Grounded Simulation: faithful, not just fluent
LLM “synthetic users” produce fluent interviews that teams find easy to distrust. Grounded Simulation is a first-principles architecture for keeping them faithful, and auditable.
01What Grounded Simulation is
Grounded Simulation is an architecture for LLM-based synthetic user research: simulated participants grounded in behavioural science, interviews kept hypothesis-blind by construction, and every finding traced back to interview evidence before a report may claim it. I coined the term in an April 2026 SSRN working paper, Grounded Simulation: Task-Preserving LLM Personas for Synthetic User Research (doi:10.2139/ssrn.6503241), which is a preprint and has not been through peer review.
Two nearby terms get confused with it. Grounded theory is a method for analysing qualitative data that humans produced, and it shares nothing here beyond a word. "Grounded user simulation" in agent benchmarking means an agent tied to a task environment so its actions can be scored, which is a different problem with a different pass mark. This architecture asks what constrains a simulated participant before it speaks, and what you can check after it does.
02The fluency trap
Ask a model to play a user and it will talk for as long as you let it. Transcripts come back readable, quotable, easy to paste into a deck. That is the danger. A synthetic interview that sounds right does more damage than one that visibly falls apart, because somebody will ship on the strength of it.
Three failure modes keep recurring in this kind of work. Prime the model with your hypothesis and it hands your hypothesis back. Ask for a persona and you get a caricature rather than somebody contradictory and situated. Then a stakeholder asks how you know, you have a transcript instead of a chain of evidence, and the room either throws the work out or believes it for the wrong reason.
The goal isn't a model that sounds like a user. It's a study you can audit in thirty minutes.
03Hypothesis-blind personas, and what they cost you
Say you want to know whether delivery charges are why people abandon your checkout. Run that synthetically the ordinary way and you write a persona prompt that already carries the answer: a price-sensitive shopper who watches shipping fees. You interview it. It talks about shipping fees. What you have measured is your own prompt.
Generating personas blind cuts that loop before the interview starts. Whatever builds your participants never sees the question under test. It sees who you are studying, rendered through published instruments rather than adjectives you chose: trait facets, cultural dimensions, a memory model that makes recall decay the way human recall decays. Your hypothesis stays in a different file. The interview meets the persona cold, and if delivery charges still come up, they come up because of who the participant was built to be rather than because you told the model what you were hoping for.
That costs you something, and it is worth being honest about which thing. You can no longer nudge a persona back on track when it says something inconvenient, because that nudge is exactly what contaminates the study. You will also get themes you did not order, and some of them will be about your onboarding when you asked about checkout.
You can copy the discipline without any of the architecture around it. Build your participants in one session, write your hypothesis in another, and never let the second file touch the first. That takes an afternoon to set up and it closes the most common route to a self-fulfilling study.
04What an audit actually looks like when you open it
Open a theme in a finished study and you do not find a tidy line of prose. You find the quotes that produced it, the interview turns that pulled those quotes out, the hypothesis that was live at the time, a count of how many times the theme was refuted rather than supported, and how confident the system was, computed instead of asserted.
Read the refutation count first. A theme that participants pushed back on almost as often as they endorsed it means the people you are studying disagree with each other, and a report that summarised it as a finding has quietly deleted the interesting part. That count is the cheapest signal in the whole artifact and the one people skip.
The pipeline I work on at Articos runs more than 100 substages and 1,669 passing tests. On its own that number just asks you to be impressed, and you should not be. The version that tells you anything is much smaller: one test fails any theme that cannot be linked to a transcript span. Themes cannot enter a report on the strength of sounding true, and a number like 1,669 is only the honest cost of building that many places to inspect.
05Auditability is not validity
Here is where the architecture stops. Provenance tells you what produced an output. An audit trail tells you whether somebody other than the author can inspect that. Whether the claim is true, representative and useful is a third thing, and the first two settle none of it. A theme can trace perfectly to a real transcript line and still describe nobody who exists.
So the promise is narrower than it sounds, and narrower is the point. Grounded Simulation does not make a simulated participant true. What it makes is a study you can open, disagree with in specific places, and correct, which is the only way a synthetic method improves rather than piling up confident mistakes. A study nobody can open is not wrong. It is unfalsifiable, which is worse.
If you are weighing a synthetic study against a real one this quarter, the useful thing to ask is not which one sounds better. Work out what you would need to see inside the audit before you would let it change a roadmap, then check whether the tool in front of you can show you that at all.
06Status and reading
The formal write-up is live on SSRN, abstract ID 6503241, and it is a working paper rather than a reviewed paper. Treat this page as the plain-language version. The paper sits with my other work on the research page, the architecture running in production is in the Articos case study, and the broader case for inspectable systems is in auditable AI research.
07References
- Grounded Simulation: Task-Preserving LLM Personas for Synthetic User Research
- Out of One, Many: Using Language Models to Simulate Human Samples
- Using Large Language Models to Simulate Multiple Humans and Replicate Human Subject Studies
- CoMPosT: Characterizing and Evaluating Caricature in LLM Simulations
- Towards Understanding Sycophancy in Language Models
- On the Social Psychology of the Psychological Experiment
- Blind analysis: Hide results to seek the truth
- Evaluating Large Language Models in Generating Synthetic HCI Research Data: a Case Study
- "Simulacrum of Stories": Examining Large Language Models as Qualitative Research Participants
- LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals
- τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains
- PROV-DM: The PROV Data Model
08Further study
Branching by what you are actually trying to do next.
If you are about to run your first synthetic study.
- Blind analysis: Hide results to seek the truth — the afternoon-sized discipline section 03 says you can copy; physicists have run it for decades.
- Towards Understanding Sycophancy in Language Models — know the leak you are closing before you close it; the model agrees with you by training, not by accident.
- Evaluating Large Language Models in Generating Synthetic HCI Research Data: a Case Study — what fluent synthetic data tempted researchers to do in 2023; the temptation has not changed.
If you are deciding whether to believe one.
- "Simulacrum of Stories" — the strongest case against synthetic participants, from qualitative researchers themselves; read before trusting any vendor.
- Using Large Language Models to Simulate Multiple Humans and Replicate Human Subject Studies — hyper-accuracy distortion: a simulation can fail by being too correct, which no fluency check catches.
- Artificial intelligence and illusions of understanding in scientific research — Messeri & Crockett, Nature 2024; believing a result for the wrong reason, at the scale of science.
If you are building the personas themselves.
- Out of One, Many — condition on real human data, not adjectives; where algorithmic fidelity was defined and tested.
- LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals — interviews as grounding evidence, the deepest version of instruments-over-adjectives.
- Replication and Analysis of Ebbinghaus' Forgetting Curve — Murre & Dros, PLOS ONE 2015; the shape human recall decay actually has, if your memory model claims to match it.