Believing in simulated virtual scents produces reported olfactory sensations

Abstract

Predictive processing (PP) refers to the brain’s ability to incorporate present and prior sensory information about states of the body and the environment, enhancing various cognitive functions including perception, motor control, decision-making, and theory of mind. The PP framework conceptualizes the brain as a probabilistic prediction engine that continuously generates top-down predictions about the causal structure of the world. This approach has provided significant explanatory power, offering what many consider a first glimpse of a unified theory of the mind. However, olfactory perception and sensation in the context of PP remains an understudied subject. We present substantial evidence for olfactory sensations as being constructed via predictive processes. We discovered, in a pre-registered and controlled behavioral experiment, that when participant’s short-term priors of olfactory sensations are manipulated by instilling a belief that olfactory stimuli have been modelled as part of a virtual reality experience, participants report diverse olfactory sensations. The results demonstrate how mind’s capability of constructing reality extends from previously studied sensory modalities such as vision or taste to olfaction. The study initiates the framework for refined experimental designs on predictive olfaction utilizing virtual stimuli and potential applications for therapeutic treatments of individuals with deficiencies in olfactory functioning or neurogenerative diseases.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-25418-1.

Introduction

The predictive processing (PP) paradigm, along with its leading neuroscientific model of neural processes, predictive coding (PC), posits a strong connection between prediction and perception. Both, PP and PC, claim our brains begin the process of transforming sensory information into coherent perceptual experience before actual physical exposure to stimulus^1–5.

The concept that the observer’s mind actively contributes to the construction of perception has a long history. The philosophical roots of PP could be traced all the way to idealism, which emphasizes the mind’s influence on epistemology (knowledge) and experience. However, German physicist Herman von Helmholtz’ (1821–1894) constructive theory, according to which human perception is not solely based on sensory input but also involves the brain’s unconscious inference that utilizes prior knowledge, is often recognized as the founding premise of modern PP^6,7. Contemporary PP proposes that the brain acts as an active hypothesis-testing mechanism, countering the noisy and ambiguous nature of sensory input by generating models of the causes behind sensory events. While much of this predictive foundational knowledge is acquired through learning, the brain is also shaped and pre-configured to some extent by evolutionary processes^6,8,9.

According to PP, perception is being constructed from the dynamic interaction between top-down predictions about the causes of sensory input and bottom-up prediction errors^10–12. If the sensory data matches the predictions, no further processing is needed. If there is a mismatch, bottom-up prediction errors signal the need to update the brain’s generative model, leading to refined and adjusted perceptions. Prior knowledge is thus incorporated into brain’s continuous “workflow” of testing different perceptual hypotheses. This represents a substantial conceptual shift, as PP diverges significantly from traditional models that depict sensory information as progressing along a linear, bottom-up pathway through increasingly complex feature detectors¹³. For instance, according to the hierarchical feedforward model (HFM) of object recognition in the visual system, visual input processing begins with neurons in the primary visual cortex (V1) detecting basic features, such as edges and orientations. These features are then progressively combined in higher visual areas, such as V2 and V4, to extract more complex attributes like textures, shapes, and contours. Ultimately, in regions like the inferotemporal cortex, these integrated features culminate in the recognition of whole objects, such as faces, tools, or animals¹⁴.

However, higher-level configurations influence the perception of lower-level features, contradicting the idea of strictly feedforward processing^13,15. In relation, it has been identified that in the human visual cortex there are generally more feedback connections than forward connections¹⁶. Also, many cognitive processes related to memory, imagination, and future prediction, while traditionally considered distinct, are now acknowledged being supported by overlapping predictive neural systems, such as the hippocampus’s role in the implicit and continuous prediction of the environment which enhances our ability to construct scenes that extrapolate beyond the immediate sensory input¹⁷.

Instead of direct opposition to previous models such as HFMs, contemporary PP should be seen as an addition or expansion to the prior understanding provided by them. Cognitive models such as HFMs have provided foundational insights into how sensory systems build complex representations from basic input and they have also inspired computational approaches, such as convolutional neural networks (CNNs), which mimic this hierarchical layered processing structure¹⁶. However, the view is limited in its ability to account for the dynamic, interactive nature of perception. In underemphasizing the role of top-down processes, such as prior knowledge, expectations, and contextual influences, which are now recognized as integral to perceptual experience, such models are not painting the full picture on how perception works.

