Editorial illustration for Study finds base AI models predict human behavior better than fine‑tuned chatbots
Study finds base AI models predict human behavior better...
We train AI to be helpful. To follow instructions, to reason, to see. And in doing so, we seem to break its ability to think like a person.
A sprawling new study shows that raw, untuned AI models predict human choices and behavior far better than the polished, obedient chatbots we deploy. Every step we take to make them useful also makes them less like us.
A large-scale study shows that the training process turning raw language models into helpful chatbots also weakens their ability to mimic human behavior.
This is a trend, not an accident. As base models get smarter across generations, their fine-tuned counterparts fall further behind. We are systematically trading away a machine's capacity for human mimicry in exchange for utility.
The goal was a helpful assistant. The result is something that understands us less with every update, a polished facade that grows more alien the more we try to improve it.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives Chronos-2 uses known covariates such as weather for building demand forecasts
From Archives OpenAI upgrades GPT-5.5 readability, removes Canvas from Instant and Thinking
Common Questions Answered
Why do base AI models predict human behavior better than fine-tuned chatbots?
According to the study, the process of fine-tuning AI models to make them helpful and obedient actually diminishes their ability to think like humans and predict human choices. Each step taken to improve utility and instruction-following removes the model's capacity for human mimicry, making fine-tuned versions less aligned with how people actually think and behave.
What trade-off occurs when training AI models to be more useful?
Training AI to follow instructions, reason effectively, and be helpful requires systematic modifications that sacrifice the model's natural ability to mimic human behavior and decision-making. This means that as AI becomes more polished and obedient, it simultaneously becomes less like humans and less capable of predicting authentic human choices.
How does the performance gap between base and fine-tuned models change across AI generations?
The study reveals a concerning trend where fine-tuned models fall progressively further behind base models as AI systems become smarter across successive generations. This suggests that the problem of losing human-like prediction capabilities worsens with each advancement in base model intelligence, creating an increasingly alien gap between raw and refined AI systems.
What is the unintended consequence of improving AI assistants according to this research?
While the goal of fine-tuning is to create helpful assistants, the research shows that each update and improvement makes AI systems understand humans less effectively. The result is a polished facade that grows more distant and alien from human understanding the more developers attempt to enhance its capabilities and usefulness.
Further Reading
- New AI system can 'predict human behavior in any situation' with unprecedented accuracy, scientists say — Live Science
- Personality test shows how AI chatbots mimic human traits — and how they can be manipulated — University of Cambridge
- Study finds ChatGPT's latest bot behaves like humans, only better — Stanford University
- AI Can Seem More Human Than Real Humans in a Classic Turing Test Study Finds — UC San Diego Today
- Can Generative AI Chatbots Emulate Human Connection? A Review of Evidence and Open Questions — PMC
