Introspective inference counteracts perceptual distortion

The study investigates how humans perform reality testing when perceptual experiences are distorted, focusing on the cognitive mechanism termed perceptual insight: incorporating knowledge about internal distortions to infer the true state of the world. Reality testing is impaired in psychosis (e.g., hallucinations believed to be real), yet formal frameworks and suitable paradigms are lacking. The authors hypothesize that when experiencing a strong perceptual distortion (motion after-effect, MAE), observers can use introspective knowledge of the distortion to adjust their inferences about stimulus categories, effectively decoupling perception from belief. They propose a Bayesian model where insight manifests as adjustments at an intermediate inferential stage (shift in the perceptual-decision variable) rather than as a late response bias, and design a task to test this dissociation with behavioral and physiological readouts.

Prior reality-testing research has mainly used: (1) source memory tasks (self vs other) relying on episodic memory, which do not capture in-the-moment processes; and (2) imagery-based perceptual decision tasks leveraging signal detection, which depend on imagery ability and limit control and generalizability. Broader Bayesian work has shown optimal integration of external sensory information and feedback, but not integration of introspective knowledge about internal distortions without external feedback. Previous studies demonstrate that changes at different decision stages yield distinct patterns in choice, RT, confidence, and pupil dilation, enabling dissociation between inference-level and response-level biases. Work on MAE and sensory adaptation supports early sensory encoding changes causing perceptual biases, while microstimulation and decision-neuroscience studies link intermediate decision variables to confidence and RT shifts in tandem.

Overview: Two in-person experiments (N=22 each after exclusions) employed a motion after-effect (MAE) paradigm with a counterclockwise-rotating Archimedean spiral adaptor to induce a robust perceptual distortion. A 2×2 design crossed Adapt vs No-Adapt (rotating vs static first spiral) with response instruction See vs Believe (report perceived vs inferred true motion direction of the test stimulus). Each trial collected left/right choice, binary confidence (high/low), and reaction time (RT). Experiment 2 added eye tracking (fixation enforcement, pupillometry). No feedback during the main task; participants were trained to understand the MAE and its illusory nature. Participants: Experiment 1 recruited 25, 3 excluded for failing to experience the illusion; 22 analyzed (11 male/11 female; median age 26, range 21–50). Experiment 2 recruited 26, 4 excluded (medication, no illusion, incomplete), 22 analyzed (6 male/16 female; median age 24.5, range 18–34). IRB-approved; informed consent obtained. Stimuli: Full-screen Archimedean spirals (ω1=30, ω2=3). Adaptor temporal frequency ω3=9 (temporal freq 1.5 cycles/s). Spatial frequency ~1.16 cycles/° (Exp1) and 1.14 cycles/° (Exp2). Adaptor velocity ~1.29–1.31°/s. Test stimuli had variable ω3 translating to velocities: Exp1 adaptive staircase generating speeds approximately within [-0.285, 0.285]°/s (arb. units ~[-0.5, 0.5]); Exp2 uniformly sampled ω3∈[-0.3, 0.3] (velocities ~[-0.044, 0.044]°/s), 121 trials/condition. Task and trial structure: Trials: fixation (Exp1: 890 ms; Exp2: 1000 ms), adaptor or static control (3000 ms), test spiral (≤1000 ms; in Exp2 always displayed full 1000 ms). Prompts signaled choice and then confidence. See and Believe trials were blocked. Adapt blocks included extended pre-block adaptation (Exp1: 15 s; Exp2: 30 s). In Exp1, condition order fixed (Adapt-See, Adapt-Believe, break, No-Adapt-See, No-Adapt-Believe). Exp2 randomized No-Adapt block order; Adapt-See preceded Adapt-Believe. Each condition: Exp1 120 trials (2×60); Exp2 121 trials (2 blocks). No feedback during main task. Training and checks: Multi-part instruction and practice ensured understanding of MAE and Believe mapping; quizzes; in Exp2, additional training with feedback on actual test speed (relative to adaptor) and accuracy criterion in practice Adapt-Believe (≥70%). Illusion reproduction tasks before and mid-experiment captured expected MAE strength and direction. Eye tracking and pupillometry (Exp2): EyeLink 1000 Plus at 1 kHz; fixation enforced within 3 dva during fixation/adaptor/stimulus (5 s) and within 24 dva during 3 s post-stimulus; blink violations restarted trials, ensuring controlled gaze and minimized blinks (8 s no-blink window per trial). Pupil preprocessing included artifact removal (speed threshold via MAD), interpolation, bandpass filtering (0.01–10 Hz), baseline subtraction (last 400 ms of fixation), and normalization. Decision-related windows: stimulus-locked (2000–2500 ms post-stimulus) and response-locked (500–1000 ms post-response). Behavioral analyses: Psychometric functions (Gaussian CDF) fitted per condition with parameters μ (bias), σ (noise); a single lapse rate λ shared across conditions (total 9 parameters). Bias from confidence and RT curves estimated by bin of minimum confidence/maximum RT. Nonparametric statistics (Wilcoxon signed-rank, Spearman correlations) with bootstrapped 95% CIs. Bayesian computational modeling: Extended perceptual decision model with three stages: early encoding xN(s−A,σ), intermediate inference computing log-posterior ratio d, and late stage choice (threshold k_choice) and confidence (threshold k_confidence). Perceptual-insight model incorporates knowledge of distortion A at inference via a shifted likelihood p(x|s,A) parameterized by μ_likelihood; late-compensation model omits A and adjusts k_choice. Additional variants allowed category prior shifts or hybrid effects. Model fitting jointly to choices and confidence per condition via MLE with sampling-based likelihood (500 samples), optimized with Bayesian adaptive direct search (BADS). μ_encoding fixed from No-Adapt-See and Adapt-See psychometric μ; free parameters included σ (per condition), k_confidence (per condition), and condition-specific μ_likelihood depending on model. Model comparison via AIC/BIC. RT and pupil modeling: Generalized linear mixed-effects models (GLMEs) predicted ranked RT or pupil area from ranked |d| (model-derived uncertainty) vs ranked |s|, including condition fixed effects and random slopes/intercepts. Moving-window GLMEs for pupil time courses. Drift-diffusion modeling (DDM): PyDDM used to fit joint choice-RT distributions per condition. Base model: mean drift rate, decision bound, nondecision time; variants added starting-point bias, drift-rate bias, or both. Drift scaled linearly with |stimulus|. Parameter bounds: drift [0,20], bound [1,3], nondecision [0,2], start bias [-1,1], drift bias [-1,1]. Model selection via AIC/BIC. Control pilot: Response-bias manipulation (“If unsure, press Left”) with No-Adapt-See and Adapt-See, plus No-Adapt-Bias and Adapt-Bias to test isolated response-stage bias (N=7 completers).

