The Signal Inversion Effect
How Authentic Cognition Mimics Deception: Evidence from Trial Testimony, Criminal Confessions,
and Global Belief Systems
Working Paper — Preprint — March 2026
Abstract
This paper presents converging evidence for the Signal Inversion Effect: a systematic pattern
whereby authentic cognitive and linguistic behaviours associated with truthfulness are misidentified
as indicators of deception by human observers. Across four studies we find: (1) disfluency is
significantly higher in truthful trial testimony than deceptive testimony (d = 0.60, p = .004), a
medium effect 6× larger than the median deception cue; (2) false confessions show massive linguistic
distancing, with impersonal pronoun contractions up to 12× higher and second-person references
7.6× higher than true confessions; (3) 91.3% of publicly believed deception cues are empirically
unrelated or inversely related to actual deception (21/23 cues inverted; binomial test p < .0001;
weighted inversion index = 0.81); and (4) human deception detection averages 54% accuracy (lie
detection specifically: 47%, below chance), while algorithmic classification using linguistic features
achieves 63.5–83%. Demeanour-based credibility assessment does not merely fail to detect
deception — it systematically penalises truthful speakers whose authentic cognitive effort produces
the signals misread as guilt.
Keywords: deception detection, credibility assessment, signal inversion, disfluency, false
confessions, belief-reality inversion, demeanour evidence
2.1 Method
We analysed 121 video transcripts from the Real-Life Trial Dataset (Pérez-Rosas et al., 2015): 60 truthful and
61 deceptive clips from actual court proceedings, with ground truth established through verdicts and
post-conviction exonerations. Eight linguistic variables were coded per 100 words. Non-parametric group
comparisons (Mann-Whitney U) were used due to non-normal distributions.
2.2 Results
Table 1. Group comparisons (Mann-Whitney U). ** p < .01.
Variable Truth M Truth SD Decep M Decep SD U p d
Disfluency/Fillers 5.14 4.21 3.03 2.68 2380.5 .004** 0.60
Negative Emotion 0.02 0.17 0.13 0.52 1711.0 .103 -0.27
First-Person Pron. 6.13 4.38 7.15 4.16 1561.5 .164 -0.24
Experiencer Framing 0.60 1.51 0.34 0.94 1934.0 .448 0.21
Certainty Markers 0.35 1.25 0.76 3.15 1687.5 .265 -0.17
Truthful speakers produced significantly more disfluencies (M = 5.14/100 words) than deceptive speakers (M
= 3.03), U = 2380.5, p = .004, d = 0.60, rpb = .290. This effect is 6× larger than the median deception cue
effect (d = 0.10) in DePaulo et al.'s (2003) meta-analysis of 158 cues. A multi-variable logistic regression
achieved 63.5% classification accuracy (10-fold CV), compared to 54% for human judges.
3.1 Method
We reanalysed published data from Rizzelli, Kassin, and Gales (2021): 37 confessions proven false through
DNA exoneration (Innocence Project) and 98 confessions presumed true (FBI case files). Per-confession word
rates were computed from Appendix A raw counts and tested with chi-square.
3.2 Results
Table 2. Word frequencies in true vs. false confessions (highest discriminators).
Marker True (per conf.) False (per conf.) Ratio χ² p
'you' 9.4 71.3 7.59× 3903.7 < .001
'that’s' 1.7 19.3 11.67× 1288.6 < .001
'it’s' 1.2 14.6 12.17× 993.1 < .001
'but' 5.2 19.6 3.78× 610.7 < .001
Impers. pron. total 49.0 253.6 5.18× 10701.9 < .001
The second-person pronoun 'you' appeared 7.59× more frequently in false confessions (χ²(1) = 3903.7, p <
.001), reflecting the interrogation dynamic: innocent individuals orient toward the interrogator who supplied
the narrative ('you said,' 'you told me'). Impersonal contractions 'that’s' (11.7×) and 'it’s' (12.2×) show that
even while confessing, innocent people frame events impersonally — linguistic distancing that occurs below
conscious awareness.