A critical need for further empirical research in PP lies in better understanding the anticipatory neural and behavioral mechanisms of perceptual information encoding, as these processes remain only partially understood. Several prominent models of sensory perception suggest that the brain forms pre-stimulus templates, or “search images”, prior to encountering a stimulus^4,18,19. Numerous related studies have demonstrated the role of anticipatory attention in both the visual and auditory systems^20–23. More recent research has extended these findings to show that anticipatory attention or top-down search processes also influence the perception of specific visual objects in higher-order visual cortex^24–27. Recently, Pierzchajlo et al. (2024) found that olfaction is highly dependent on vision, however also suggesting that processing olfactory prediction errors relies on cross-sensory resources in a way that visual prediction errors do not²⁸. However, although some research designs seek to manipulate the prior expectations of participants, this has been conducted with long-term subtle conditioning to stimulus. The effects of manipulation on short term expectations has remained an open question. This is unfortunate, given the fast, dynamic, complex, and competitive environments in which humans and other animals must experience, process and rely on senses such as vision or olfaction. Short term expectations play a crucial role in enhancing the predictive mechanisms for selective attention, helping to filter out sensory noise and focus on what is relevant in specific action-oriented events or settings^29,30. Considering the need for rapid decision-making, the most compelling example of a putative prediction error signal is the robust neural response elicited by the omission of an expected stimulus. Numerous studies involving both humans and animals have been conducted to investigate this, however mainly concentrating on the visual and auditory modalities^31–35.

Olfaction is a unique sensory modality. It is important to acknowledge that while the role of olfaction is still being fully delineated, established theoretical frameworks suggest its functions extend beyond basic perception. These hypotheses propose that olfaction evolved to support spatial navigation³⁷, help organisms evaluate the hedonic value of an odor for approach or avoidance behaviors³⁸, and facilitate social interactions^38,39. Several structural features also set olfaction apart from other sensory systems. The peripheral olfactory system exhibits low substrate specificity, as individual receptors can detect multiple odorants, and individual odorants can be recognized by multiple receptors⁴⁰. Olfactory information has been recognized to be primarily projected ipsilaterally to central structures and was previously thought to be the only sensory input directly integrated into cortical regions with little prior processing in the thalamus⁴¹.

Neuroscientific or psychological research on PP and PC in odor perception and olfactory system has been unfortunately scarce. There are still some exceptions. Freeman (1981) proposed that “search images” in the rabbit olfactory bulb act as filters that enhance sensitivity to expected odors without reducing responsiveness to others⁴². These early studies suggest that olfactory perception is strongly shaped by expectations. Some studies have identified similar kinds of anticipatory neural events in the olfactory bulb, piriform cortex, and orbitofrontal cortex, also indicating that predictions are made before odors are actually/physically encountered^43,44. In recent psychological research, some of the most interesting and relevant evidence of the constructive and predictive side of olfaction has come from the studies on the influence of color and language on the perception of wine aromas^41,45,46. This research has focused on investigating what is known as the olfactory perceptual bias. Findings from these studies underline how color (and language) influences perception of odor and taste, also showing strong evidence that conceptualization and categorization of a beverage such as wine is not achieved sole by relying on sensorial information form the nose and mouth⁴⁷.

