Surprising sounds influence risky decision making

The study investigates whether task-irrelevant sensory prediction errors (unexpected auditory events) systematically influence risky decision making in humans. While unpredictable sensory events are salient and can attract attention, prior work leaves open whether their influence reflects mere distraction (increased decision noise) or induces stereotyped, value-independent behavioral responses (e.g., uncertainty seeking and option switching). The authors hypothesize that incidental auditory surprise will increase risk taking and decrease choice perseveration, potentially reflecting mechanisms linked to neuromodulatory systems (e.g., dopamine, norepinephrine). They leverage a risky decision-making paradigm where option values vary unpredictably across trials and are independent of the auditory sequences, allowing separation of incidental sensory effects from value-based influences.

Prior literature suggests that unexpected events can motivate exploratory behavior and facilitate detection of environmental changes. Task-irrelevant sensory unpredictability has been associated with impaired performance via distraction (e.g., visual salience changes, auditory frequency or loudness changes). However, evidence from learning paradigms shows that uncertainty can induce exploration, and neuromodulatory systems may link surprise to behavior. Dopamine neurons respond to value-neutral sensory surprises and identity violations; pharmacological manipulations of dopamine can increase value-independent gambling propensity and reduce choice perseveration. Noradrenergic modulation can affect risk taking in rodents when salient cues accompany outcomes. These findings motivate testing whether incidental sensory prediction errors induce value-independent biases in risk taking and perseveration in a decision task without learning.

Design: Seven online experiments (total n = 1600) conducted via Gorilla Experiment Builder with participants recruited on Prolific (fluent English, age ≥18). Headphone screening ensured standardized audio presentation. Compensation was flat for Experiments 1–6; Experiment 7 used incentive-compatible payment. Ethical approval from Yale IRB (protocol #2000028824). Task: On each trial, participants chose between a risky and a safe option aiming to maximize points. Risky gambles had equal probabilities of win/loss; outcomes appeared after a 1.3 s delay and were shown for 1 s. Three trial types: Gain (sure gain vs risky larger gain or 0), Mixed (0 vs risky gain or loss), Loss (sure loss vs risky larger loss or 0). Options varied unpredictably across trials. Median RT across experiments ≈ 1.34 s. Auditory manipulation: Before options in every trial, participants heard a six-tone sequence (100 ms tones, 400 ms ISI) composed of A=1000 Hz, B=1500 Hz, C=500 Hz tones. In Experiments 1–4 and 7, 75% of trials used a common sequence and 25% used a rare sequence; sequences differed in the final tones. In Experiments 5–6, two sequences occurred equally often (50/50), either deterministically alternating (Exp. 5) or randomly shuffled (Exp. 6). Participants were informed about common/rare identities and identified the common sequence with ~95.5% post-task accuracy in Exps 1–4. Experiment specifics:

• Exp. 1 (n=200): Common AAAAAA (75%); Rare ending deviant AAAAAB or AAAAAC (25%). Risky option always on left. 192 trials (48 unique trials repeated 4×; each unique trial once Rare, thrice Common).
• Exp. 2 (n=200): Same sequences as Exp. 1; risky/safe sides alternated every 10 trials. 120 trials.
• Exp. 3 (n=200): Rare sequences ended on standard tones (AABAAA; 25%); common sequences ended with deviant (AABAAB; 75%). 120 trials.
• Exp. 4 (n=200): Rare sequence AAAAAA (25%); common AAAAAB (75%). 120 trials.
• Exp. 5 (n=200): Balanced 50% AAAAAA and 50% AAAAAB alternating deterministically. 120 trials.
• Exp. 6 (n=200): Balanced 50% AAAAAA and 50% AAAAAB shuffled randomly. 120 trials.
• Exp. 7 (pre-registered; n=400): Replication of surprise effects with incentive-compatible payment and no on-screen earnings display; included post-task belief query regarding whether tones predicted win probabilities. Design matrix: 48 unique trials (16 Gain, 16 Mixed, 16 Loss) constructed from certain amounts and multipliers; Experiments 2–7 used 120 trials assembled from sets ensuring coverage of multipliers with 1:3 rare:common assignment and pseudorandom order enforcing 3 rare per block of 12. Modeling: Choices fit per participant via Maximum Likelihood using Prospect Theory utilities with a softmax. Base parameters: loss aversion λ, risk aversion for gains α_gain and losses α_loss, and inverse temperature μ (or σ for stochasticity formulations). To capture value-independent influences:
• Risky Bias model adds a riskybias parameter to softmax; Perseveration model adds a perseveration parameter biasing repetition of prior choice [persev × (Choice_risky(t−1) − Choice_safe(t−1))].
• Difference parameters δ capture changes on Rare vs Common (e.g., δ_riskybias, δ_persev); alternative models allowed δ on value-dependent parameters, lapse rate, or stochasticity. Model comparison used AIC; simulations and parameter/model recovery assessed identifiability and goodness-of-fit. Nonparametric statistics (Wilcoxon signed rank/sum) used throughout; Bayes Factors computed where appropriate. No multiple-comparisons corrections were applied. Data and code available on OSF (pre-registration for Exp. 7: OSF DOI 10.17605/OSF.IO/PA75Q).

