TWO ANSWERS, ONE VOICE
The Most Valuable Thing an AI Can Say Is "I Don't Know"
In a preregistered BCG experiment, the same GPT-4 interface helped on one task set and hurt on another.
- Inside the study frontier: significantly higher quality
- On one task selected outside it: 19% less likely correct
- The experiment measured outcomes, not model self-knowledge
SEVEN HUNDRED CONSULTANTS
The edge was drawn from outcomes, not appearance
Dell'Acqua, McFowland, Mollick et al. named the jagged frontier: apparently similar tasks can land on opposite sides of measured capability.
- 758 BCG consultants; three randomized conditions
- Inside: 18 tasks, 12.2% more completed, 25.1% faster, significantly higher quality
- Outside: one selected task, AI users 19% less likely correct
- A frontier in this experiment, not a universal capability map
The product does not label the edge
A smooth response is not a boundary signal. Without a tested refusal policy, the operator sees an answer, not which side of the evaluated frontier produced it.
NOTHING TO LEARN FROM
No evaluation support for the required capability
- Unsupported capability is an evaluation boundary, not automatically OOD
- OOD requires a case and a specified reference distribution
- Surface resemblance is not evidence that the required capability was tested
Call this an unsupported capability when relevant evaluations do not cover the required work. Use out of distribution only when a case is outside a specified reference distribution; neither label means absent from training data.
RISK AND UNCERTAINTY
Uncertainty must be taken in a sense radically distinct from the familiar notion of Risk, from which it has never been properly separated.
Frank Knight, Risk, Uncertainty and Profit, Part I, Chapter I; Online Library of Liberty PDF p.14 (1921)
Knight's economic distinction is an analogy here, not a machine-learning mechanism: when no reliable class or test supports an answer, responsible odds are unavailable.
THE WINS YOU CAN MEASURE
Abstention starts with a scope contract
- Customer support: 5,179 agents; issues per hour +14% overall, +34% for novices
- Scoped coding experiment: one JavaScript HTTP-server task, completed 55.8% faster
- Each result names the task, population, metric and check
These bounded wins remain real. The discipline is to state where the evidence holds, then pass the strategy question to the value that survives imitation.
FIND A VERIFIER
Answer only where the check can answer back
- 01Generate a candidate
- 02Run an external check tied to the claim
- 03Answer only if the check passes
- 04If no adequate check exists, abstain or escalate
A verifiable space has a test outside the model. Abstention marks the branch where that test is absent or fails; it is not itself proof that the model knows why.
TWO ENEMIES OF THE CHECK
A refusal policy can fail in two directions
- False acceptance: the system answers and is wrong
- False refusal: the system abstains when it could answer correctly
- Measure both on held-out tasks, by domain and model version
- Tune the threshold to the cost of each error
Abstention is a decision rule, not a virtue. A system that never refuses hides risk; one that always refuses has no coverage.
THE MAP NO ONE SHIPS
Ship the risk-coverage curve
Geifman and El-Yaniv formalised the reject option as a risk-coverage tradeoff. Their classifier result supplies a design pattern, not proof that a language model is calibrated.
- Coverage: the share of evaluated cases the system answers
- Selective risk: the error rate among those answers
- Lower answered-case risk generally costs coverage on the evaluation set
- Choose a threshold from the harm budget; retest after model or domain changes
THE COORDINATE
Abstention is a coordinate only after calibration
No validated boundary, no autonomous handoff. Past the tested edge, prediction stops and optionality begins.
- Set the acceptable error rate and minimum useful coverage
- Validate the refusal threshold on representative held-out tasks
- Route refusals to a named human, source or tool; log outcomes and retest each version