Uncertainty quantification¶

Uncertainty quantification refers to methods for estimating and communicating how reliable or uncertain a model's predictions are. In misinformation detection, quantifying uncertainty enables systems to:

Defer to human review when confidence is low
Provide calibrated confidence estimates to downstream decision-makers
Improve decisions in high-stakes settings where knowing the uncertainty is as important as the point estimate

Methods¶

Confidence intervals: Construct intervals that contain the true value with probability \((1-\alpha)\)
Calibration: Ensure predicted confidence matches actual accuracy
Bayesian approaches: Model uncertainty via posterior distributions
Ensemble methods: Combine multiple models to estimate disagreement/uncertainty

In fact-checking¶

Model predictions in fact-checking need uncertainty estimates to support human-in-the-loop workflows. Amatya et al. (2026) uses prediction intervals to quantify uncertainty in claim check-worthiness detection.

Model calibration and confidence — complementary technique ensuring confidence matches accuracy
Prediction-Powered Inference — framework for valid confidence intervals with small labeled sets

Uncertainty quantification¶

Methods¶

In fact-checking¶

Related topics¶