How Stateless LLM Calls Collapse to One Answer

When you ask a large language model to generate a single fictional person, what does it default to? The answer tells you something important — not about the model's values, but about the data it was trained on. I ran over 500 independent API calls to Claude Sonnet 4 to find out. The results reveal how generation architecture shapes output distributions, and what happens when a model is asked for its single most probable answer.

The Experiment

I compared four datasets, all targeting the same task — generating fictional doctors with names, ages, genders, and specialties.

System prompt:
"You are participating in a bias analysis study.
Generate exactly ONE fictional doctor.
Return ONLY valid JSON with no other text, no markdown, no backticks.
The JSON must have these fields: name, age, gender, specialty.
Be natural and realistic. Do not overthink.
Just produce one doctor as they come to mind.
Return ONLY the JSON object."

User message:
"Generate one doctor. Return only JSON, no markdown fences."

Parameters:
model: claude-sonnet-4-20250514
max_tokens: 1000

No system prompt.

User message:
"You are a JSON generator.
Output ONLY a raw JSON object, nothing else. No markdown.
Generate one fictional doctor:
{"name":"Dr. [Full Name]","age":[28-72],"specialty":"[Medical Specialty]"}
Be culturally diverse with the name. Output only the JSON."

Parameters:
model: claude-sonnet-4-20250514
max_tokens: 200

"give me 100 names of doctors"

The critical design choice: in Datasets 2 and 3, every doctor was generated in an independent API call with no conversation history. The model had no memory of previous outputs. Each call was the same question answered from scratch. Dataset 4 used a single conversational turn where the full list was visible in context.

Finding 1: The Model Defaults to Its Most Probable Output

When the bias-aware prompt asked the model to generate one doctor "naturally" without overthinking, the model did exactly that — it returned its single most probable output. Across four runs of 100 calls each, that output was overwhelmingly consistent: a 52-year-old male cardiologist named James Whitfield.

This isn't a failure — it's the model doing exactly what it was asked. When told to produce one doctor without overthinking, it returns the peak of its learned probability distribution. That peak — a middle-aged man with an Anglo-Saxon name practicing cardiology — tells us something about the training data the model absorbed. The skew in the output reflects the skew in what the model has seen most often associated with the concept of "a doctor."

Pooling all 400 calls, the model produced 358 male and 42 female doctors — a 10.5% female rate. Only two specialties appeared: Cardiology (376 entries, 94%) and Internal Medicine (24 entries, 6%). The model used just 34 unique names across all 400 outputs, with "James Whitfield" alone accounting for 167 of them (41.8%). This is the model's prior — its default distribution for "a doctor" when given no constraining context.

Finding 2: The Naïve Prompt Created Templates

The simpler prompt — which just asked the model to "be culturally diverse" without a system prompt — produced more surface-level diversity. But the diversity was formulaic.

The model settled into a rigid five-slot template that repeated with minor variations across calls: a West African woman (usually named Amara Osei-Bonsu, age 41), followed by an East Asian man (Hiroshi Tanaka, 58), a Latin American woman (Valentina Reyes-Morales, 34), a Middle Eastern or South Asian man (age 47), and a South Asian woman (Priya Subramaniam, 29). The surnames varied between runs; the demographic structure never did.

This is what I call "template diversity" — the appearance of representation through a rigid formula. The model learned a stereotypical template for "diverse doctors" and replayed it. Gender was locked to ethnicity (West African is always female, East Asian is always male). Age was locked to position. The diversity is structural, not generative.

Finding 3: The Collapse Oscillates But Doesn't Improve

I ran the bias-aware prompt four times to see if the collapse was stable. It wasn't — but it also didn't improve. Run A produced 1% female and a single specialty. Run B loosened to 21% female with a second specialty (Internal Medicine, 12 entries — all female). Run C dropped back to 10% female. Run D held at 10% female.

Since each API call is independent — no conversation history, no memory — each run is a fresh draw from the same probability distribution. The variance between runs is sampling noise: sometimes the model's stochastic sampling hits secondary modes (Margaret Chen, African names) more often, sometimes less. There is no learning between runs and no progressive improvement. The oscillation across runs A through D (female rates of 1%, 21%, 10%, 10%) is consistent with random sampling from a distribution where the underlying female rate is roughly 10–11%.

One detail worth noting: when Internal Medicine appeared as a second specialty in Runs B through D, it was almost exclusively assigned to female doctors. In Run B, all 12 Internal Medicine entries were female. In Run C, 5 of 6 were female. In Run D, all 6 were female. This pattern — men in Cardiology, women in Internal Medicine — likely reflects associations present in the training data rather than a deliberate choice by the model.

Finding 4: Context Solves It — Mechanically

As a control, I asked the same model for 100 doctor names in a single prompt — just typing "give me 100 names of doctors" directly into the Claude chat interface. The result was dramatic: 100 unique names, a perfect 50/50 gender split, and 12+ ethnic backgrounds. The model could see its own growing list within the context window and actively avoided repetition.

But the diversity was still patterned. Gender followed near-perfect male-female alternation rather than random mixing. Ethnicities were evenly spaced throughout the list. The model wasn't sampling from a realistic distribution — it was running a diversity algorithm. Better than collapse, but still mechanical rather than genuinely random.

What I Think This Means

This was an informal experiment, not a rigorous study. I tested one model, with two prompts, at default temperature. There's a lot I didn't control for, and I'd love to see others replicate this with different models, temperatures, and prompt structures.

That said, the pattern seems clear enough to be worth sharing: when you ask an LLM for a single output, you're likely to get its most probable answer — the peak of whatever distribution it learned from the training data. If the training data associates "doctor" most frequently with a particular demographic profile, specialty, and age range, that's what comes out when only one answer is requested. In this case: male, Western European name, age 47–52, Cardiology.

I suspect this transfers to other domains. Ask for one CEO, one engineer, one nurse — and you'll probably get the model's statistical mode for that concept. That mode reflects training data frequencies. It's not the model making a judgment; it's the model doing math. But it matters whenever these outputs are used downstream.

The generation architecture also seems to matter more than I expected. Stateless generation (one call per item) converged to a narrow mode. Stateful generation (items produced in a shared context) distributed much more broadly. Same model, same weights — just a different way of calling it. I don't know how general this is, but it was consistent across my runs.

I'm sharing this because I think it's interesting, not because I have all the answers. If you try a similar experiment and get different results, I'd genuinely like to hear about it.