In the PP paradigm, perception is many times described as “controlled hallucination”⁶. The “hallucination” refers to the constructive ability of our minds in producing sensations on the base of predictions, while “control” refers to the minds integration powers in producing models that allows for useful interactions with the world. Although integration relies many times on information flow from multiple sensory modalities, this is not always the case. For example, although psychological studies have argued that people in general are poor at identifying substances by olfaction in the absence of cues from other sensory modalities such as vision⁴⁸, experiencing odors in complete omission of stimulus (known in medical literature as “phantosmia”) is known to be common for many psychiatric disorders, such as depression, schizophrenia, and bipolar disorder, as well as neurological diseases such as epilepsy or migraine^49–52. Recently, Covid-19 infection has been reported also to cause decreased ability to smell (“hyposmia”), and the absence of the ability to smell (anosmia), as well as persisting auditory and visual hallucinations^53–55. Although the linkage between prediction and hallucination is interesting, however, at this point it is sufficient enough to recognize that in the framework of PP, the occurrence of perceptual hallucinations are considered a predictable outcome when the perceptual system generates prediction errors, leading to the emergence of illusory yet vivid and seemingly real perceptual phenomena, such as visual or olfactory hallucinations. Similarly, in the PC framework, these processes can be understood as the prediction based neural activations and/or mappings in the brain.

Hallucinations in general have been studied the most in relation to the visual system. Visual hallucinations (VH), or visions, have been defined as visual percepts that are experienced when fully conscious but in the absence of the corresponding external stimulus⁵⁶. VH are common in a number of neurodegenerative disorders. According to recent studies, up to 93% of people with Lewy body dementia report VH⁵⁷, as do up to 75% of people with Parkinson’s disease^58,59. However, especially people with psychosis report high rates of hallucinations across modalities such as auditory, olfactory and tactile senses. It is however well know, that even without a severe neurodegenerative disorder the loss of the sense of smell is associated with distinct deterioration in the quality of life and depressive symptoms⁵³. This underscores the significance of investigating sensory hallucinations associated with olfaction and their relation to PP. Notably, in the medical evaluation, diagnosis, and treatment of olfactory function, despite significant efforts to develop standardized and validated tests, the primary treatment approaches remain olfactory training with actual stimuli, pharmacological interventions such as corticosteroids, or invasive procedures like surgery^53,60,61. Since these approaches do not adequately account for the predictive nature of olfactory perception, they may risk being either overly invasive or only moderately effective as treatment options.

Virtual reality as a tool for enhancing ecological validity in perceptual research

The past neuroscientific research, primarily focusing on visual and auditory modalities, has been overlooking important sensory systems like olfaction, touch, and interoception. Also, while lab studies are essential for understanding specific neural mechanisms related to PP and PC, their artificial and simplified nature frequently lacks ecological validity.

Virtual reality (VR) is a technology that allows the user to experience physical and/or behavioral simulations (e.g., interacting with objects and environments)⁶². It offers researchers an enhanced possibility to test hypotheses by simulating experiences that otherwise would be difficult or even impossible to create⁶³. VR makes use of virtual environments to present digitally recreated real-world activities to participants via immersive head-mounted displays. Virtual environments provide experimental control and dynamic presentation of stimuli in ecologically valid research designs, since they allow for controlled presentations of emotionally engaging stimuli to enhance affective experience and social interactions^64,65. The ecological validity of VR has already been demonstrated in multiple fields, such as social-, clinical- and affective neuroscience and neuropsychology^66–70.

Although the usage of VR in investigating olfactory function has been scarce, some studies have proposed it as a compatible methodological choice for studying olfaction. Javerliat et al. (2022) aimed to address the lack of standardized tools for olfactory experiments in VR by developing an open-source olfactory display capable of diffusing scents. Their device could successfully create three significantly different perceived odor intensities, offering a reproducible design for VR olfaction research⁷¹. Similarly, Niedenthal et al. (2023) created their own novel olfactometer device for studying olfaction in VR. Their device was successfully tested in a VR wine-tasting game, results indicating that such a device could be used for delivering scents and successfully integrating them into a VR experience⁷². The recent findings of Trabanelli et al. (2025) also suggest, through a behavioral, neural, and immune study, that humans anticipate potential infection threats by activating the immune system without any physical contact when confronting infectious VR avatars³⁶. None of these studies however directly focused on the predictive nature of olfaction.

This study highlights notable limitations in research on predictive processing in relation to olfaction: 1) a limited volume of research on olfactory perception, (2) a scarcity of behavioral research designs examining the impact of manipulating short-term prior expectations of perception, (3) the need for valuable insights for developing non-invasive, more effective treatments for restoring olfaction by better aligning therapeutic strategies with the brain’s olfactory sensory processing mechanisms, and (4) a lack of research using new, ecologically valid methods, such as VR. This study was designed to help address these gaps.

