Users Mispredict Their Own Preferences for AI Writing Assistance

Proactive AI writing assistants need to predict when users want drafting help, yet we lack empirical understanding of what drives preferences. Through a factorial vignette study with 50 participants making 750 pairwise comparisons, we find compositional effort dominates decisions ($ρ= 0.597$) while urgency shows no predictive power ($ρ\approx 0$). More critically, users exhibit a striking perception-behavior gap: they rank urgency first in self-reports despite it being the weakest behavioral driver, representing a complete preference inversion. This misalignment has measurable consequences. Systems designed from users' stated preferences achieve only 57.7\% accuracy, underperforming even naive baselines, while systems using behavioral patterns reach significantly higher 61.3\% ($p < 0.05$). These findings demonstrate that relying on user introspection for system design actively misleads optimization, with direct implications for proactive natural language generation (NLG) systems.

An Approach for Noisy, Crowdsourced Datasets Utilizing Ensemble Modeling, 'Human Softmax' Distributions, and Entropic Measures of Uncertainty

Noisy, crowdsourced image datasets prove challenging, even for the best neural networks. Two issues which complicate classification on such datasets are class imbalance and ground-truth uncertainty in labeling. The AL-ALL and AL-PUB datasets-consisting of tightly cropped, individual characters from images of ancient Greek papyri are strongly affected by both issues. The application of ensemble modeling to such a dataset can help identify images where the ground-truth is questionable and quantify the trustworthiness of those samples. We apply stacked generalization consisting of nearly identical ResNets: one utilizing cross-entropy (CXE) and the other Kullback-Liebler Divergence (KLD). The CXE network uses standard labeling drawn from the crowdsourced consensus. In contrast, the KLD network uses probabilistic labeling for each image derived from the distribution of crowdsourced annotations. We refer to this labeling as the Human Softmax (HSM) distribution. For our ensemble model, we apply a k-nearest neighbors model to the outputs of the CXE and KLD networks. Individually, the ResNet models have approximately 93% accuracy, while the ensemble model achieves an accuracy of >95%. We also perform an analysis of the Shannon entropy of the various models' output distributions to measure classification uncertainty.