Deception Detection: Using Machine Learning to Analyze 911 Calls

This study examined the use of machine learning in detecting deception among 210 individuals reporting homicides or missing persons to 911. The sample included an equal number of false allegation callers (FAC) and true report callers (TRC) identified through case adjudication. Independent coders, unaware of callers’ deception, analyzed each 911 call using 86 behavioral cues. Using the random forest model with k-fold cross-validation and repeated sampling, the study achieved an accuracy rate of 68.2% for all 911 calls, with sensitivity and specificity at 68.7% and 67.7%, respectively. For homicide reports, accuracy was higher at 71.2%, with a sensitivity of 77.3% but slightly lower specificity at 65.0%. In contrast, accuracy decreased to 61.4% for missing person reports, with a sensitivity of 49.1% and notably higher specificity at 73.6%. Beyond accuracy, key cues distinguishing FACs from TRCs were identified and included cues like “Blames others,” “Is self-dramatizing,” and “Is uncertain and insecure.”

Researchers analyzed 210 balanced 911 calls (105 false allegation, 105 true report) with callers identified through case adjudication. Independent blind coders rated 86 behavioral cues, which were analyzed using random forest machine learning with k-fold cross-validation and repeated sampling.

Background & Context

Deception detection in emergency 911 calls represents a critical challenge at the intersection of law enforcement, forensic psychology, and machine learning. When individuals report homicides or missing persons to emergency services, first responders must rapidly assess the credibility of the caller while the investigation clock is ticking. Research consistently shows that humans detect deception at barely above-chance levels—approximately 54% accuracy in everyday situations—yet the stakes in 911 calls couldn't be higher.

This 2026 study by Markey and colleagues builds on their extensive prior research examining behavioral cues in high-stakes deception contexts. Earlier work from their Interpersonal Research Lab identified that deceptive 911 callers display distinctive emotional and behavioral signatures, but manual human coding of these patterns is labor-intensive and subjective. The application of machine learning to automate and potentially improve deception detection in this context addresses a significant practical need in criminal investigations.

The study also emerges against a backdrop of controversy surrounding law enforcement training programs that claim to identify deceptive callers based on specific behavioral indicators—programs whose core claims have repeatedly failed to replicate in controlled research. By employing rigorous machine learning methods with cross-validation, this research offers a data-driven alternative to subjective intuition or unvalidated training protocols.

Research Design & Methodology

The study analyzed 210 individuals reporting either homicides or missing persons to 911, with an equal number (105 each) of false allegation callers (FAC) and true report callers (TRC) identified through case adjudication. Independent coders who were unaware of each caller's deception status analyzed every 911 call using 86 distinct behavioral cues. The sample was deliberately balanced to avoid bias toward either outcome category.

Ground truth determination followed established protocols in 911 deception research: callers were coded as TRC if the case was ruled a suicide, a third party was convicted of charges in which the caller was not involved, or the caller was exonerated through alibi verification, forensic evidence, or other physical evidence—consistent with investigative and judicial outcomes used in prior 911 research. The 86-cue behavioral coding system drew from previous 911 studies, general deception research, and interpersonal behavior assessments.

The analytical approach employed sophisticated machine learning techniques:

• Random forest classification model with k-fold cross-validation and repeated sampling to ensure robust generalization
• Separate analyses conducted for homicide calls versus missing person calls to examine context-specific performance
• Feature importance analysis to identify which behavioral cues most strongly discriminated between deceptive and truthful callers
• Calculation of sensitivity (true positive rate), specificity (true negative rate), and overall accuracy metrics

Results & Key Findings

Across all 911 calls, the random forest model achieved an overall accuracy rate of 68.2%, with sensitivity at 68.7% and specificity at 67.7%—substantially exceeding chance-level performance and human judgment accuracy reported in prior studies. This represents a meaningful improvement over the approximately 54% accuracy typically observed in human deception detection.