• Rare auditory sequences increased risk taking when rare sequences ended with deviant tones (Experiments 1–2): • Exp. 1: Risky choices +2.06 ± 1.09% (mean ± SEM) on Rare vs Common, p = 0.005. Overall risky choice rate 53.16 ± 1.17%. • Exp. 2: Risky choices +4.03 ± 1.06%, p < 0.001. Overall 47.54 ± 0.98%; slower RTs vs Exp. 1. • Effect present across trial types in combined n=400: Gain +3.09 ± 1.01% (p < 0.001), Mixed +2.75 ± 0.97% (p = 0.010), Loss +3.17 ± 0.88% (p < 0.001). • Effect present both when model-predicted P(risky) ≥ 50% (+1.61 ± 0.83%, p = 0.005) and < 50% (+3.67 ± 0.88%, p < 0.001). • δ_riskybias > 0 in both Exps 1–2, consistent with value-independent increase in gambling propensity:
  • Exp. 1: δ_riskybias = 0.125 ± 0.052, p = 0.004.
  • Exp. 2: δ_riskybias = 0.195 ± 0.056, p < 0.001.
• Rare auditory sequences decreased choice perseveration (rate of staying with previous option) in Exps 1–2: • Exp. 1: Stay rate Rare−Common −3.69 ± 0.81%, p < 0.001 (overall stay 59.10 ± 0.72%). • Exp. 2: −3.62 ± 0.89%, p < 0.001 (overall stay 54.91 ± 0.63%). • Present across trial types (combined n=400): Gain −3.51 ± 0.89% (p < 0.001), Mixed −4.63 ± 0.87% (p < 0.001), Loss −2.68 ± 0.82% (p = 0.006). • Present when model predicted staying and not staying: −5.14 ± 0.81% (p < 0.001) and −1.57 ± 0.73% (p = 0.019), respectively. • δ_persev < 0 in both Exps 1–2 in the winning model:
  • Exp. 1: −0.149 ± 0.041, p = 0.001.
  • Exp. 2: −0.126 ± 0.048, p = 0.010.
• Alternative explanations ruled out by modeling: • Lapse Rate Difference model: δ_lapse > 0 (0.050 ± 0.013, p < 0.001) but failed to reproduce increased risk taking and decreased staying across trial types. • Stochasticity Difference model: δ_σ = 0.054 ± 0.044 (p = 0.156, BF01 = 8.42); simulations failed to reproduce observed effects. Full Risky Bias Perseveration Difference model preferred by AIC.
• Dissociation by auditory statistics (Experiments 3–4, n=400): when rare sequences ended on standard tones and common sequences ended on deviant tones, the risk-taking increase disappeared while perseveration decrease persisted. • Exp. 3: Risk-taking Rare−Common −2.50 ± 1.83% (p = 0.51); staying −6.00 ± 1.02% (p < 0.001). δ_riskybias −0.020 ± 0.091 (p = 0.913, BF01 = 17.40); δ_persev −0.314 ± 0.058 (p < 0.001). • Exp. 4: Risk-taking −3.23 ± 1.43% (p = 0.075, BF01 = 1.07); staying −4.27 ± 0.87% (p < 0.001). δ_riskybias −0.067 ± 0.072 (p = 0.575, BF01 = 11.48); δ_persev −0.172 ± 0.049 (p < 0.001).
• Balanced/predictable sequences eliminate both effects (Experiments 5–6, n=400): no differences in risk taking or staying between AAAAAA and AAAAAB sequences across all trial types; δ_riskybias and δ_persev not different from zero in both experiments. • Exp. 5: risk-taking −0.98 ± 0.96% (p = 0.494, BF01 = 7.55); δ_riskybias −0.015 ± 0.053 (p = 0.932, BF01 = 17.13); δ_persev 0.007 ± 0.040 (p = 0.914, BF01 = 17.53). • Exp. 6: risk-taking 0.93 ± 1.12% (p = 0.635, BF01 = 8.98); δ_riskybias 0.029 ± 0.057 (p = 0.698, BF01 = 15.70); δ_persev 0.007 ± 0.035 (p = 0.802, BF01 = 17.49).
• Belief control (Experiment 7, pre-registered, n=400): replicated positive δ_riskybias (0.109 ± 0.055, p = 0.028) and negative δ_persev (−0.208 ± 0.040, p < 0.001). Effects persisted after excluding 89 participants who believed tones predicted rewards: δ_riskybias 0.102 ± 0.058 (p = 0.041), δ_persev −0.180 ± 0.047 (p < 0.001). Overall, surprising sounds produced value-independent increases in risk taking and reductions in perseveration that depend on auditory statistical structure, not on increased decision noise or lapses, and are dissociable across manipulations.