Main aim, hypotheses and an overview of the study

The main aim of this study is to test the effects of short-term prior expectations on odor perception in a VR setting (addressing limitation 1 and 2). The hypothesis is, that when short-term prior expectations of participants in the experimental group are altered by planting a belief that their VR experience also includes simulated scents, they will report and describe experiences of odors related to the VR stimulus (Amazonian jungle) since the expectations become integrated into their current scent-model. We also hypothesize that participants in the control group, whose expectations aren’t being manipulated, won’t give descriptions since they rely on their existent non-altered scent-model. The hypothesis was examined in adherence to preregistration by observing both the nominal coded open questions and continuous self-reported realism in olfactory quality as predicted by manipulation condition. Both dependent variables were tested separately and by controlling for effects of other independent variables.

The manipulation of expectations for the experimental group is achieved by giving them both a textual manipulation (information of simulated scents in their upcoming VR experience) and a visual manipulation (presenting a faux olfaction modelling apparatus). The design is explained in depth in the upcoming section Experimental setup. This design creates a methodology for studying anticipatory predictive processing of olfaction with VR, that is both ecologically valid and creates possibilities for innovative future therapeutic applications (addressing limitation 3 and 4).

Methods

Ethics & inclusion statement

This study was conducted in accordance with the ethical guidelines for research with human participants and followed principles that prioritize maximizing benefits and minimizing potential harms. The research was preapproved by the Ethics Committee in the Humanities and Social and Behavioral Sciences of the University of Helsinki (statement 67/13.9.2023). Participants gave their written consent before onset of the experiment and received a movie ticket as compensation. At the conclusion of the study, participants answered questionnaires and were then fully debriefed regarding the true nature of the experiment. All research data was anonymized to avoid stigmatization, incrimination, discrimination or any other personal risk to participants. The research included local researchers throughout the research process, and roles and responsibilities were agreed amongst collaborators ahead of the research.

Data availability and statement

For transparency and scientific replication, this research was preregistered through the Open Science Framework (OSF): https://osf.io/94ugn/?view_only=7a38487cfb6f472f814b0a94bff1b3d9). The data and complete statistics of the statistical models and their syntax as well as power calculations can be found from the project public GitHub repository: https://github.com/berginator/predictiveolfaction-study.

Participants

Thirty adult participants (21 female) took part in the experiment (university students, university employees and faculty). Majority of the participants completed the experiment in Finnish, while four participants received instructions and forms in English. None of the participants reported neurological (e.g. migraine, epilepsy, sensory hypersensitivity) or cardiac diseases or disorders, medication affecting the central nervous system, mental illness (excluding non-acute history of mild depression or anxiety) and learning difficulties or other developmental disorders (e.g. reading difficulties, ADHD).

Statistical power

Sample size was determined according to stopping rule laid out in preregistration: we set a cut-off point after 30 participants, and an evaluation of effect against the main hypothesis. If there is no effect, the experiment will be terminated. In addition, post-hoc statistical power was estimated using a simulation-based approach to determine the likelihood of detecting a significant difference between the control and manipulation groups. The simr package in R was utilized to perform power calculations on a linear mixed model with a fixed effect for group (control vs. manipulation) and a random intercept for participants. The model was designed to capture individual variability in responses, with hypothesized effect sizes and variance parameters set based on prior expectations.

The power simulation was conducted with 100 iterations, testing whether the inclusion of the group variable significantly improved model fit compared to a null model. The results indicated an estimated power of 87.00%, suggesting that an optimized sample size of 30 (15 participants per group) provides sufficient power to detect a large effect size with an alpha level of 0.05. This power level is above the conventional threshold of 80%, indicating a high probability of identifying a true effect between groups, if present, in this experimental setup.

