A convincing persona is not yet a consistent person
Changing one detail at a time reveals how synthetic personas respond to personality and knowledge limits.
A synthetic persona can sound like a believable person. A harder question is whether its answers change in the right way when one detail about that person changes. We tested that question by creating pairs of personas that differed in just one line.
Change one property and compare the answers
The study used 32 synthetic people across four kinds of task and created 72 pairs of descriptions. Each pair differed in one detail, such as a personality trait or educational background. We asked both versions the same questions so we could look at the effect of that one change.
The comparison asked two things at once. Would an edited detail affect a question related to it? And would answers to unrelated questions remain reasonably stable? Both matter if a persona is meant to represent a consistent individual.
For example, changing a description from outgoing to reserved might reasonably affect an answer about seeking social situations. It would be harder to interpret if the same edit also changed unrelated answers. The paired design makes that wider effect visible.
Personality changes spread beyond the intended trait
An edit to extraversion changed the directly related answer in the expected direction for all eight pairs tested on that item. But the same edit also changed answers about openness, agreeableness and neuroticism. The change appeared to influence the overall impression of the person, rather than only the intended trait.
Knowledge edits behaved differently. Changing a person's educational and reading background changed none of 60 factual answers across the tested tasks. Some unrelated answers changed instead. Simply describing a person as lacking knowledge did not reliably limit what the model could answer.
| Test | Result | What it suggests |
|---|---|---|
| Question directly related to extraversion | 8 of 8 pairs changed in the expected direction | The targeted trait affected this answer |
| Other personality questions | Answers also changed | The effect was not limited to the edited trait |
| Factual answers after knowledge edits | 0 of 60 changed | A written knowledge limit did not control these answers |
Make access to knowledge part of the design
A description of a person and a limit on the information they can use are different things. A model may still answer from its training even when a prompt describes someone who would not know the answer. This gives us a more concrete design question: what should the simulated person actually be able to access?
One next step is to test what happens when a persona is actually given or denied access to a relevant fact, rather than merely described as knowing or not knowing it. Controlled sampling and unrelated questions would help separate the effect of that access from other changes.
The experiment shows why a plausible description alone is not enough to establish a consistent synthetic person. The relationship between a persona's background, available knowledge and decisions needs to be tested directly.