Seminar M.Sc Students
Subjective Utility and willingness to pay for AI Assistance
Maximiliano Niemetz, M.Sc. student in the School of Industrial & Intelligent Systems Engineering
Advisors: Prof. Joachim Meyer
Abstract: Misinformation detection poses a complex classification challenge where human judgment is often constrained by uncertainty, time pressure, and cognitive bias.
While artificial intelligence is increasingly used as a decision support tool to flag suspicious items, the effectiveness of such hybrid decision systems depends not only on the technical accuracy of the algorithm, but on how users subjectively perceive, trust, and economically value that support under different decision environments.
This work presents a study combining a normative decision model based on Signal Detection Theory with a behavioral experiment to investigate how people evaluate AI assistance in a fake-news verification task. Participants completed a series of scenarios with systematically varied payoff structures and performance metrics.
In each scenario, they evaluated the contribution of a human editor, evaluated the contribution of a perfectly accurate AI system, and indicated their maximum willingness to pay for the AI support.
The results show that the true usefulness of an AI tool and its perceived financial value are two different things.
This highlights that getting people to adopt and pay for AI requires more than just proving the technology is accurate.
Bio: Maximiliano Niemetz is an M.Sc. student in the School of Industrial and Intelligent Systems Engineering at Tel Aviv University, specializing in Data Science. His research, supervised by Prof. Joachim Meyer, focuses on human-AI decision-making and the economic valuation of AI assistance in fake-news detection. By combining normative decision models based on Signal Detection Theory with behavioral experiments, his work investigates how people perceive, trust, and price AI support across different uncertain decision environments
AI Explanations and External Context in Misinformation Detection: Effects on Reliance and Human-AI Team Performance
Amy Vogel , M.Sc. student in the School of Industrial & Intelligent Systems Engineering
Advisors: Prof. Joachim Meyer
Abstract: As AI-based fact-checking tools become increasingly used to support misinformation detection in news and media contexts, understanding what drives appropriately calibrated human reliance on these systems is critical. We report a controlled online experiment (N = 275) examining how AI explanations and external source context affect reliance on an AI decision support system in a misinformation detection task. Participants classified 30 political statements drawn from PolitiFact as misinformation or accurate. Items were selected from a preliminary study so that one third were easy to classify, one third intermediate, and one third difficult. They received recommendations from an 80% accurate Wizard-of-Oz AI system, in a 2 (explanation: present vs. absent) × 2 (context: present vs. absent) between-subjects design, with item difficulty varied within subjects.
AI explanations and external context independently increased over-reliance on AI recommendations. Explanations raised incorrect acceptance without affecting overall AI acceptance or correct acceptance. Context showed a similar pattern: rather than facilitating independent judgment, it also increased over-reliance and significantly reduced team accuracy. Task difficulty further amplified over-reliance for difficult items. Across all conditions, human-AI team accuracy did not exceed the AI-alone baseline. Overall, participants were substantially more likely to accept an incorrect AI recommendation when the AI incorrectly labeled a true statement as misinformation, compared to when it incorrectly labeled actual misinformation as accurate.
These findings suggest that adding information cues — even those designed to support informed judgment — can lead people to trust AI systems more, rather than make better decisions, with implications for the design of AI-assisted fact-checking tools and for the broader study of human-AI trust and reliance in decision support contexts.
Bio: Amy Vogel is an M.Sc. student in the School of Industrial Engineering at Tel Aviv University, supervised by Prof. Joachim Meyer. She holds a Bachelor of Science in Engineering from MIT’s Department of Civil and Environmental Engineering. Her research focuses on reliance and trust calibration in human-AI interactions, particularly the impact of explanations and external context on over-reliance in AI-assisted misinformation detection. She is also a Data Analyst at HoneyBook, where she works on product analytics and the evaluation of AI-driven features.
RADAR-BASED PHYSICAL PRESENCE VERIFICATION FOR COUNTERING BOT-DRIVEN COMPUTER MISUSE
Amichai Gelkop, M.Sc. student in the School of Industrial & Intelligent Systems Engineering
Advisors: Prof. Pavel Ginzburg & Prof. Amichai Painsky
Abstract: Modern personal computers can remain active after authentication even when the authorized user is no longer physically present, creating opportunities for bot-driven misuse, unauthorized automation, and remote activity. Conventional security mechanisms — including passwords, multi-factor authentication, and software monitoring tools — verify digital credentials or software behavior, but do not continuously verify whether a real authorized user is physically present near the device. This thesis presents a local physical presence verification mechanism based on millimeter-wave (mmWave) radar sensing, designed to serve as a hardware-level complement to existing cybersecurity defenses.
A Texas Instruments AWR1642 mmWave sensor was mounted near a laptop and directed toward the expected user seating location. The sensor acquired frame-based point cloud data, which were converted into an 18-dimensional feature vector spanning spatial, Doppler, reflectivity, and region-of-interest descriptors. Three machine learning classifiers — Random Forest, Gradient Boosting, and RBF-kernel SVM — were trained and evaluated. Random Forest was selected as the primary model due to its highest AUC-ROC (0.9814), strong interpretability via permutation feature importance, and robustness to feature scale differences.
Under baseline conditions (user present versus clean absence), the radar-only classifier reached 97% accuracy and an F1 score of 0.98. To evaluate practical security robustness, a structured sequence of physical spoofing attacks was then applied. A pillow placed at the seating position was sufficient to fool the baseline model, which classified it as present in nearly 100% of cases. Hard negative retraining with pillow samples restored specificity to 99.2%. A foil-covered pillow was then introduced to exploit the model's residual dependence on reflectivity, again causing near-total failure (specificity 2.4%) until foil samples were added to the training set, restoring specificity to 89.5%. Finally, motion was added to the foil-covered object to mimic human micro-motion, reducing post-retraining specificity to 84.9% — the most challenging tested condition, with AUC-ROC of 0.98.
These experiments reveal a systematic measure-and-countermeasure dynamic: each new spoofing layer breaks the model until representative examples are included in training, after which performance largely recovers. The pattern has clear implications for deployed systems, which must be continuously updated against novel physical attack strategies.
To address the fundamental identity limitation of radar alone — a real person cannot be distinguished from an unauthorized intruder — radar-camera fusion was introduced. A laptop camera stream was combined with the radar classifier using a late fusion rule. Four conditions were evaluated: legitimate user present, truly absent, photograph spoofing (a printed photo of the user at the seating location), and a different unauthorized person. The fused system achieved 91.7% overall accuracy, correctly identifying 84.7% of photograph attack frames and 98.8% of unauthorized-person frames — scenarios that neither modality could resolve alone.
These results demonstrate that mmWave radar sensing, when combined with adversarial retraining and sensor fusion, provides a practical physical-layer signal for presence-aware computer security. The approach complements existing authentication mechanisms by providing continuous, privacy-preserving physical evidence of user-attended operation.
Bio: Amichai Gelkop, is an M.Sc. student in the School of Industrial Engineering at Tel Aviv University, For the past year and a half have been working with Prof. Pavel Ginzburg focusing on different usage of radar signals.
