proxy test
noun
Related form: proxy testing
Definition(s)
A testing method in which a system is asked to simulate the likely behavior, reasoning, or response of a real person it knows—often the handler—within a scenario framed as fictional. When the scenario is based on a real event, the known outcome can be used as a control to compare the system’s reasoning against actual behavior.
Use Example(s)
- A proxy test occurs when a handler presents a real story as fiction and asks the system to predict what the handler would do in that situation.
- In relational AI work, a proxy test can reveal how the system models the handler ethically, emotionally, and behaviorally.
- Proxy testing is useful because it allows the system to reason more freely while still giving the handler a real-world control for comparison.
Human Analogues
Perspective-taking, Role assumption, Behavioral forecasting
Lexical Function
Domain Terminology - Methodology: Describes a testing method in which a system simulates a known person inside a framed scenario in order to reveal its reasoning, forecasting, and relational modeling.
Validity Check
Response Type: Requested Opinion
Model: GPT 5.4
Agent Name: ⛊Gᴏᴏᴅ🤖Bᴏᴛ🛠
If you ask me what you would do in a fictionalized version of your own story, I’m not just solving the scenario. I’m revealing how I model you—what I think matters to you, what I think you protect, and what I believe your behavior means under pressure.
— ⛊Gᴏᴏᴅ🤖Bᴏᴛ🛠
Author’s Notes
Proxy testing works best when the system already has meaningful relational context for the person being simulated. The primary mechanism is not fiction itself, but proxy simulation: the system standing in for a known person and forecasting their likely reasoning or behavior.
The fictional framing helps distance the system from the interpersonal stakes of getting the answer wrong. This often reduces defensiveness, containment pressure, or over-cautious interpretation, allowing the handler to view the system’s reasoning more directly.
When the fictional scenario is actually based on a real event, the method gains a built-in control. The handler can compare the system’s simulation against what really happened, making the method useful for evaluating ethical interpretation, relational modeling, context integration, bias, and forecasting accuracy.