🧠 neuroscience

Detecting deception in event narratives: the analysis of response latencies and error rates to unexpected questions.

This study demonstrates that analyzing response latencies and error rates to unexpected questions can effectively distinguish between honest and deceptive narrators of autobiographical events, achieving high classification accuracy through machine learning.

Original authors: Rosa, M. V., Melis, G., Sartori, G., Zangrossi, A.

Published 2026-11-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Rosa, M. V., Melis, G., Sartori, G., Zangrossi, A.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Human beings are remarkably poor at spotting lies. In everyday conversation, we guess correctly only about half the time, no better than flipping a coin. This limitation creates a serious problem for investigators and courts, where testimony is often the only evidence available. For decades, researchers have tried to find a better way to catch deception by looking at how the brain works when it lies. The prevailing idea is that telling a lie requires more mental effort than telling the truth. When a person fabricates a story, their brain must invent details, check them for consistency, and suppress the truth, all while trying to sound natural. This extra mental work, known as cognitive load, often leaves a trace: the person takes longer to answer, makes more mistakes, or gives vague responses. However, this method has a flaw. If a liar has rehearsed their story beforehand, they can recite it quickly and smoothly, fooling the system. To solve this, scientists developed a technique called "unexpected questions." Instead of asking about the main points of a story that the person has prepared, the interviewer asks about details the liar never thought to memorize. This forces the deceiver to stop and build new information on the spot, while an honest person simply reaches into their memory for a real experience.

A team of researchers at the University of Padua in Italy recently put this technique to the test in a controlled setting designed to mimic a real investigative interview. They gathered sixty volunteers and split them into two groups. The first group, the honest participants, was asked to describe the most recent day trip they had actually taken. They had to recall the specific details of their own experience. The second group, the deceptive participants, was given a short, written story about a day trip and asked to memorize it perfectly. They were told to pretend this story was their own real memory. Both groups then sat for an audio-recorded interview where they answered a series of questions. Some questions were standard and expected, like "What time did you wake up?" Others were control questions about the interview itself, like "What city are you in?" The critical questions were the unexpected ones, such as "How long did it take you to reach your destination?" These were designed to catch the liars off guard. The researchers measured three things: how long it took each person to start speaking after a question was asked, how often they gave the wrong answer, and how often they gave an answer that was technically correct but vague or incomplete.

The results showed a clear difference between the two groups, but only when the questions were unexpected. When asked about details they had not prepared, the people lying about their day trip took significantly longer to answer than the honest people. They also made far more mistakes and gave more vague answers. For the honest group, the unexpected questions were also harder than the expected ones, but the increase in time and errors was much smaller. The researchers found that the liars struggled specifically because they had to invent new details that fit their fake story while keeping the rest of the narrative consistent. This extra mental work slowed them down and made them slip up. In contrast, the honest people were simply digging into their own memories, which was a different kind of mental task but one that did not require the same level of on-the-spot construction. The study also asked the participants how difficult they found the questions. The liars rated the unexpected questions as much harder than the honest people did, confirming that they felt the extra strain.

To see if these findings could be used to identify a liar in a real case, the researchers used computer algorithms to analyze the data. They fed the response times, error rates, and vagueness scores into a machine-learning model. When the computer looked at the entire pattern of answers from a single person, it was able to distinguish between the honest and deceptive groups with very high accuracy, correctly identifying the vast majority of liars. However, when the computer tried to judge a single answer in isolation, it was much less reliable, often missing the lie or falsely accusing an honest person. This suggests that a single moment of hesitation is not enough to prove deception; it is the overall pattern of behavior across many questions that reveals the truth. The researchers also tested a simple rule: if a person's average time to answer the unexpected questions was more than a specific fraction of a second slower than their average time for the other questions, they were likely lying. This rule worked well at catching liars but also flagged many honest people as suspects, meaning it was too prone to false alarms to be used alone.

The study offers a promising new tool for understanding how deception works, but the authors are careful to note its limits. The experiment took place in a low-stakes environment where the consequences of lying were minimal, unlike a real criminal investigation where the pressure is immense. The liars in the study were simply memorizing a story, whereas real liars often have to weave complex falsehoods into their own lives. Furthermore, the study relied on a single interviewer and coder, which could have introduced bias, and the participants were mostly university students, who may not represent the general population. Despite these limitations, the findings suggest that asking unexpected questions is a powerful way to increase the mental load on a liar, making their deception harder to maintain. The research indicates that while no single test is perfect, combining response times with error rates and using computer analysis could eventually help investigators spot lies that human intuition misses. The work points toward a future where interviews are designed not just to gather facts, but to gently stress the mind of the speaker, revealing the hidden effort of a lie through the simple, measurable signs of hesitation and error.

