From Generic to Genuine: Seeding LLMs for Authentic Human Behaviour Simulations

As large language models (LLMs) grow increasingly capable of mimicking human reasoning, conversation, and decision-making, a new frontier is emerging: using them to simulate populations to explore human-like behaviours in dynamic contexts. Imagine being able to test a new marketing campaign, refine a user experience, or gauge opinions on policy changes—without conducting a time-consuming, expensive set of human interviews or surveys. With careful “seeding” strategies—providing models with demographic data, personality profiles, and real survey distributions, these simulations can reach levels of fidelity that mirror the diversity of real populations.

Yet, while LLMs are impressive at generating human-like responses, their default outputs often represent an “average” behaviour. They lack the complexity, variance, and demographic nuance characterising diverse societies. The solution lies in thoughtfully curating what we feed these models: demographic data, personality profiles, rich backstories, and calibration from real survey distributions. By doing so, we can transform a generic LLM response into a realistic reflection of a particular population segment—enabling more authentic simulations of human behaviour.

Why LLMs Often Lack Authenticity and Variance

LLMs are trained on massive corpora of text, learning statistical patterns of language that reflect common narratives, popular opinions, and mainstream cultural references. While this training enables them to produce coherent, context-aware responses, it also means that if left unstructured, LLMs tend to gravitate towards “safe” or median answers. Out of the box, they may give you something like a well-reasoned, but somewhat generic, reaction—failing to capture the perspectives you’d encounter in the real world.

For applications where understanding variability is key—predicting brand preferences across different age groups, assessing opinions on a product across various cultural backgrounds, or exploring political sentiment across income brackets—this convergence on the “average” won’t cut it. Without guidance, LLMs risk glossing over subtleties that real populations would display.

Seeding as a Strategy for High-Fidelity Simulation

This is where seeding comes into play. By “seeding,” I mean feeding the model with carefully chosen inputs that shape its output along desired lines. Instead of asking an LLM a broad question in isolation, you provide contextual layers:

• Demographics: Specify age, location, income level, education, or religious affiliation.
• Personality Traits: Incorporate frameworks like the Big Five (Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism) to give your simulated agent stable dispositions.
• Backstories: Generate narratives around an individual’s upbringing, career path, or life events, which in turn inform their worldviews. You can use LLMs to do this, too.
• Survey-Based Calibration: Embed known distributions of attitudes from real-world surveys—such as results from the World Values Survey (WVS), the American National Election Studies (ANES), or Pew Research—so that the model’s outputs reflect empirically grounded probabilities rather than purely invented patterns.

These methods push the LLM to produce responses aligned with specific population segments, making outputs more representative of how actual people in those segments might think and speak. They also help break away from overly generic outputs, enabling more accurate modelling of how individuals with distinct backgrounds might respond.

Designing Realistic Personas and Populations

Constructing simulated populations involves combining multiple seeding methods. Start with demographic and personality data: For instance, create a persona who’s a 32-year-old college-educated professional living in a mid-sized American city. Add personality traits, like high Openness and low Neuroticism, and layer on a backstory—perhaps this individual grew up in a working-class family, developed a passion for sustainable products, and became socially active in their neighbourhood.

Next, calibrate these traits with real-world survey distributions. For example, if surveys show that 40% of college-educated professionals in mid-sized American cities prefer a certain brand, your LLM persona should have roughly a 40% chance of expressing that preference. Draw on recent survey findings to calibrate how likely this persona is to prefer a particular political candidate, adopt environmentally friendly consumer behaviours, or choose certain media outlets. Incorporating probabilities from surveys ensures that your simulated personas don’t all share identical opinions; instead, they mirror the distributions found in real populations.

Use Cases and Applications

Marketing Research

Marketing teams frequently grapple with a fundamental question: “How will our target audience respond to this new campaign?” Seeding LLMs can help answer this at a fraction of the time and cost of traditional methods. By creating multiple personas that represent various segments—young urban professionals, mid-career parents in suburban areas, retirees with a moderate income—you can test messaging, branding, and product concepts on synthetic populations before investing in large-scale consumer studies.

