- The paper proposes Cognitive Load Correlation (CLC), a within-session measure linking content complexity to typing latency, with genuine composition estimated at ρ = 0.3–0.6 versus below 0.15 for transcription.
- The paper designs a non-intrusive system that collects only quantized keystroke timing, claiming under 0.5% cognitive-signal loss at 5 ms resolution while reducing biometric re-identification risk.
- The paper analytically estimates 85–95% discrimination accuracy and argues that cognitive signatures are harder to forge, but emphasizes that controlled within-subject experiments and validation across diverse users remain necessary.
The paper proposes reframing authorship verification as a human-computer interaction problem grounded in the cognitive process of writing rather than the textual output. It argues that keystroke timing during genuine composition carries "cognitive signatures"—planning pauses, clausal burst structure, revision patterns—that are absent when text is transcribed mechanically or pasted from an LLM, and that these signatures can be collected non-intrusively within ordinary writing interfaces while limiting biometric leakage through evidence quantization.
Motivation and problem statement
The paper positions itself against three failing defenses of authorship integrity: output-based AI-text detection, watermarking, and proctoring. Citing information-theoretic barriers as model outputs converge toward human distributions (Sadasivan et al., 2023), it notes estimates that up to 16.9% of peer-reviewed conference submissions contain AI-modified content [Liang2024]. The central claim is that all output-centric approaches share a common flaw—they ignore the process—and that the writing interface already captures the temporal trace of composition: inter-key intervals (IKIs) reflecting planning, translating, and revising stages in the sense of Flower and Hayes' cognitive process theory [FlowerHayes1981].
Cognitive signatures
Drawing on Salthouse's distinction between transcription typing (perceptual-motor) and composition (language generation and monitoring) [Salthouse1986], the paper characterizes three temporal fingerprints: planning phases with extended pauses of 1,000–5,000 ms at discourse boundaries; translating bursts aligned with clauses, with intra-burst IKI around 100–200 ms exhibiting $1/f$-like structure; and revision sequences qualitatively distinct from single-character typo corrections [Leijten2013]. The ScholaWrite dataset reports IKI values 40–60% higher in planning-labeled intervals than in translation intervals [Nguyen2024].
The core contribution is the Cognitive Load Correlation (CLC): the within-session correlation ρ between content complexity and typing latency. During genuine composition, CLC is positive and moderate to strong (ρ=0.3–$0.6$); during transcription it is near zero (ρ<0.15). Because CLC is computed within-session rather than against population thresholds, it partially mitigates confounds from typing proficiency—a slow typist pausing before difficult content still yields positive CLC.
Framework design
The proposed system is a lightweight background component (browser extension, plugin, or LMS feature) that collects timing metadata only—never key identities or screen content—quantizes IKIs, computes CLC and entropy estimates, and produces a verifiable evidence record. Privacy is addressed via quantization at r=5\,ms, exploiting a timescale separation: motor-identifying features degrade meaningfully above roughly 15\,ms resolution (Killourhy and Maxion report EER increases at coarse clock resolutions), while cognitive features live at scales two to three orders of magnitude larger. At r=5\,ms the paper claims less than 0.5% loss of cognitive signal while substantially degrading re-identification capability; differential privacy is offered as an optional stronger guarantee at approximately 15% accuracy cost per 0.10 decrease in ϵ [Loya2021].
Evaluation and stated assumptions
The evaluation is explicitly analytical, not experimental. No single dataset combines composition logs with content-complexity annotations at scale, so results synthesize complementary sources: ScholaWrite (10 subjects), the Aalto 136M-keystroke corpus (168,000 participants) [Dhakal2018], neuroQWERTY (85 subjects), and KLiCKE (7,209 subjects). Under three stated assumptions—generalizability of ScholaWrite CLC distributions, validity of transcription baselines, and accurate estimation of content complexity via n-gram surprisal or syntactic depth—the paper estimates 85–95% discrimination accuracy, exceeding statistical power of 0.99 for documents over 1,500 words. This figure should be read as a theoretical bound: a controlled within-subjects study comparing composition and transcription remains unperformed, and the paper says so plainly.
Population stratification is acknowledged as a material limitation. Hunt-and-peck typists may produce planning-like pauses during routine motor execution; second-language writers exhibit elevated IKI at lexical access points that could inflate CLC even during transcription; touchscreen and voice-input modalities produce fundamentally different or absent signals; motor impairments confound cognitive-motor separation. Fairness evidence is mixed but encouraging: Jiang et al. found negligible demographic fairness differences for keystroke-based nonauthenticity detectors, with small gender-related effects [Jiang2024].
Adversarial robustness
The paper's most distinctive argument concerns why cognitive signatures resist attacks that defeat motor-level keystroke biometrics. Motor features occupy a low-dimensional timing space demonstrably forgeable: human mimics achieve at least 80% bypass success [Tey2013], and Monaco reports ≥99.8% evasion against five classifiers using timing-forgery attacks [Monaco2026]. Cognitive features, by contrast, are claimed to be structurally harder to forge because forging a positive CLC requires simultaneous satisfaction of multi-level constraints: entropy-driven letter-level micro-variation tied to the typist's internal LLM [Crump2019], syntactically structured pause placement [Plank2016], content-correlated bursts, and realistic revision behavior. Supporting evidence includes Mehta et al.'s result that models distinguishing genuine composition from AI-transcribed text achieved F1 scores exceeding 97%, while human evaluators performed near chance (~45%) [Mehta2025].
The paper concedes the residual attack path directly: a motivated adversary who pre-composes text, memorizes it along with a complexity-derived "pause map," and rehearses its typing could plausibly produce positive CLC. The estimated cost is 30–60 minutes of preparation per 1,500-word document—comparable to writing the text outright, which the authors argue collapses the incentive. Multi-session consistency checks (comparing within-writer CLC variance to population norms) are proposed to raise detection reliability as observed sessions accumulate, though this claim about escalating reliability is itself analytical rather than demonstrated.
Ethics and deployment
Grounded in Nissenbaum's contextual integrity framework [Nissenbaum2004], the paper proposes layered transparency (status indicator, on-demand detail panel, plain-language disclosure), opt-in consent with genuinely voluntary participation in educational settings and alternative verification pathways such as oral defense, minimal data retention (aggregate CLC score plus cryptographic commitment after decision), and a biometric parity mode calibrating thresholds to individual motor profiles for users with impairments or atypical input modalities.
Limitations and open questions
Three limitations bound the contribution. First, the headline 85–95% accuracy estimate derives entirely from cross-dataset synthesis under assumptions that have not been jointly tested; no controlled experiment compares composition and transcription within the same participants. Second, the empirical base skews toward university-educated adults on desktop keyboards, leaving cross-population validation—for non-native writers, mobile users, and individuals with motor impairments—as an explicit prerequisite for deployment. Third, the adversarial robustness argument, while structurally persuasive, rests on the assumption that entropy-driven micro-timing cannot be deliberately manufactured; whether rehearsed adversaries introduce detectable performance artifacts, and how reliably multi-session consistency tests flag fabricated profiles, remain empirically open.
Conclusion
The paper offers a coherent theoretical reframing of authorship verification as process attestation, anchored in a well-grounded literature on writing-process cognition and keystroke dynamics, and notable for its candor about the analytical nature of its evaluation. Its central claim—that CLC separates composition from transcription more reliably than output-level signals can separate human from machine text—is plausible and supported by convergent published statistics, but it awaits validation in a within-subjects study with concurrent complexity annotation. The privacy-preserving design and the adversarial asymmetry argument are substantive contributions; both would be strengthened considerably by empirical measurement on quantized data across diverse populations.