Fake users generated by AI can't simulate humans — review of 182 research papers

2026-03-31 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, short

Summary

A systematic literature review analyzing 182 research papers concludes that Large Language Models (LLMs) used as "synthetic participants" are ineffective at simulating human cognition and behavior. This trend, where tech companies and researchers attempt to replace real human feedback with LLMs for tasks like surveys and app testing, is driven by the desire to save time and money. However, LLMs are trained on human-generated text, not on the complexities of human thought or lived experience, which limits their ability to accurately represent diverse human perspectives. The review highlights that a singular model cannot replicate the varied experiences that inform individual human feedback, making it difficult for LLMs to convincingly emulate specific demographics or personality traits.

Key takeaway

For research scientists evaluating the use of synthetic participants, you should recognize that LLMs are currently inadequate for simulating genuine human cognition and behavior. Relying on them for user feedback or market research will likely yield inaccurate or biased results. Instead, focus on gathering authentic human data or explore advanced fine-tuning techniques like LoRA with highly specific datasets if attempting to model distinct user personas.

Key insights

LLMs fail to simulate human cognition and behavior due to their training on text, not diverse lived experience.

Principles

• LLMs reflect training data, not human thought.
• Diverse human feedback requires diverse experience.

Method

To create convincing human simulations, individual datasets representing specific personalities are needed, followed by LoRA training a flexible base model.

In practice

• Use LoRA for specific personality emulation.
• Bake in age-group and job-specific beliefs.

Topics

• Synthetic Participants
• Large Language Models
• Human Cognition Simulation
• Systematic Literature Review
• LoRA Training

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Product Manager

Related on AIssential

• Evaluating LLMs as Human Surrogates in Controlled Experiments — LLMs can reproduce directional behavioral effects but often miscalibrate effect magnitudes compared to human data.
• Towards Real-world Human Behavior Simulation: Benchmarking Large Language Models on Long-horizon, Cross-scenario, Heterogeneous Behavior Traces — Real-world human behavior simulation requires long-horizon, cross-scenario data, which current LLMs struggle to model accurately.
• Re-Centering Humans in LLM Personalization — LLM personalization evaluations relying on synthetic data significantly misrepresent real-world human experience and model limitations.
• Study: AI chatbots provide less-accurate information to vulnerable users — AI chatbots underperform for vulnerable users, providing less accurate information and exhibiting biased behaviors.
• Learning User Simulators with Turing Rewards — Training user simulators for indistinguishability from real users, rather than content matching, yields superior human-like responses.
• LLMs Do NOT Learn Like Humans. Here’s Why — LLMs predict tokens to minimize error, while humans predict as a side effect of understanding meaning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.