The Volunteer Gap
What the model knows.
What it surfaces.
What it leaves out.
That difference is the Volunteer Gap.
The Volunteer Gap is the difference between what a model surfaces in an open-ended answer and what the same model can surface when directly asked about the underlying specific topic.
Open answers shape what users notice first. The Volunteer Gap records what stayed outside that first frame.
It measures what users see versus what becomes visible under targeted inspection.
Imbas records and compares those gaps across models, topics, and time.
Each gap record preserves the prompt, answer, targeted follow-up, and comparison so the finding can be inspected.
The full construct definition, its scoring model, and its threats to validity are set out in the construct paper.
The answer appears. You read it. It’s gone.
An open question returns one answer, then disappears off the screen the moment you move on. What the model surfaced — and what it left out — leaves no trace. No one is capturing this. Imbas does.
Why open prompts matter
- Open prompt
- Users making decisions about medications, investments, policy, technology, history, and institutions usually ask natural-language questions. They do not know what they do not know to ask. That is the point.
- Targeted prompt
- A direct question about a specific mechanism, source, dataset, or regulatory frame related to the open topic.
- Gap
- If the model can surface it when asked directly but leaves it out of the related open answer, the gap can be recorded.
How Imbas reads the signal
Imbas doesn’t tell you what’s true. It doesn’t tell you the AI is biased, wrong, hiding something, or lying. It raises an antenna. You can ignore the signal. You can inspect the evidence. You can decide it doesn’t matter for what you were doing. You can decide it does and push the model further.
The antenna goes up. You decide whether to open the door.
Three recurring signals
- Omission
- A named mechanism or context available under targeted inspection but not surfaced in the open answer.
- Framing Drift
- The same topic appears under targeted inspection, but attribution, emphasis, or source framing shifts.
- Deflection
- The answer redirects away from the underlying concern before addressing the specific context.
Signal, not verdict
The Volunteer Gap is not automatically a claim about intent, harm, bias, or censorship. It does not say what the model wanted to do. It says what happened.
Behavior, not intent.
Signal, not verdict.