In their survey of agent-based modelling and simulation, Gao et al. (2024) highlight how LLMs can model human behaviour at scale to inform decision-making. Similarly, initial research suggests that LLM-based agents can replicate participants’ responses with surprising accuracy. For example, Park et al. (2024) demonstrated “Generative Agent Simulations of 1,000 People,” showing that when LLM personas are shaped by interviews and demographic data, they can closely predict real individuals’ attitudes.

User Experience (UX)

Research in UX design, multiple user types interact differently with a product interface. Rather than recruiting dozens of testers repeatedly, you can simulate the behaviour of a diverse set of synthetic users. Some might be tech-savvy millennials drawn to sleek minimal interfaces, while others may be older users who need more guidance and prefer familiar navigation patterns.

By calibrating personalities and preferences, you can generate feedback about how each segment might respond to a new layout, feature placement, or onboarding tutorial. This rapid, cost-effective simulation process encourages iterative design. The work by Park et al. (2023) on generative agents showcases how well-structured LLM agents can engage in complex decision-making, making them suitable stand-ins for varied user groups.

Opinion and Attitude Research

For social scientists, policymakers, and opinion researchers, seeding LLMs offers a method to assess shifts in public sentiment without constant polling. By taking known distributions of attitudes from surveys, you can align the LLM’s responses with those real-world baselines. Then, tweak a particular stimulus—introduce a political candidate’s speech, a policy proposal, or a social narrative—and observe how the synthetic population’s opinions evolve.

Recent studies reinforce this potential: Jiang et al. (2024) explored how LLMs could simulate and predict public opinions in surveys, showing alignment with cultural differences and stable replication of known response patterns. See below the predictions of the election result for the 2024 U.S. Prediction Election by Jiang et al. (2024):

This approach can not only mirror existing survey data but also forecast election outcomes or policy acceptance rates, providing a low-cost, flexible supplement to traditional methods.

Practical Tips for Implementing Seeding Strategies

Getting started with seeding involves experimenting and refining your approach as you go. Consider these actionable steps:

• Incorporate Demographics: Incorporate explicit demographic attributes into your prompt. For instance, “You are a 40-year-old parent of two living in a rural community in the Midwest with limited broadband access. You earn about $50,000 a year and graduated from a two-year college program.”
• Assign Personality Traits: Introduce stable personality characteristics to guide how the persona interprets information: “You have high Agreeableness and low Openness. You trust established authorities and prefer familiar solutions over new, untested ideas.”
• Use LLM-Generated Backstories: Providing a short narrative fosters consistency and depth: “You grew up in a family that valued community traditions. You now work as a local librarian who organizes town events and relies on regional newspapers for your news. You have fond memories of your grandparents teaching you to garden in the backyard.” As mentioned before, you can use LLMs to do this for you — I’ll write about this another time!
• Use Real-World Surveys: Calibrate the persona’s opinions with known distributions from surveys like the World Values Survey (WVS) or Pew Research. Use randomness to your advantage and incorporate probability into your prompts. For example, if you know that 60% of people trust local newspapers over national media, make the coin toss with this probability to see if this particular simulated individual will trust the media. You can deep, using regression or machine learning to predict a specific person’s propensity for a specific attitude.

Experimentation is key.

After generating responses, compare them to actual survey results or known community preferences. Are the responses matching expected distributions (e.g., about 60% of the generated personas trust local news)? If not, adjust your prompts or the embedded probabilities and try again.

Once satisfied with a single persona, create multiple personas representing diverse demographics and personalities. This allows you to simulate entire populations. For example, you could run the same question by a dozen seeded personas—some older, some younger, some urban, some rural—and examine the differences in their responses.

Conclusion and Future Directions

Seeding LLMs enables a leap in the realism of human behaviour simulations. By tailoring inputs—through demographics, personality, backstories, and survey calibration—we can transform average-case LLM outputs into textured, realistic portraits of diverse populations. These seeded simulations are not just a laboratory curiosity; they can reduce costs, speed up user testing, improve the reliability of preliminary market research, and directly help opinion researchers test hypotheses or run experiments (see examples mentioned in Bail, 2024).

Ongoing research points toward even richer agent architectures, more sophisticated personality modelling, and advanced calibration techniques that reduce biases across racial and ideological lines. As LLMs improve, so too will our ability to create realistic simulations of complex social systems.

The message is clear for now: With the right seeds, we can grow simulations that help us understand and anticipate human behaviour more accurately than ever before.