Catching lies with AI

People lie every day, from harmless white lies like calling in sick when you weren’t, to more serious deceptions that can end up in front of a court, but it can often be difficult to figure out if someone is telling the truth.

A group of scientists based at the University of Sharjah in the United Arab Emirates have found that AI might be the solution to this truth detection problem, but only if the machines can account for cultural and gender differences.

“Our objective was to conduct a comprehensive review of publications focusing on the computational prediction of deception, particularly using Machine Learning approaches,” the scientists write.

The team conducted a meta-analysis of 98 papers published from 2012 to 2023 which used Machine Learning-based approaches and deep learning algorithms to identify liars. 

In particular, the study focused on convolutional neural network (CNN) programs. These networks are a type of machine learning program that mimics the role of the visual cortex in the brain to recognise images. 

These papers were then compared to conventional approaches to lie detection that don’t involve using AI such as diagnostic questioning, expert analysis and using evidence. 

As part of the study, the team also analysed 35 short videos and two hours of footage.

“We conducted a comprehensive analysis on deception detection, providing a clear overview of the field contributions and limitations,” says the authors. 

Deception detection is a growing field with many scientists eager to learn more in hopes that it can lead to a more objective understanding of human behaviour. 

Accurate deception detection is also critical for areas in society like the legal system, where consequences for lying can be substantial. 

“Mistaking lies for the truth, or vice versa, in such situations can have significant repercussions for the individuals involved and society at large,” the authors note. 

When compared to the traditional methods of lie detection, AI and CNN programs had improved efficiency in spotting deceptive statements and lies.

“Over half of the papers achieved an accuracy exceeding 83%, a notably high performance highlighting the effectiveness of the machine and deep learning models employed,” write the authors.

“Overall, ML-driven approaches have shown comparable performance to traditional methods while achieving improved efficiency.”

While this may seem promising, their study also identified potential issues that may arise using AI-based deception detection methods.

One of the major limitations of using these algorithms to detect deception is their inability to consider the role of gender, culture and language. 

“Not considering culture, language and gender could limit the generalizability of the findings to diverse populations”, the authors write. 

The team outlined that this may be due to the small scale and lack of diversity within the datasets that many of these machine-learning programs are developed with. 

The team from the University of Sharjah’s research has been published in Expert Systems with Applications

Please login to favourite this article.