 13% that one was 1.2%“.
  • “Maybe the difference between 13% and 1.2% is the difference between we ignore this sub-population and there are sizable minority that we would care about.”
  • The ladder: “as you climb up the ladder, you take on more risk of issues. But it also gets more ambitious.”
    1. Possibility: “What might happen? What could happen? We’re not going to put a probability label on it.”
    2. Qualitative: attitudes and “chat outcomes”, “safer because we’re pretty good at modeling this already, given the interviews”.
    3. Quantitative: “histograms, bar charts, where the difference between 5% and 10% might really matter”.
    4. Multi-agent: “town… an entire market simulation”.
  • Possibility “mostly generally works today” if the chain of events is plausible on inspection.
  • Qualitative “mostly works today if you have rich enough data… I would never say that you should do this instead of actually engaging with the communities”.
  • Quantitative: “consider you have 100 ideas… use simulation to narrow it down to five… And then actually a B test the remaining five on real people”.
  • Multi-agent: “I don’t think that we’re really ready for that in decision-making quite yet.” “Multi-agent can be right, but it also can be wrong. And you don’t know how to distinguish those.”
  • Mitigations: in-domain data in memory; use the rough-edged rungs; “validate important questions on a small subsample”.

📚 Deeper theory

  • Why the ladder is ordered by error tolerance. At the possibility rung the output is a hypothesis a human judges: false positives are cheap. At the quantitative rung the output is an estimate that the human acts on, so errors are costly. At the multi-agent rung the output is an emergent quantity that depends on every agent and every interaction, so errors compound and cannot be isolated.
  • The 13% versus 1.2% case shows two different things. The ranking of age groups was right; the magnitude was off by a factor of ten. For a decision about “which segment is largest” the model was fine; for “is this segment worth a product?” it was not. The same output can pass one test and fail another.
  • His “if individuals are accurate, the emergent outcome should be accurate” step is a hope, not a theorem. In complex systems, small errors in each part can be amplified (sensitive dependence), or can cancel; “phase changes” are the case where aggregate behaviour jumps†. He notes both views and declines to claim either. That is a candid epistemic position.
  • The funnel is a statistical idea. Use a noisy, cheap test to screen, and a precise, expensive test to confirm. It works if the cheap test does not discard the best idea too often. That is a recall question about the simulator that his talk does not measure. The 2026 Pew and Verasight tests (Part 1) are the nearest evidence on that question, and they are not encouraging for the quantitative rung. (partly checked: Pew and Verasight; the recall point is the page’s inference)
  • The rung structure maps onto the Anand page’s “forecast” idea: it works inside a regime, degrades beyond it, and must be checked against reality.

Where it is weak. The ladder is Bernstein’s judgement, not a tested typology. “Mostly works today” for qualitative attitudes rests on the individual-level results of the 1,000-person study; later work (Part 1) finds weaker individual-level prediction in other settings. Also, qualitative outputs from a model can be persuasive and therefore over-trusted by non-technical readers, which is the risk the rough-edged label hides.

📚 Roots: Validation and the epistemology of models

The problem. A simulation is a model, not the world. When is a model good enough to act on?

The point. Bernstein’s ladder is a practical version of an old rule: the use of a model, not the model alone, decides the validation needed.

The chain of thinkers.

  • George Box (1976/1987): “All models are wrong, but some are useful.” The question is for what.†
  • Karl Popper (1934): a claim is scientific if it can fail a test. The possibility rung makes no testable claim; the quantitative rung does.†
  • Joshua Epstein (2008), “Why model?”: sixteen non-predictive reasons (explain, guide data collection, illuminate dynamics, discover new questions, bound outcomes, train practitioners). This gives the possibility rung and the training use (idea 8) a respectable logic.†
  • Validation in ABM (Windrum, Fagiolo, Moneta, 2007; Grimm et al., “pattern-oriented modelling”, 2005): ways to check a model against several patterns at different scales.†
  • The Lucas critique (1976): a model fit to past behaviour fails when policy changes behaviour. For a what-if machine, an agent bank trained on how people did behave may fail in the new regime the decision creates.†
  • Goodhart’s law (1975) and performativity (MacKenzie, 2006): if many firms act on the same simulator, its use changes the market it models.†
  • Complex systems (Anderson, “More Is Different”, 1972): macro behaviour is not simply the sum of micro behaviour. This is why Bernstein hedges on the top rung.†

What they leave out. Who bears the risk of a wrong simulation. A manager who “narrows to five” bears it silently, because ideas the simulator rejected are never tested.

