People are more error-prone after committing an error

People often respond more slowly immediately after committing an error (post-error slowing). Classic accounts attribute this to increased caution (e.g., higher decision thresholds or lower baseline activation), which should yield higher accuracy along a common speed-accuracy tradeoff. However, empirical work frequently finds that accuracy is unchanged or even reduced after an error, especially when intervals between successive trials are short. This has prompted proposals that errors trigger an initial orienting response or transient global inhibition that impairs processing before later, adaptive control can aid performance. A key limitation in prior work is reliance on free response time (RT) tasks, where RT conflates stimulus-response preparation with response initiation, making it difficult to isolate whether post-error changes reflect slower cognitive processing versus strategic delays in initiating responses. The present study asks why accuracy is often poor after errors despite slower responding and for how long such deficits persist, by dissociating preparation from initiation to characterize post-error processing directly.

Prior studies and models explain post-error slowing as increased response caution: neural network accounts posit reduced baseline response activation; evidence accumulation (drift-diffusion) accounts posit increased decision thresholds. These predict improved accuracy post-error along the same speed-accuracy function. Contradictory evidence shows accuracy often does not improve, particularly at short inter-trial intervals, motivating maladaptive accounts. Orienting-based theories (e.g., Notebaert et al.; Wessel’s adaptive orienting theory) propose that unexpected outcomes evoke an automatic, transient inhibitory response disrupting motor and cognitive processing, followed by adaptive control that may enhance attention. However, these frameworks often infer cognitive dynamics from RT distributions and assume RT reflects total processing duration, with strong assumptions (single accumulator, static drift). Recent motor control research shows preparation and initiation are independent, challenging inferences drawn from free RT alone and motivating interrogation methods that fix initiation to probe preparation over time.

Design and paradigm: Four online experiments (Prolific) used a forced-response 4-alternative forced-choice stimulus-response task. On every trial, participants were cued to initiate a keypress at a fixed time (about 2000 ms after trial start) using a rectangle fill cue, while stimulus onset varied uniformly between 0 and 2000 ms. This manipulation treated preparation time (PT = stimulus onset to response) as an independent variable to sample the time course of stimulus-response processing while holding response initiation fixed. Participants responded with designated keys mapped to four stimuli. Training: Participants first learned stimulus-response (S-R) mappings (60 trials, unconstrained RT with feedback), then practiced fixed-timing responses (20 trials) to respond exactly when the cue completed at 2000 ms with timing feedback. Main task: 10 blocks of 40 trials with fixed response timing and variable stimulus onset. Feedback and ITI varied by experiment:

• Experiments 1–2: Armenian letter stimuli; timing feedback for 1000 ms; ITI = 1000 ms; response keys f/g/h/j.
• Experiment 3: Colored circles; timing feedback only if mistimed (400 ms); ITI = 0 ms; response keys d/f/j/k.
• Experiment 4: Colored circles; timing feedback only if mistimed (400 ms); ITI = 2000 ms; response keys d/f/j/k. Participants: Analysis samples per figures: Exp1 n=46; Exp2 n=33; Exp3 n=47; Exp4 n=46. All protocols approved by University of Michigan IRB; informed consent obtained. Preprocessing: Analyses focused on test phase. Trials with mistimed responses (RT <1900 ms or >2100 ms) were excluded to ensure fixed initiation and avoid post-error slowing of response emission; this excluded 36.9% (Exp1), 30.6% (Exp2), 40.3% (Exp3), and 35.8% (Exp4). First trials of each block were excluded. Preparation time was rescaled to [0,1] by dividing by 2000 ms. Previous-trial outcome coded as −0.5 (previous correct) vs +0.5 (previous error). Behavioral analyses: Sliding-window conditional accuracy functions (100 ms window, 1 ms step) were computed as a function of PT, separately by previous-trial outcome. Hierarchical Bernoulli regressions (brms) assessed effects within PT ranges (e.g., PT <500 ms; PT >1000 ms), reporting posterior medians, 95% credible intervals (CI), and probability of direction (pd). Response preparation model: A hierarchical Bayesian model (Stan) captured the relation between PT and accuracy with minimal assumptions:
• μ: mean latency to complete stimulus-based response preparation (speed).
• σ: trial-to-trial variability of preparation latency (variability).
• β: efficacy, the probability that the prepared (correct) response is expressed once preparation is complete; 1−β reflects slips of action.
• α: probability of a correct response if preparation is incomplete (guessing rate). If preparation is incomplete at response time, responses are random (chance-weighted by α). Previous-trial outcome (error vs correct) entered as a covariate affecting μ, σ, and β via hierarchical delta parameters. Weakly informative priors were used. Model fit was assessed via posterior predictive checks and out-of-sample predictive performance (ELPD) comparing the full model (Δβ included) against a restricted model (no Δβ).

