what is selective abstention?

selective abstention is a classifier's option to decline a call. instead of being forced into real-versus-AI on every input, the detector can return a third outcome, “uncertain” or “i don't know,” when the evidence won't support a verdict. machine learning also calls this the reject option, selective classification, or selective prediction. the idea is old. C.K. Chow's 1970 paper in IEEE Transactions on Information Theory laid out an optimum rejection rule and showed that the error rate falls as you let a model reject more of its inputs.

the risk-coverage trade-off

abstaining buys reliability by answering less. El-Yaniv and Wiener formalized this for modern machine learning in 2010 (JMLR), naming the central dynamic the risk-coverage trade-off. coverage is the fraction of inputs you choose to answer. selective risk is the error rate measured over only those answered inputs. push coverage down by abstaining on the hardest cases and selective risk tends to fall with it. Geifman and El-Yaniv (NeurIPS 2017) carried the reject option into deep networks, defining a selection function that emits the prediction when it fires and “don't know” when it doesn't, then summarizing the whole behavior as a risk-coverage curve.

the gains can be large in the right setting, and they belong to specific studies rather than to detection in general. Geifman and El-Yaniv (NeurIPS 2017) reported that on top-5 ImageNet classification a selective model could guarantee 2% error with probability 99.9% while still answering almost 60% of inputs. that is an image-classification result from 2017, not an AI-content-detection number, and it does not transfer to verdicts on whether an image was generated. the honest takeaway is structural, not numeric. a model that can abstain answers less often, more reliably, on the cases it keeps.

abstaining versus deferring

abstaining and deferring are related, not identical. plain selective abstention withholds the answer and stops there. the theory of learning to reject (Cortes, DeSalvo, and Mohri, 2016) trains a classifier and a rejection function together, with consistency and calibration guarantees. learning to defer goes one step further: instead of only stepping back, the model routes the hard case to another decision-maker, often a human, and is trained accounting for that expert's accuracy and biases. Madras, Pitassi, and Zemel introduced it in 2018, and Mozannar and Sontag (ICML 2020) gave it a consistent surrogate loss. deferral adds an outside party to the loop. abstention does not.

the reject option has reached AI-image detection directly. Yumlembam et al. (Scientific Reports, 2025) built a rejection mechanism that scores each sample's uncertainty and rejects it when the score crosses a threshold tuned by particle-swarm optimization. on unseen generators, their combined uncertainty measure rejected roughly 70% of the samples the classifier had gotten wrong, filtering out most of the misclassifications. that figure is their specific experimental result on their own pipeline, not a general guarantee, and not a number any detector can promise.

the theory is honest about one catch. Chow's optimal trade-off assumes the true class probabilities are known. real detectors only estimate their own confidence, so the theoretical optimum is not directly reachable, and a miscalibrated confidence score can push a model to abstain on the wrong cases. this is why calibration matters as much as the abstention rule itself, and why the precise claim is “answers less, more reliably” rather than “more accurate” full stop.

this is the principle behind amige. returning “uncertain.” when the panel disagrees, or an image looks unlike anything the engine has characterized, say a generator it has not seen yet, a forced call would only produce a confident-sounding mistake. amige. routes each scan to the detectors strongest for it, runs a panel of independent detectors built by different teams, names the likely maker as a best guess rather than a fact, and recalibrates as new generators land. when the reads conflict it degrades to “uncertain” instead of averaging the disagreement into a falsely precise number. abstaining there is a feature, and every verdict stays a probabilistic estimate. see how the whole thing fits together in the machine.