The findings demonstrate that task-irrelevant sensory prediction errors have systematic and dissociable influences on human decision making under risk: they increase a value-independent propensity to choose risky options and reduce value-independent choice perseveration. These effects persist across gain, mixed, and loss contexts and across trials where the model predicts either risky or safe choices, indicating that they are not mediated by changes in choice stochasticity or attentional lapses. Modeling shows that two value-independent parameters—riskybias and perseveration—capture these influences more parsimoniously than alternatives involving lapse rates or stochasticity changes. Dissociation across auditory statistics suggests distinct mechanisms: decreased perseveration arises from recognition of rare sequences per se (global sequence-level surprise), whereas increased risk taking depends on the rarity of local deviant features (e.g., deviant tones or rare tone-to-tone transitions). This aligns with theories linking sensory surprise to neuromodulatory systems, particularly dopamine, which responds to value-neutral sensory deviations and has been implicated in value-independent gambling bias and reduced perseveration. The results suggest that incidental sensory surprise can engage exploration-related processes (directed for risk taking; random for decreased perseveration) even when such behavior is not adaptive in the task. These effects have broad implications: in real-world environments rich with unpredictable sensory events (e.g., trading floors, casinos, busy urban settings), incidental sounds could bias risk preferences and choice switching independent of option values, introducing a previously unrecognized source of variability in consequential decisions.

Across seven experiments (n=1600), rare, task-irrelevant auditory sequences biased risky decision making in two ways: they increased value-independent risk-taking propensity and reduced value-independent perseveration. These effects were dissociable via auditory statistical manipulations, eliminated when sequences were fully balanced/predictable, and persisted in participants who reported that tones did not predict reward. Computational modeling identified parsimonious value-independent parameters capturing these influences, arguing against explanations based on attentional lapses or changes in choice stochasticity. Future work should: (1) test causal neuromodulatory mechanisms using pharmacology and neurostimulation/optogenetics in animal models timed to sensory surprises; (2) examine generalization across sensory modalities and tasks involving learning; (3) investigate boundary conditions and individual differences; and (4) evaluate ecological validity in field settings (e.g., financial trading, consumer behavior, casinos) and potential clinical applications assessing relationships between sensory prediction errors and mental health.

• Power for detecting hypothesized negative risk-taking effects in Experiments 3–4 was limited due to higher overall frequency of deviant tones (reduced rarity), with authors noting substantially larger samples would be required to detect such effects.
• No corrections for multiple comparisons were applied in statistical analyses, increasing the chance of Type I error for multiple tests.
• Mechanistic inferences about neuromodulators (e.g., dopamine) are indirect; no neural measurements or causal manipulations were included.
• Generalization beyond auditory modality and to real-world environments remains to be empirically demonstrated.
• Online data collection, while controlled with headphone screening, may introduce variability in audio presentation and participant environments.