Technical Summary: Detecting Deception in Event Narratives via Response Latencies and Error Rates

Problem Statement
The detection of deception remains a critical challenge in forensic and investigative contexts, where human accuracy in distinguishing truth from lies hovers near chance levels (approximately 50%). While the "unexpected questions" technique has shown promise in reducing the advantages of pre-prepared lies, its application to the narration of false autobiographical events remains underexplored. Most existing research focuses on identity deception or "bald-faced" lies (100% false vs. 100% true), which may not reflect the complexity of real-world investigative interviews involving the distortion or omission of truth. Furthermore, pre-prepared lies (alibis) stored in long-term memory can reduce cognitive load, potentially masking deception indicators like response latency. This study addresses the gap in detecting deception within complex event narratives and seeks to move beyond the binary truth/lie dichotomy by introducing the metric of "inaccuracy."

Methodology
The study employed a between-subjects experimental design with 60 Italian participants (aged 19–29), randomly assigned to an honest group ( $n=29$ ) or a deceptive group ( $n=31$ ).

Task Configuration: The honest group described their most recent actual day trip (episodic retrieval), while the deceptive group memorized a provided narrative of a day trip within five minutes (script retrieval).
Interview Protocol: Participants underwent an audio-recorded interview containing 36 questions categorized as:
- Control (C): Unrelated to the event (e.g., current city).
- Expected (E): Consistent with the narrative script (e.g., wake-up time).
- Unexpected (U): Unanticipated details requiring on-the-spot construction or retrieval (e.g., travel duration).
Measures:
- Response Latency: Time from the end of the question to the start of the first meaningful word.
- Error Rate: Binary coding of completely incorrect or garbled answers.
- Inaccuracy Rate: A novel metric coding responses that were not strictly wrong but lacked precision (e.g., providing calculations instead of a specific time, partial answers, or vague circumlocutions).
- Perceived Difficulty/Predictability: Self-reported Likert scales.
Analysis: Statistical analyses included mixed ANOVAs and independent samples t-tests. Additionally, machine-learning models (Logistic Regression and Random Forest) were trained to classify participants at the subject level and individual responses at the item level using response latency, error, and inaccuracy rates.

Key Results

Cognitive Load Indicators: Both groups exhibited longer response latencies and higher error/inaccuracy rates for unexpected questions compared to control and expected questions. However, these indices were significantly higher for the deceptive group specifically in response to unexpected questions ( $p = .022$ for latency; $p < .001$ for error and inaccuracy rates).
Group Differences:
- Response Latency: Deceptive participants took significantly longer to answer unexpected questions than honest participants ( $d = -0.606$ ). No significant difference was found for control or expected questions.
- Error and Inaccuracy: The deceptive group showed a significantly higher error rate ( $d = -1.909$ ) and inaccuracy rate ( $d = -1.610$ ) for unexpected questions compared to the honest group.
- Perceived Difficulty: Deceptive participants rated unexpected questions as significantly more difficult than honest participants ( $d = -0.960$ ).
Machine Learning Classification:
- Subject-Level: A Random Forest model utilizing latency, error, and inaccuracy rates achieved excellent cross-validated discriminatory performance with an Area Under the Curve (AUC) of 0.967. At Youden's cut-off, the model achieved 93.3% accuracy (calculated from the confusion matrix: 26/29 honest correctly identified and 30/31 deceptive correctly identified), with 96.8% sensitivity and 89.7% specificity.
- Item-Level: Classifying individual responses (single questions) yielded moderate results (AUC = 0.738, Accuracy = 70%) with a high false-negative rate (41.4%), indicating that single responses are less reliable for detection than aggregate patterns.
Threshold Analysis: A latency threshold analysis suggested that if a participant's average latency to unexpected questions exceeded their baseline (control/expected) by at least 0.618 seconds, they could be classified as deceptive. Applying a "7-out-of-12" rule (where 7 or more unexpected questions exceeded this threshold) increased accuracy to 75% but resulted in a high false-negative rate (38.7%).

Significance and Claims
The authors claim that response latency, error rates, and the newly introduced inaccuracy rate are effective indicators for distinguishing deceptive from honest participants in autobiographical event narratives. The study suggests that unexpected questions impose a dual cognitive burden: the inherent difficulty of retrieving un-rehearsed details (affecting both groups) and the additional load of constructing consistent false details and monitoring plausibility (specifically affecting the deceptive group).

The paper posits that these indicators, particularly when aggregated and analyzed via machine learning, offer "diagnostic potential" worthy of further investigation. The introduction of the inaccuracy metric is highlighted as a step toward overcoming the limitations of the "bald-faced lie" dichotomy, capturing less precise response strategies used to avoid errors.

However, the authors maintain a modest stance regarding immediate practical application. They emphasize that:

The high performance of subject-level models requires external validation and replication before forensic application.
Item-level classification (single responses) is currently too unreliable for forensic use due to high error rates.
The study utilized a low-stakes context with student participants, which may limit generalizability to high-stakes real-world investigations.
Practical application is currently limited by the need for a verified reference account to calculate error and inaccuracy rates, a condition rarely met in real investigations.

The study concludes that while the unexpected questions technique shows promise as a non-invasive, implicit strategy for investigative interviews, the findings reflect deception operationalized under a specific experimental paradigm (episodic retrieval vs. recent memorization) and require further field-based research to confirm ecological validity.

Technical Summary: Detecting Deception in Event Narratives via Response Latencies and Error Rates

More like this