Automated verbal lie detection has accuracy limits around 75 percent
A persistent accuracy ceiling in automated verbal deception detection
Computation and Language
Summary
Detecting when someone is lying by analyzing their words has been a big challenge. The authors looked at 25 years of studies that tried to make computers better at spotting lies from what people say. They found that the best accuracy is about 74%, and this hasn’t really improved even with new machine learning technologies. Most studies didn’t use truly verified lie or truth labels, and often the tests weren’t done on completely separate data. This suggests that current approaches face a natural limit in how well they can detect deception from speech alone.
What this means in practice
- •For law enforcement analysts: Evaluate the practicality and limits of automated lie detection tools based on speech analysis for investigative support.
- •For customer service quality teams: Assess the reliability of automated speech-based deception indicators to inform training and compliance monitoring.
A survey. It maps existing work.
Authors
Riccardo Loconte, Jonas Festor, Zane Fatjanova, Mariam Bolkvadze, Bennett Kleinberg
Abstract
Automated methods have been proposed to overcome the limitations of human verbal deception detection, but evidence remains fragmented across disciplines. We systematically reviewed 25 years of research (289 reports, 6,136 classification models) and meta-analyzed 3,653 models nested within 97 datasets. Pooled accuracy was 74.4% (95% CI: 71.2%-77.4%) with substantial heterogeneity. Accuracy was driven by methodological quality (ground truth, data source, class balance, evaluation procedure) more than by model complexity: the adoption of embeddings and large language models has not translated into improved predictive performance. Only 12.46% of reports used data with verifiable ground-truth, and only 23.96% of models were evaluated on independent data. The pooled accuracy aligns with meta-analyses of manual approaches, suggesting a ceiling of 70-75%, unlikely to be lifted by current research conventions.