Back to Glossary
AI Digital Twin
Synonyms: AI clone, digital human, synthetic persona, AI double, customer digital twin
Do not index
Definition
An AI digital twin is an AI model that acts for one real person, built from that person's data: interviews, survey answers, writing, or voice. The term comes from manufacturing, where twins track machines with live sensor data. In UX it means person-level twins, from a creator's AI clone to a research twin that predicts a customer's answers.
Use cases
Twins speak in a real person's name. When they're wrong, no one notices until someone acts on the answer.
- The research shortcut: A team surveys synthetic respondents in place of real product managers. 93% of synthetic PMs used UX terms, compared with 44% of real PMs (NN/g).
- Unclear identity: A fan chats with a creator's AI clone, assumes it's the creator, and follows advice the creator never gave.
- Reused data: Interviews collected for one study get reused to build twins for another, without the participants' agreement.
How it's used in practice
- Build from individual data: Stanford's twins, built from 2-hour interviews, matched survey answers 85% as well as people matched their own answers 2 weeks later. Demographic-only twins scored lower.
- Label the twin every time: Call it "Mai's AI assistant" in the first message.
- Get written consent with limits: Agree on what the twin may say and how to delete it. Microsoft requires written permission from voice talent for custom voices.
- Test against the real person: Ask the twin and the human the same 10 questions first.
Clean results are a warning sign. Synthetic respondents vary far less than real people. If every twin agrees, you've lost the outliers research exists to find.
Challenges & limitations
- Bias carries over: Twins repeat bias from their source data, though interview-based twins showed less.
- Twins get out of date: People change their minds, and the twin keeps the old views until someone rebuilds it.
Free resources
- Synthetic Users: If, When, and How to Use AI-Generated "Research" (NN/g) — where AI stand-ins help and where they mislead
- Digital Twins: Simulating Humans with Generative AI (NN/g) — Raluca Budiu on how twins differ from synthetic users and how they're built
- Evaluating AI-Simulated Behavior: 3 Studies (NN/g) — accuracy numbers from finetuned, interview-based, and synthetic-user studies