Performance varied significantly by call type, revealing important contextual effects:

• Homicide reports: 71.2% accuracy, with 77.3% sensitivity but slightly lower 65.0% specificity
• Missing person reports: 61.4% accuracy, with substantially lower 49.1% sensitivity but notably higher 73.6% specificity
• The model demonstrated better detection of false allegation callers in homicide cases but was more conservative (fewer false positives) in missing person cases

Key behavioral cues that distinguished false allegation callers from true report callers included "Blames others," "Is self-dramatizing," and "Is uncertain and insecure." These findings align with the research team's prior discovery that deceptive callers were self-dramatizing, moody, worried, and related narratives that lacked structure, clarity, and focus, whereas honest callers were forthright, candid, focused, cooperative, and provided a coherent sequence of events.

Discussion & Significance

The machine learning approach demonstrates that automated behavioral analysis can exceed both chance-level and typical human performance in detecting deception in 911 calls. The 68.2% overall accuracy represents approximately a 14-percentage-point improvement over the ~54% baseline human detection rate, which translates to substantial practical value when multiplied across thousands of emergency calls. The model's ability to consistently apply objective criteria without fatigue or bias offers advantages over human judgment.

The differential performance across call types reveals important insights about the nature of deception in different emergency contexts. The higher sensitivity for homicide calls (77.3%) suggests that false allegation callers in homicide scenarios may exhibit more pronounced or consistent behavioral signatures that machine learning can detect. Conversely, the lower sensitivity but higher specificity for missing person calls (49.1% and 73.6%, respectively) indicates the model is more conservative in this context—possibly because genuine missing person reports encompass more behavioral variability, or because the cognitive load and emotional presentation differ when someone has actually gone missing versus been killed.

This research contributes to the growing body of evidence that multivariate, data-driven approaches to deception detection substantially outperform single-cue heuristics or untested training protocols. The identification of specific discriminative cues provides investigative value beyond simple classification, offering law enforcement concrete behavioral indicators to inform investigative priorities without replacing human judgment entirely.

Limitations & Considerations

The study's reliance on case adjudication outcomes (convictions, exonerations) as ground truth introduces potential limitations. Legal outcomes, while the best available standard, do not always perfectly align with actual guilt or innocence given the imperfections of the criminal justice system. Some truly guilty callers may not be convicted, and wrongful convictions, while rare, do occur. This means the model is technically learning to predict legal outcomes rather than absolute truth.

The 86 behavioral cues were coded by human raters, introducing a layer of subjective interpretation despite efforts to use blind coders and establish inter-rater reliability. The machine learning model is therefore constrained by the quality and consistency of human coding. Additionally, the sample consisted of calls from public archives and media sources, which may not fully represent the broader population of 911 emergency calls—potentially skewing toward higher-profile cases.

The balanced 50/50 design (equal FAC and TRC) does not reflect the natural base rate of deceptive calls in real-world emergency dispatch scenarios, where the vast majority of calls are likely truthful. Deploying this model operationally would require careful calibration to account for base rates and the costs of false positives (wrongly suspecting honest callers) versus false negatives (missing deceptive callers).

Practical Applications

This research offers a proof-of-concept for machine learning-assisted triage in 911 call analysis, particularly for homicide reports where the model achieved over 71% accuracy. Law enforcement agencies could potentially implement such systems as investigative support tools—flagging calls with behavioral patterns consistent with deception for prioritized investigative follow-up, rather than as definitive determinations of guilt. The identified discriminative cues (blaming others, self-dramatizing, uncertainty) provide concrete behavioral indicators that investigators can consider alongside other evidence.

For polygraph examiners and forensic interviewers, this work reinforces the value of comprehensive behavioral assessment using multiple cues rather than relying on single indicators. The success of the machine learning approach demonstrates that deception in high-stakes contexts produces detectable multivariate signatures. However, the sub-optimal performance in missing person cases (61.4% accuracy) underscores that no single approach works universally across all contexts, and human expertise remains essential for integrating automated analysis with investigative context, evidence, and legal standards.