Counterfactual Probe
What would have to change for a different answer?
An answer arrives with no sense of its stability: users can't tell whether a recommendation is robust or one assumption away from flipping. Counterfactual Probe exposes the levers — which inputs, moved how far, would change the verdict — and lets the user pull them: drag the budget 15% higher and watch the recommendation flip, with the boundary marked. Asking "why" gets a rationalization; probing "what would change your mind" shows the actual decision surface.
Framing
The problem
An answer arrives with no sense of its stability — users cannot tell whether a recommendation is robust or one assumption away from flipping.
The pattern
Expose the levers: which inputs, moved how far, would change the verdict — and let the user pull them, with the decision boundary marked on the control.
Why chat breaks here
Asking "why" in chat gets a rationalization and "what if" gets a fresh essay — neither shows the actual decision surface or where it flips.
Risks
Counterfactuals on fabricated reasoning are theater — the levers must be causally tied to the actual computation, or honestly labeled as approximations.
Avoid when
Creative or generative tasks with no decision boundary to probe.
Use when
Decisions rest on the answer and users need to know how fragile it is before acting on it.
DOPE evaluation
- Directability
- The user probes the decision directly — drag a lever, watch the verdict respond — instead of prompting for hypotheticals
- Observability
- Each input shows its distance-to-flip, so the fragile assumptions are visible before anyone acts on the answer
- Predictability
- The boundary is marked on the control: the user knows where the flip happens before crossing it
- Explainability
- Sensitivity is ranked — which factor actually drives this verdict — replacing prose rationalization with testable structure
Closest neighbours
- Google · What-If Tool (PAIR) (Google) — Probe an ML model's decisions by editing datapoints and finding counterfactual twins — the closest shipped implementation of pull-the-lever-and-watch-the-verdict, aimed at practitioners rather than end users.
- Fiddler AI · Counterfactual explanations (Fiddler) — Model-monitoring platform surfacing which feature changes would flip a prediction — decision-boundary visibility as an enterprise observability product.
- Adverse-action reason codes (US consumer credit (regulated)) — Credit denials must state the principal reasons — a legally mandated, decades-old counterfactual receipt ("what would have to be different"), proving the pattern predates AI.
FAQ
When should I use the Counterfactual Probe pattern?
Decisions rest on the answer and users need to know how fragile it is before acting on it.
When should I avoid the Counterfactual Probe pattern?
Creative or generative tasks with no decision boundary to probe.
What problem does Counterfactual Probe solve?
An answer arrives with no sense of its stability — users cannot tell whether a recommendation is robust or one assumption away from flipping.
Why is chat the wrong fit for this?
Asking "why" in chat gets a rationalization and "what if" gets a fresh essay — neither shows the actual decision surface or where it flips.
Related patterns
- Extends: Reasoning Trace — The trace shows the claimed path; the probe tests whether that path actually bears weight.
- Often paired with: Confidence Signals — Stated confidence vs measured stability — the probe checks whether the confidence is earned.
- Often paired with: Second Opinion — Vary the inputs vs vary the model — two independence tests on the same answer.