Speaker Introduction and Human Lie Detection Limitations

• Riccardo: psychologist, PhD candidate studying lie detection abilities
• Humans tell ~2 lies per day according to scientific estimates
• Human lie detection accuracy: chance level (~50%, coin flip equivalent)
• 2006 meta-analysis of 108 studies findings:
  • Naive judges: 54% accuracy
  • Experts (police, prosecutors, psychologists): 55% accuracy
  • Experience alone insufficient for accurate lie detection

Applications for Lie Detection Technology

• Criminal investigations: assessing suspect reliability during interviews
• Personal relationships: detecting partner infidelity
• Security contexts: border crossings, airport screening
• Political statements: identifying politician authenticity
• Recruitment processes: distinguishing genuine vs strategic responses
• Social media: combating scams, identity theft, fake news

Technical Approach: Large Language Models

• Large Language Models (LLMs): generate natural language mimicking human communication
• Fine-tuning process: specialized training after general education (metaphor: students specializing in law/medical school)
• Model used: Google's FLAN-T5 ("Flanny")
• Training datasets: 3 existing datasets containing truthful/deceptive statements
  • Personal opinions dataset
  • Future intentions dataset
  • Third dataset (type unspecified)
• Data collection method: participants instructed to tell truth or lie about specific topics
  • Example: Vietnam vacation story (truthful vs fabricated)

Research Study Design and Results

Experiment 1: Single Dataset Training

• FLAN-T5 fine-tuned on each dataset separately
• Results: 70-80% accuracy range
• Outperformed human benchmarks

Experiment 2: Cross-Dataset Generalization

• Model trained on 2 datasets, tested on third
• All 3 possible combinations tested
• Results: accuracy dropped to ~50%
• Demonstrates poor generalization across contexts

Experiment 3: Combined Dataset Training

• Model trained on all 3 datasets combined
• Results: accuracy recovered to ~80%
• Shows successful generalization when exposed to diverse examples

Key Findings and Implications

• LLMs can effectively classify deceptive statements when properly trained
• No universal deception rule exists across contexts
• Linguistic deception cues are context-dependent
• Generalization possible with diverse training exposure
• More replication studies needed before real-world deployment

Future Vision: Integrated Lie Detection Technology

• National security applications: enhanced safety measures
• Political transparency: identifying authentic vs strategic statements
• Security screening: detecting malicious intentions pre-incident
• Corporate recruitment: distinguishing genuine passion from strategic responses
• Social media protection: automatic fake news flagging with credibility scores
• Automated content verification systems

Ethical Concerns and Risks

• Blind reliance on AI outputs without critical thinking
• Increased likelihood of false accusations based on AI judgments
• Loss of interpersonal trust in society
• Risk of abandoning human judgment and critical evaluation
• Previous studies confirm people's tendency to over-rely on AI decisions

Proposed Solution: Interpretable AI

• AI should provide explanations alongside conclusions
• Analogy: online reviews showing both star ratings and detailed feedback
• Users need clear, understandable reasoning behind AI decisions
• Technology should augment, not replace human judgment
• Maintain human agency in decision-making processes
• Balance between AI assistance and critical thinking preservation

Conclusion and Call to Action

• AI lie detection future requires technological advancement + ethical understanding
• Focus on enhancing rather than replacing human judgment
• Commitment to interpretability and responsible development
• Avoid blind technological reliance
• Pursue truth through deep understanding and ethical application