Across all four experiments, accuracy was lower following an error, including at long preparation times when a correct response should have been prepared:

• Conditional accuracy: For PT >1000 ms, strong post-error accuracy deficits were observed (e.g., Exp1 b = −0.63, CI [−0.98, −0.25], pd = 1.0; Exp2 b = −0.50, CI [−0.93, −0.05], pd = 0.99; Exp3 b = −0.87, CI [−1.20, −0.49], pd = 1.0). For PT <500 ms, effects were small or absent (near chance accuracy).
• Experiment-level overall post-error effects on accuracy (free of PT stratification) were negative: Exp1 b = −0.28, CI [−0.41, −0.15], pd = 1.0; Exp2 β = −0.11, CI [−0.24, 0.04], pd = 0.93; Exp3 β = −0.42, CI [−0.62, −0.22], pd = 1.0; Exp4 β = −0.23, CI [−0.43, −0.04], pd = 0.99. Model-based inferences (hierarchical response preparation model):
• Core parameters (group-level): Exp1 μ ≈ 523 ms (CI [502, 544]), σ ≈ 140 ms (CI [109, 175]), β ≈ 0.97 (CI [0.96, 0.98]). Exp3 μ ≈ 434 ms (CI [385, 479]), σ ≈ 202 ms (CI [155, 265]), β ≈ 0.94 (CI [0.92, 0.96]). Exp4 μ ≈ 434 ms (CI [396, 471]), σ ≈ 214 ms (CI [175, 257]), β ≈ 0.96 (CI [0.94, 0.97]). Exp2 similar to Exp1 (μ ≈ 512 ms, σ ≈ 141 ms, β ≈ 0.97).
• Efficacy decreases after errors (Δβ < 0): • Exp1: Mdiff = −0.02, CI [−0.03, −0.01], pd = 0.999. • Exp2: Mdiff = −0.01, CI [−0.02, −0.002], pd = 0.99. • Exp3: Mdiff = −0.05, CI [−0.08, −0.03], pd = 1.0. • Exp4: Mdiff = −0.01, CI [−0.03, −0.00], pd = 0.98.
• No credible evidence that preparation speed (μ) or variability (σ) slowed after errors: Exp1 μ Mdiff = 10 ms, CI [−40, 50], pd = 0.68; σ Mdiff = 10 ms, CI [−60, 50], pd = 0.59. Exp2 μ Mdiff = 10 ms, CI [−20, 50], pd = 0.71; σ Mdiff = 30 ms, CI [−20, 70], pd = 0.84. Exp3 μ Mdiff = 0 ms, CI [−50, 50], pd = 0.50; σ Mdiff = −20 ms, CI [−90, 60], pd = 0.73. Exp4 μ Mdiff = 30 ms, CI [−10, 80], pd = 0.93; σ Mdiff = 0 ms, CI [−60, 70], pd = 0.54.
• Inter-trial interval (ITI) modulated effect size on efficacy: Δβ larger at ITI = 0 ms (Exp3: −0.0535, CI [−0.0802, −0.0314]) than at ITI = 2000 ms (Exp4: −0.0141, CI [−0.0312, −0.00097]), indicating partial recovery with longer ITIs, but deficits persisted even with 2 s.
• Persistence across trials: Post-error effects on β were ‘reset’ by an intervening correct trial (little or no N−2 effect when N−1 was correct), but could persist/compound if the intervening trial was also an error (significant N−2 effects in some experiments).
• Nature of errors: Increased slips of action were predominantly perseverative repeats (repeating the previous button press).
• Timing errors: Mistimed responses and key-press errors were independent; timing errors did not drive post-error accuracy deficits.
• Model comparison (ELPD): Full model including Δβ outperformed a restricted model without Δβ across all experiments (relative weights: Exp1 0.900; Exp2 0.663; Exp3 0.928; Exp4 0.794).

Fixing response initiation with a forced-response paradigm allowed direct interrogation of stimulus-response preparation over time. Despite ample preparation time (up to 2 s), participants were less accurate immediately after errors. Modeling showed that post-error deficits are not due to slower or noisier cognitive processing (μ, σ unchanged), but to reduced efficacy (β): prepared responses were less likely to be expressed correctly, yielding more action slips, often perseverative. These findings challenge accounts that attribute post-error slowing solely to increased caution along a fixed speed-accuracy curve. Instead, the observed slowing in free RT paradigms may be an adaptive initiation delay that compensates for a transient, error-induced impairment in the ability to translate prepared decisions into correct actions. Although adaptive orienting mechanisms may enhance attention after initial impairment, they did not eliminate the post-error efficacy loss observed here. Longer ITIs attenuated but did not abolish the deficit, and effects could persist across consecutive error trials but reset after an intervening correct response. The results suggest a changed relationship between speed and accuracy after errors, with an elevated propensity for slips of action rather than a uniform shift along a common tradeoff function.

Across four experiments using a forced-response design, people were more error-prone immediately after committing an error, even with ample time to prepare a response. Computational modeling revealed that post-error impairments reflect reduced efficacy in expressing prepared responses (increased slips of action), not slower or more variable preparation. These findings revise interpretations of post-error slowing: delays in free RT may be adaptive strategies to counteract impaired efficacy rather than evidence of slowed cognitive processing. Future work should identify the cognitive and neural sources of efficacy loss (e.g., action selection vs. mapping retrieval), determine how adaptive orienting and control interact with these impairments, test generalization across task domains (with and without conflict/distraction), and refine models that jointly capture preparation and initiation decisions.

• Distributional assumption: The model assumes preparation latency is approximately normally distributed. Although fits were good under forced deadlines, deviations such as increased skewness could, in principle, affect parameter interpretations.
• Paradigm generalizability: Forced-response tasks may alter engagement and processing relative to free RT paradigms; participants must monitor timing cues, potentially increasing task demands.
• Selection vs initiation independence: Conclusions hinge on the assumption that preparation and initiation are independent parameters; while supported by prior motor control work, this may not fully capture all decision contexts.
• Pre-/post-error control analyses: With accuracy as the sole dependent variable under forced initiation, standard free-RT controls (e.g., examining pre-error RT) cannot be directly replicated, making it harder to fully rule out multi-trial attentional lapses. Nevertheless, independence between timing and key-press errors argues against a general inattentiveness account.
• Task domain: Prior adaptive orienting evidence often comes from conflict tasks; effects and mechanisms may differ in tasks without distractors.