• MAE illusion and compensation:
  • Experiment 1: Adapt-See vs No-Adapt-See showed a leftward psychometric shift indicating clockwise perceptual bias (μ difference: z=4.108, p<0.001, r=0.619). Adapt-Believe vs Adapt-See showed a rightward corrective shift (compensation) (z=2.642, p=0.008, 95% CI [0.823, 3.602], r=0.398).
  • Experiment 2: Replicated with enforced fixation. Robust compensation in Adapt-Believe vs Adapt-See (z=4.107, p<0.001, r=0.619). No significant σ difference between Adapt-See and Adapt-Believe (z=−1.412, p=0.158). Participants often overcompensated (compensation index >1: z=3.945, p<0.001, r=0.595).
• Shifts in tandem across measures indicate intermediate inference adjustment:
  • Psychometric, confidence, and RT curves shifted in tandem for MAE and compensation (Fig. 4). Biases correlated across individuals for Adapt-See and for Adapt-Believe (all correlations significant; e.g., 0.63<ρ<0.85, p<0.002 for Adapt-Believe).
• Bayesian model comparison favored the perceptual-insight model:
  • Joint fits to choice and confidence selected the likelihood-shift (perceptual-insight) model over late-compensation (k_choice), prior-shift, hybrid, and other alternatives by AIC/BIC (Fig. 5B).
  • Winning model captured choice and confidence curves (Fig. 5C) and produced condition-specific μ_likelihood shifts under Adapt-Believe (μ_likelihood differed between Adapt-See and Adapt-Believe: z=4.107, p<0.001, r=0.619). Condition differences also found in σ_encoding (z=−3.457, p<0.001, r=−0.521) and k_confidence (z=2.613, p=0.009, r=0.394).
  • Changes in μ_likelihood strongly correlated with psychometric μ shifts between Adapt-See and Adapt-Believe (Spearman ρ=0.95, p<0.001), whereas changes in σ_encoding and k_confidence did not (p>0.634).
• RTs track model-derived decision uncertainty:
  • GLME: ranked |d| predicted ranked RT (β=−10.023, p<0.001; adjusted R^2=0.615), outperforming |s|-based model (β=−8.927, p<0.001; adjusted R^2=0.541) and showing no residual condition effects when using |d|.
• Pupillometry corroborates intermediate uncertainty signal:
  • Decision-related windows showed significant interaction between stimulus strength and condition (stimulus-locked F(10,210)=4.105, p<0.001; response-locked F(10,210)=3.073, p=0.001), with visible peak shifts between Adapt-See and Adapt-Believe.
  • GLMEs: ranked |d| significantly predicted pupil area (stimulus-locked β=−7.631, p<0.001, adjusted R^2=0.376; response-locked β=−5.389, p<0.001, adjusted R^2=0.346), remaining significant controlling for RT (β=−2.693, p=0.007; adjusted R^2=0.386). |d| outperformed |s| models (|s| β=−3.835, p<0.001; adjusted R^2=0.319). Effects held within Adapt-only analyses.
• DDM supports intermediate-stage bias (starting-point):
  • Model with starting-point bias best fit choice-RT data across conditions (BIC). Conditional response functions showed bias strongest at shortest RTs, consistent with starting-point biases.
  • Adapt-See vs Adapt-Believe differed in mean drift rate and starting-point bias. Differences in starting point strongly correlated with MAE compensation and μ_likelihood changes (ρ≈0.85–0.89, p<0.001), and not with other parameters; no significant differences found in decision bounds (arguing against late-stage response changes).
• Control experiment (response bias): Instructional bias (“If unsure, press Left”) produced isolated psychometric shifts without corresponding confidence/RT shifts; explained by k_choice changes and not μ_likelihood (Supplementary Fig. S9), dissociating response bias from perceptual insight.

