The survey of the trustworthy-machine-learning field left one thread deliberately hanging: the constructive one. Most of that literature explains why models assert too freely, or measures how often they are wrong. But a smaller and more interesting strand asks the opposite question. Instead of trying to make a model right more often, it asks when a model should decline to answer at all. That research turns out to be the machine-learning field's own approach to the problem at the centre of this work, arrived at independently and from a different direction, and it is worth following properly rather than leaving as a footnote.
Selective prediction: a model that can decline
The first idea is selective prediction, also called classification with a reject option. A selective model is permitted to abstain: on an input where it is likely to be wrong, it may return no answer rather than a forced guess. The value of this is captured by what the field calls the risk-coverage tradeoff (Geifman and El-Yaniv, 2019). Coverage is the fraction of inputs the model chooses to answer; risk is its error rate on those it does answer. The two trade against each other: allow the model to abstain more, and the error rate on what remains falls. A selective model can therefore offer a guarantee that a model forced to answer everything cannot, namely a bounded error rate on the cases it accepts, at the cost of leaving some cases to someone else.
The relevance to this work is direct. Selective prediction is, in the machine-learning register, the recognition that not every claim should be acted on, and that a system is safer when it can mark some of its own outputs as ones it should not stand behind. That is the same instinct that motivates marking a claim's basis: the difference between a system that emits every output in the same confident voice and one that can distinguish the outputs it will vouch for from the ones it will not. Where selective prediction stops short is that its decision is still made purely on the model's own terms, from the model's own estimate of its own likelihood of error. It knows when to be quiet, but it does not yet ask what the silence or the answer is for.
Learning to defer: the decision depends on who receives it
The second idea closes exactly that gap, and it is the more striking of the two for this work. Learning to defer extends abstention into something relational. A model that learns to defer does not simply decline when uncertain; it hands the decision to a downstream expert, and, crucially, it makes that choice by weighing its own expected performance against the expert's (Mozannar and Sontag, 2020). It should predict when it is likely to do better, and defer when the expert is likely to do better. The deferral is not a function of the model's uncertainty alone. It is a function of the comparison between the model and whoever, or whatever, receives the handoff.
This is a significant move, because it makes the right action depend on the receiver. The same input, with the same model uncertainty, may warrant a confident answer in one setting and a deferral in another, according to who is downstream and what they are equipped to do. The machine-learning literature reached, on its own and for its own reasons, the proposition this work builds its discipline around: that whether an output should be acted upon is not a property of the output in isolation, but of the output in relation to the next decision-maker and the action they are about to take.
Why this matters for what follows
These two threads are the field's own partial answer to the failure the mechanism argument described. If generation is optimised to assert rather than abstain, selective prediction and learning to defer are the constructive correction: mechanisms by which a system can decline to assert, and can decide to defer based on what happens next. They are encouraging company for this work, because they show the central intuition is not idiosyncratic. A mature body of research independently concluded that the safe thing is to let a system withhold, and to condition withholding on the downstream consequence.
But they also mark the boundary of what that research provides, and therefore the room this work has to occupy. Selective prediction and learning to defer are training-time constructions internal to a model: they shape a single model's behaviour, and they operate on that model's own estimates. Neither carries a claim's basis across a boundary to a party that did not train the model and cannot see its internals. Neither survives the derivation of one claim from another, or the handover of a value from a system that can defer to one that cannot. They answer the question for the model that produced the claim; they do not answer it for the owner who receives it, later, cold, with only the value in hand. That receiver is where this work begins, and the fact that the machine-learning field has already validated the underlying instinct, without extending it across the boundary, is precisely the opening the discipline is built to fill.
References
Geifman, Y. and El-Yaniv, R. (2019). SelectiveNet: A deep neural network with an integrated reject option. Proceedings of the 36th International Conference on Machine Learning. arxiv.org/abs/1901.09192
Mozannar, H. and Sontag, D. (2020). Consistent estimators for learning to defer to an expert. Proceedings of the 37th International Conference on Machine Learning. arxiv.org/abs/2006.01862