Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2201.11260

To what extent should we trust AI models when they extrapolate?

Roozbeh Yousefzadeh·Xuenan Cao
View original source ↗Read source document ↗Export metadata ↓

Abstract

Many applications affecting human lives rely on models that have come to be known under the umbrella of machine learning and artificial intelligence. These AI models are usually complicated mathematical functions that map from an input space to an output space. Stakeholders are interested to know the rationales behind models' decisions and functional behavior. We study this functional behavior in relation to the data used to create the models. On this topic, scholars have often assumed that models do not extrapolate, i.e., they learn from their training samples and process new input by interpolation. This assumption is questionable: we show that models extrapolate frequently; the extent of extrapolation varies and can be socially consequential. We demonstrate that extrapolation happens for a substantial portion of datasets more than one would consider reasonable. How can we trust models if we do not know whether they are extrapolating? Given a model trained to recommend clinical procedures for patients, can we trust the recommendation when the model considers a patient older or younger than all the samples in the training set? If the training set is mostly Whites, to what extent can we trust its recommendations about Black and Hispanic patients? Which dimension (race, gender, or age) does extrapolation happen? Even if a model is trained on people of all races, it still may extrapolate in significant ways related to race. The leading question is, to what extent can we trust AI models when they process inputs that fall outside their training set? This paper investigates several social applications of AI, showing how models extrapolate without notice. We also look at different sub-spaces of extrapolation for specific individuals subject to AI models and report how these extrapolations can be interpreted, not mathematically, but from a humanistic point of view.

Explore related subjects

cs.LGcs.AImath.OC

Keep this discovery

BibTeXRIS

Roozbeh Yousefzadeh, Xuenan Cao. 2022-01-27. To what extent should we trust AI models when they extrapolate?. https://arxiv.org/abs/2201.11260

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

AUC Maximization from Biased Positive-unlabeled Data with Confidence

Maximizing the area under the receiver operating characteristic curve (AUC) is a standard approach to imbalanced binary classification. Although positive and negative data are required for maximizing the AUC, negative data are often difficult to collect in some real-world applications due to privacy concerns or the need for specialized expertise to annotate them. Thus, AUC maximization from positive and unlabeled (PU) data has been attracting attention. Existing methods assume that labeled positive data are unbiased samples from the true positive distribution. However, this ideal assumption is often violated in practice. In this paper, we propose a method to maximize the AUC from biased PU data. To address the bias, our key idea is to exploit {\it confidence}, i.e., the probability that an instance is positive, associated with the small number of labeled positive data. We derive an estimator of the AUC risk using biased PU data with confidence, enabling AUC maximization under such bias. We further show that the rewritten AUC risk induces a Bayes-optimal AUC ranking even when the available confidence is any strictly increasing transformation of the true posterior probability. We experimentally show the effectiveness of our method on eight real-world datasets.

cs.LG↗
arXiv2026

Measuring the Value of World-Model Updates: A Counterfactual Utility Protocol for Continual Adaptation

Continual world models must decide whether new data justify changing the model. Fixed replay schedules and prediction-error triggers specify when to update, but neither reveals the value of an individual update: one deployment run cannot show how the same model would have performed at that moment had it held its parameters. We introduce the fork ledger, which branches a deployment stream at pre-registered decision points into matched update and hold continuations under common random numbers. It evaluates both continuations on the same episodes and records $\Delta R = R_{\mathrm{update}} - R_{\mathrm{hold}}$. Always applying one fixed update mechanism lowers return on all three simulated control tasks: CartPole ($-144.0$; checkpoint-bootstrap $95\%$ CI $[-185.4,-116.1]$, against a converged return near $650$), Walker ($-82.8$; $[-101.1,-61.7]$) and Cheetah ($-18.6$; $[-29.0,-6.6]$). Divergence is an outcome of applying the update, so the estimand counts every attempted fork; restricted to the $693$ of $720$ that did not collapse, CartPole and Walker are unchanged in sign ($-113.4$ and $-82.1$) and Cheetah becomes unresolved ($-3.9$; $[-17.5,+13.0]$). The task is the unit of inference: each contributes $240$ attempted forks over five pretrained checkpoints crossed with two drift directions. The ledger makes counterfactual utility observable for a fixed mechanism, allowing triggers to be judged by the updates they select rather than by surprise detection alone.

cs.LG↗
arXiv2026

When More Is Not Better: Component Anti-Synergy in a P300 Speller

P300 brain-computer interface (BCI) spellers can provide hands-free communication for people with severe motor impairments. Modern pipelines combine multiple individually promising components, often assuming that 'more-is-better'. We tested this assumption using a four-component full-factorial experiment varying the inclusion of Euclidean Alignment (EA), xDAWN spatial filtering, subject calibration, and language model priors on a public P300 dataset. Performance was evaluated using accuracy, repetitions, and information transfer rate (ITR) with mixed-effects models. Results show that the value of components is conditional rather than additive. Calibration was the strongest singular contributor, while EA compensated for its absence in zero-calibration settings. Adding independently useful components could also reduce performance, revealing component anti-synergy. Contrary to conventional wisdom, LM support was not universally beneficial: its effect depends strongly on the strength of the underlying EEG pipeline, while results from a larger LM showed a similar pattern. Together, these findings challenge maximal 'all-on' pipeline design and highlight the value of selecting spatial and language-support components according to the quality of available EEG evidence.

cs.LG↗