Your AI twin may be a stereotype with your name on it.
Digital twins built from hundreds of answers still struggled to reproduce the people they were meant to simulate.
Published
Evidence boundaries Solid · Inference · Speculation
In 19 preregistered studies across 164 outcomes, LLM-based digital twins trained on each participant’s prior answers correlated only weakly with the humans they modeled and showed systematic distortions.
A model that sounds personalized can still be compressing you toward population-level stereotypes.
As personal AI agents become more persistent, users may begin adapting themselves to the simplified versions their systems reflect back.
A personal model can know a lot about you and still miss the person.
Researchers trained digital twins on more than 500 prior answers from each participant, then asked the twins and the humans new questions across attitudes, judgments and behavioral intentions.
The twins were only modestly more accurate than the underlying base model and averaged a weak correlation with the humans they were supposed to represent.
The errors were not random. The researchers identified recurring distortions including insufficient individuation, stereotyping, representation bias, ideological bias and hyper-rationality.
Personalization can feel intimate long before it becomes accurate.
We may soon be represented by models that are persuasive before they are faithful.
AI twins are already being discussed for research, policy, customer simulation and personal assistance.
A system does not need to understand you deeply to produce language that feels like you. That gap between fluency and fidelity may become one of the defining design problems of personalized AI.
Think with the idea.
Take me sideways
This resembles recommendation systems. A model can become useful by predicting a compressed version of you, then gradually reinforce that version through what it chooses to show.
Cross-domain analogy, not a direct finding of the digital-twin studies.Challenge it
Weak performance today does not prove digital twins are fundamentally doomed. Better memory, richer behavioral data and continuous calibration could improve fidelity substantially.
Future-performance counterpoint.Trace it
Follow the idea from psychometrics to recommender systems, synthetic respondents, customer personas and persistent AI agents.
Conceptual lineage.Make it useful
When an AI system claims to know you, ask what evidence it has, how often it is recalibrated and which parts of your behavior it is structurally unable to observe.
Practical epistemic prompt.