4.1 Method
We systematically matched deception cues appearing in both the Global Deception Research Team (GDRT,
2006) survey of 11,227 participants across 75 countries and DePaulo et al.'s (2003) meta-analysis, yielding 23
matched cues. For each, we coded whether the publicly believed direction matched or was inverted relative to
the empirical evidence. We supplemented DePaulo effect sizes with our Study 1 disfluency finding.
4.2 Results
Figure 1. Forest plot of all 23 matched deception cues. Red = belief inverted; Green = belief aligned. Percentages show global
belief endorsement rates.
Of 23 matched cues, 21 (91.3%) were inverted — the public believes these cues indicate deception, but
empirical evidence shows either no relationship or the opposite direction. Only 2 cues (appearing nervous,
negative statements) showed alignment between belief and evidence.
Table 3. Belief–Reality Inversion Matrix statistical summary.
Statistical Test Result Interpretation
Inversion rate 21/23 = 91.3% 9 in 10 believed cues are wrong
Binomial test p < .0001 Significantly above 50% chance
95% CI [75.1%, 100.0%] Lower bound still far above chance
Weighted Index 0.806 Strong systematic inversion (scale: -1 to +1)
Belief–validity r 0.237 (p = .276) No relationship: stronger beliefs ≠ more valid
Figure 2. Inversion rates overall and by cue category. Paralinguistic cues show 100% inversion; nonverbal 91%; verbal 86%.
The Weighted Inversion Index (0.806 on a scale from -1.0 to +1.0) indicates that the most strongly endorsed beliefs are the most empirically wrong. Gaze aversion, the #1 believed cue worldwide (63.7% endorsement), has an effect size of d = 0.05 — essentially zero. The Spearman correlation between belief strength and actual validity was non-significant (r = .237, p = .276), confirming that how strongly the public believes in a cue has no relationship to its diagnostic value.
Figure 3. Belief–Reality scatter plot. The bottom-right quadrant (high belief, negative effect) represents the most consequential inversions.
Metric Value Source
Human deception detection 54% Bond & DePaulo (2006), k=206
Human lie detection 47% (below chance) Bond & DePaulo (2006)
Police / Judges ≈55% Bond & DePaulo (2006)
Algorithmic (disfluency only) 60.3% Current Study 1
Algorithmic (multi-variable) 63.5% Current Study 1
Algorithmic (3-predictor) 74–83% Rizzelli et al. (2021)
Believed cues that are wrong 91.3% Current Study 3
Disfluency effect d = 0.60 (6× median) Current Study 1
'you' in false confessions 7.59× Current Study 2
robust association with truthfulness (d = 0.60). The inversion is not marginal: 91% of matched cues are wrong (p < .0001), and the weighted index shows the strongest beliefs are the most wrong.
6.1 Implications for Legal Proceedings
When judges, jurors, or investigators assess demeanour, they draw on heuristics that are empirically inverted.
A witness who pauses, says 'um,' breaks eye contact to search memory, qualifies statements with 'I think,' and
produces a fragmented narrative is exhibiting cognitive signatures of genuine recall. These same behaviours
activate the folk model of deception, producing credibility judgments that are directionally wrong. This cannot
be solved through training: Bond and DePaulo (2006) found professional groups perform no better than
untrained members of the public.
6.2 The Behavioural Adaptation Feedback Loop
Individuals repeatedly disbelieved despite telling the truth — because their authentic speech patterns trigger
inverted credibility heuristics — may learn to suppress authentic patterns and adopt performed behaviours
(steady eye contact, fluent delivery, reduced hedging) matching observers' expectations. This means repeated
demeanour-based credibility assessment does not merely fail; it actively trains truthful speakers to
communicate in ways that resemble deceptive speech, eroding the very cues that might otherwise assist
veridical assessment. The system is self-defeating.
6.3 The Unpredictable 9%
Two cues showed alignment between belief and reality: appearing nervous (d = 0.27) and negative statements
(d = 0.21). However, even these small effects are insufficient for reliable individual assessment, and their
presence in a sea of inverted cues creates a dangerous illusion: because occasionally a nervousness-based
judgment happens to be correct, observers are reinforced in their broader reliance on a framework that is 91%
wrong. The small kernel of truth sustains the larger inversion.
