The science behind the game
Game – not evidence· Not a medical device.
What we measure, how we calculate it and, above all, what we cannot know. The scoring rules are fixed, written by people and identical on the phone and in the browser.
What a “reaction” means
When we are surprised, embarrassed or scared, the sympathetic nervous system speeds up the heart, narrows the blood vessels in the skin and increases sweating. These changes are automatic and appear just as well for the truth as for a lie.
The game measures how large the change is compared with you at rest. It does not compare people with each other and it does not judge honesty.
The signals used
Heart rate and inter-beat intervals (from PPG or a chest strap), pulse-wave amplitude (it drops when vessels narrow), electrodermal activity on Galaxy Watch8, and the accelerometer to detect movement.
Heart rate at 1 Hz, the only signal on many watches, misses fast changes. The app says so explicitly when that is all it has.
Baseline and index
Calibration takes 60–90 seconds plus a few neutral questions. From the rest windows we compute the mean and spread of each signal.
For each question we compute the z-score of each signal in the 0–15 second window after the question, combine them with fixed weights (wrist EDA counts less) and map the result to a 0–100 index. 50 means typical for you at rest; above 65 is “elevated”, above 80 “high”.
The concealed information test (CIT)
“Guess the card” follows Lykken's paradigm: one target among 4–6 irrelevant items, shown shuffled 3–5 times, at least 15 seconds apart. We compare the reaction to the target with the same person's reactions to the irrelevant items (effect d). Above 0.3 we say “larger reaction to the target”, between 0.1 and 0.3 “inconclusive”.
Meta-analyses show the CIT works well above chance in the lab, but not perfectly, and much less well outside the lab.
Signal quality
Every window gets a quality grade: movement (standard deviation of acceleration) and PPG waveform shape (Elgendi skewness index) can invalidate a window. An invalid window gets no index but “inconclusive”, and the game offers to ask the question again.
Limitations we do not hide
The American Psychological Association and the US National Research Council (2003) conclude that the polygraph lacks a solid scientific basis for detecting deception. Emotion is not guilt.
Wrist sensors are noisier than lab electrodes, wrist EDA correlates weakly with palm EDA, and coffee, alcohol, heat or laughter change every signal. That is why the results are a game and nothing more.
No machine learning in the score
The score does not use trained models. AI (optional, with consent) only generates questions and a funny summary, receiving derived scores only, without names or raw signals. Generated texts are labelled “AI-generated”.
Sources
- American Psychological Association (2004). The Truth About Lie Detectors (aka Polygraph Tests).
- National Research Council (2003). The Polygraph and Lie Detection. The National Academies Press.
- Lykken, D. T. (1959). The GSR in the detection of guilt. Journal of Applied Psychology, 43(6), 385–388.
- Meijer, E. H., et al. (2014). Physiological detection of concealed information: A meta-analysis. Psychophysiology.
- Matsuda, I., Nittono, H., & Allen, J. J. B. (2012). The current and future status of the concealed information test. Frontiers in Psychology, 3, 532.
- Elgendi, M. (2016). Optimal signal quality index for photoplethysmogram signals. Bioengineering, 3(4), 21.
- van der Mee, D. J., et al. (2021). Validity of electrodermal activity-based measures of sympathetic nervous system activity from a wrist-worn device. International Journal of Psychophysiology.
- Regulation (EU) 2024/1689 (AI Act), Article 5(1)(f).