The findings demonstrate that humans can counteract distorted perceptual experiences by incorporating introspective knowledge about internal biases into their inferential process, effectively decoupling seeing from believing. Shifts in tandem of psychometric, confidence, RT, and pupil curves between Adapt-See and Adapt-Believe are consistent with adjustments at an intermediate inference stage (shift in decision variable d), not merely response-level biases. Bayesian model comparison and parameter-behavior correlations pinpoint a likelihood-level adjustment (μ_likelihood) as the primary mechanism of compensation. RTs and pupil dilation aligned with model-derived uncertainty (|d|), reinforcing the interpretation of an inference-level process distinct from motor execution. DDM analyses converged on a starting-point bias as the decision-level signature of compensation, further arguing against late response-rule adjustments. Conceptually, the work formalizes perceptual insight—a key component of reality testing—showing that explicit knowledge of an illusion can recalibrate inference and prevent distorted percepts from dictating beliefs about the external world. This hierarchical separation between sensory encoding and higher-order inference aligns with theories of metacognition and decision-making, and suggests candidate neural substrates in decision and prefrontal regions for implementing insightful corrections of sensory biases. The approach lays groundwork for studying insight impairments in psychosis and for developing objective markers (e.g., pupillometry) of inference-level decision uncertainty.

This study introduces a formal Bayesian framework and controlled psychophysics paradigm to quantify perceptual insight as compensation for distorted perception. Healthy observers experiencing MAE adjusted their inferences about motion direction when instructed to report beliefs (Believe), yielding compensation—often overcompensation—relative to perceived motion (See). Shifts in tandem across choice, confidence, RT, and pupil dilation, together with model comparison and DDM analyses, indicate that compensation arises from intermediate inference-level adjustments (likelihood shift/start-point bias) rather than late response biases. These results establish that humans can leverage introspective knowledge to counteract internal sensory distortions, decoupling percepts from beliefs. Future work should manipulate expectations about illusion strength and encoding noise, refine task order and design, and extend the framework to characterize and remediate insight impairments in clinical populations (e.g., psychosis).

• Inference vs response-stage dissociation relies on the pattern of shifts in tandem (choice, confidence, RT, pupil) versus isolated psychometric shifts; although supported by prior and control data, some literature is mixed.
• Overcompensation observed (especially in Experiment 2) suggests potential miscalibration of illusion strength expectations (possibly influenced by training with longer adaptors); indicates suboptimality despite using an optimal strategy form.
• Fixed block orders for key conditions (Adapt-See before Adapt-Believe) could introduce practice or fatigue effects; analyses of RTs and DDM nondecision time do not support substantial confounding, but residual effects cannot be fully excluded.
• Modeling assumptions include MAE primarily affecting early encoding and simplified treatment of adaptor uncertainty (σ_A≈0 in the main model), though parameter recovery and simulations suggest robustness.
• Generalizability beyond MAE and to clinical populations remains to be tested; future designs should allow pre/post acquisition of illusion knowledge and manipulations of encoding noise (e.g., contrast) to further dissociate prior vs likelihood mechanisms.
• Sample sizes moderate (N=22 per experiment) and control pilot small (N=7 completers).