6.4 Limitations
Study 1 uses a small corpus (N = 121) with short clips. Study 2 uses aggregate counts rather than
individual-level data. Study 3 requires judgment in matching belief categories to meta-analytic cues. These
limitations notwithstanding, convergence across independent datasets, methodologies, and research groups
strengthens confidence in the general pattern.
6.5 Conclusion
Demeanour-based credibility assessment is not merely unreliable. It is systematically biased against truthful
communication. The cues authentic cognitive effort produces during genuine recall — disfluency, hedging,
fragmented narrative, gaze aversion during memory search — are precisely the cues human observers
interpret as deception. This Signal Inversion Effect operates across trial testimony, criminal confessions, and
global belief systems, and is resistant to professional training. Until credibility assessment is reformed to
eliminate reliance on demeanour evidence, truthful individuals will continue to be structurally disadvantaged
in any context where honesty is judged by appearance rather than content.
References Bond, C. F., Jr., & DePaulo, B. M. (2006). Accuracy of deception judgments. Personality and Social Psychology Review, 10(3), 214–234. Bond, C. F., Jr., & The Global Deception Research Team. (2006). A world of lies. Journal of Cross-Cultural Psychology, 37(1), 60–74. DePaulo, B. M., Lindsay, J. J., Malone, B. E., Muhlenbruck, L., Charlton, K., & Cooper, H. (2003). Cues to deception. Psychological Bulletin, 129(1), 74–118. Hartwig, M., & Bond, C. F., Jr. (2011). Why do lie-catchers fail? A lens model meta-analysis of human lie judgments. Psychological Bulletin, 137(4), 643–659. Kassin, S. M. (2017). False confessions: How can psychology so basic be so counterintuitive? American Psychologist, 72(9), 951–964. Luke, T. J., et al. (2023). What have we learned about cues to deception? A survey of expert opinions. Psychology, Crime & Law. Pennebaker, J. W., & Francis, M. E. (1999). Linguistic Inquiry and Word Count: LIWC. Erlbaum. Pérez-Rosas, V., Abouelenien, M., Mihalcea, R., & Burzo, M. (2015). Deception detection using real-life trial data. Proceedings of ACM ICMI 2015, 59–66. Rizzelli, L., Kassin, S., & Gales, T. (2021). The language of criminal confessions: A corpus analysis. Wrongful Conviction Law Review, 2(3), 205–225. Vrij, A., Granhag, P. A., & Porter, S. B. (2010). Pitfalls and opportunities in nonverbal and verbal lie detection. Psychological Science in the Public Interest, 11(3), 89–121.
Supplementary Materials Complete analysis code (Python/SciPy), raw data files (SPSS-compatible CSV), and all figure source files are available at the project repository and will be deposited on the Open Science Framework.
S1. Study 1 Classification Model
Logistic regression with disfluency as sole predictor: B0 = -0.740, B1 = 0.183, accuracy = 60.3%.
Multi-variable model (10-fold CV) using disfluency, hedging, certainty, experiencer framing, negative
emotion, first-person pronouns: M accuracy = 63.5% (SD = 11.5%). Standardised coefficients: disfluency (β =
+0.546), experiencer framing (β = +0.364), negative emotion (β = -0.366), first-person pronouns (β = -0.194).
S2. Study 3 Full Inversion Matrix
23 cues matched. Inversion status: 15 inverted-null (belief endorses a cue with no real effect), 6
inverted-opposite (belief direction is backwards), 2 aligned. Category breakdown: paralinguistic 100%
inverted (5/5), nonverbal 91% (10/11), verbal 86% (6/7). Binomial test: p < .0001. Weighted inversion index:
0.806.
Figure S1. Belief endorsement rate vs. actual cue validity. No significant relationship (Spearman r = .237, p = .276).
Published by OMXUS Publishing. Free to read. Free to share. Evidence is the credential.
Part of the OMXUS Research Series — 99 publications.
This paper serves:
#14 Cancer is 90% preventable.
For Lily and Joshua.