Check yourself.

  1. Why is “possible” a safer claim than “likely”?
  2. What can the 13% versus 1.2% example support and not support?
  3. If your simulator screens out the best idea, how would you ever find out?

8. Frontiers: look before you launch, rehearsal for conflict, and the company

In short. He ends with two uses where the what-if machine already helps: testing rules for online communities before launch, and practising a negotiation or conflict before it happens. He mentions the business activity around it, including the company that Stanford’s work spun out into.

🎙 In the talk.

  • “The set of tools around, look before you launch. So this is actually where we started.”
  • Platform designers say “everything we do is set in reaction to a dumpster fire”, and when given the tools “they actually iterated in simulation” and “fixed it before it launched”.
  • In his course: “I open the fire hose of trolls at their systems, and their goal is to make them stay upright.”
  • Soft skills: sparring partners “so that you can engage in this conflict or this negotiation… in simulation before you go and have the actual conflict”.
  • The experiment: people were randomised to watch a lecture, or the lecture plus a simulated conflict. “Both groups, critically, did equally well on a book test of what they’re supposed to do in a conflict, but only the group that had the simulation did better in an actual conflict later.”
  • “Having done the simulation reduced their likelihood of using an antisocial strategy by two-thirds.”
  • The business activity: the a16z article, and “the company that we’re spinning out is called Simile”. The close: “so much that we can learn from the opportunity to create this what if machine”, then a QR code for the course.

📚 Deeper theory

  • Why the platform case suits the possibility rung. Trolls and low-effort posts are a long tail of rare events. The aim is to find a way the system breaks, so a plausible chain of events is enough. This is a “red team” use.†
  • The conflict result is the strongest causal claim in the talk, because it is a randomised experiment with a behavioural outcome. It is a result about the training value of rehearsal, not about the realism of the simulated partner. The simulator only has to be a good enough sparring partner. A closely related Stanford paper with Bernstein as co-author describes LLM-based conflict roleplay with feedback (Shaikh et al., arXiv 2309.12309); I could not confirm it is the same experiment, and the talk does not say.
  • Knowing-doing gap. Passing a knowledge test without changing behaviour is a familiar finding in training research†. Both groups “knew” the strategy; only practice changed what they did.
  • The company. Simile is the spin-out. The 2026 Park page covers its funding and claims in detail, so this page only notes that, in the talk, it appears as “some embarrassment” at the end, and that claims by the company are the company’s own.

Where it is weak. The results come from a single experiment, summarised in a few lines. A two-thirds drop in antisocial strategy depends on what was measured, in what setting, for how long. And using simulated trolls to harden a system tests against the model’s idea of a troll.

📚 Roots: HCI design, prototyping and simulated rehearsal

The problem. Designers of systems with people in them cannot see the consequences until real people use them, by which time it is expensive to change.

The point. Bernstein’s “look before you launch” extends a design tradition of cheap prototypes that stand in for the real thing, here with synthetic users. His training example extends a second tradition: practising before the real event.

The chain of thinkers.

  • Donald Schön, The Reflective Practitioner (1983): designers learn by “conversation with the situation”; a sketch or model talks back. A simulation that reacts is a richer “talking back”.†
  • Wizard of Oz prototyping (Kelley, 1984): a human secretly plays the system to test an interface before building it. Social Simulacra replaces some of the secret humans with models.†
  • Participatory and speculative design (Scandinavian tradition, 1970s onward; Dunne and Raby, 2013): involve users and imagine futures. Bernstein himself says the simulation is no replacement for “engaging with the communities”.†
  • Online community governance (Ostrom’s design principles, 1990; Kraut and Resnick, Building Successful Online Communities, 2012): rules and moderation shape what communities become. The trolls case tests rules against abuse.†
  • Deliberate practice (Ericsson, 1993), flight simulators, and role-play in negotiation teaching (Harvard’s Program on Negotiation): skills change through rehearsal with feedback, not through lectures. The conflict result is a member of this family.†
  • Gelfand’s work on cultural tightness and looseness (2011) is the likely home of the “expert in conflict and negotiation” at the business school, though the talk gives no name.†

What they leave out. The cost of false confidence. A trainee who “wins” against an agreeable simulated partner may overrate their skill. Models’ politeness (flagged in the 2023 Park page) makes a poor adversary.

Check yourself.

  1. Why can a simulation help people learn even if its partner is not realistic?
  2. What would you measure to know whether your platform rules are better after the troll test?
  3. What does the equal score on the book test and unequal score on the real conflict show?