Experimental setup

The experiment was conducted in an ordinary room, where participants sat at a table and were instructed to “experience a virtual jungle for a one-minute period with your senses open”. Upon arrival to the laboratory, participants were pseudo randomly assigned between manipulation (n = 16) and control (n = 14) conditions. Before starting the experiment, the participants in the manipulation condition received a presentation of a faux olfactory modeling apparatus (see Fig. 1) along with a textual description in the task instructions reading as follows: “Apart from sounds, we have also modeled the jungle to have smells, in order to create a maximally entertaining VR-environment. Due to modeling, some vapors will be released to the room now”. Participants in the second control condition received equal instructions albeit in absence of the aforementioned demonstration and task instruction sentence. After reading the instructions, participants were assisted to put on a VR headset (HTC Vive Pro II). Stimulus was presented using SteamVR video player and consisted of a one-minute clip depicting the Amazonian jungle in 360°. The clip was presented only two times, resulting in a total of two minutes of exposure. This strategy was chosen to avoid habituation to stimulus. Roughly half of the participants (n = 16) completed the experiment in springtime [temperature (Celsius), M(SD) = 22.16(0.67); moisture (%), M(SD) = 26.85(3.51)] and the rest (n = 14) in autumn at different location [temperature, M(SD) = 23.39(0.55); moisture, M(SD) = 52.65(5.38)]. The experiment rooms were comparable in terms of size, lightning conditions and layout. The room was ventilated after each participant and only participants that had not noticeably used scented products were chosen for the experiment.

Fig. 1.

The faux olfaction modelling apparatus used for manipulation. White diffusor produced water steam, and the black, red and yellow cables were affixed to the bottom of the diffusor with duct tape. Raspberry Pi 3b computer on top of an empty black case provided additional cues.

Independent variables

Four self-report measures were included as covariates to assess possible confounding factors related to dependent variables. Differences in participants’ immersion during the VR-experience were measured with a modified version of scale introduced by Jennett et al. (2008), denoted onwards as JIS (⁷³). Items related to gaming were removed, resulting in 19 items on 1–5 scale with varying endpoint descriptions. Creative Experiences Questionnaire (CEQ⁷⁴) was used to assess participants’ fantasizing disposition, consisting of 25 true/false statements. Olfactory sense’s importance to participants was assessed with the Individual Significance of Olfaction scale (IO-Q⁷⁵). It consists of 20 items describing various cognitive phenomena measured on 1–4 scale, separated into three six-item factors (association, application and consequence) and a two-item factor (aggravation). It is associated with measured olfactory function (⁷⁶). Finally, a 13-item short form of Marlowe-Crowne Social Desirability Scale (MCSDS⁷⁷) was used to assess participants’ desire for social approval. All scales exhibited adequate to good internal reliability, assessed with the omega coefficient (IO-Q and MCSDS, ωt = 0.75; CEQ, ωt = 0.78; JIS, ωt = 0.87). Responses were screened for quality by calculating intra-individual response variability (IRV⁷⁸) for two scales, IO-Q and JIS. All responses lied within +−3 SD from the mean, suggesting no extremely straight lining or random responding patterns. The descriptive statistics (Table 1) reveal that all variables, except for CEQ with slight positive skewness, exhibit near-normal distributions with minimal skewness and kurtosis deviations. Low standard errors across all variables indicate precise and reliable mean estimates.

Table 1.

Descriptives of the four continuous independent variables (N = 30).

Variable	Mean	SD	Median	Min	Max	Range	Skew	Kurtosis	Se	ω
MCSDS	18.13	2.69	18.0	13	24	11	0.12	−0.63	0.4	0.75
IO-Q	50.93	6.97	50.5	35	67	32	−0.16	0.13	1.2	0.75
CEQ	8.83	4.15	8.0	2	21	19	0.59	0.43	0.76	0–78
JIS	63.43	9.27	64.5	45	80	35	−0.3	−0.95	1.69	0–87

Besides continuous measures, participants were asked about their prior usage of VR-devices (once or twice, n = 17; occasionally, n = 3), and staying in real jungle environments (n = 1). In addition, the degree of habitual caffeine consumption and elapsed time from last dosage was collected, as caffeine can induce auditory hallucinations⁷⁹. Participants were classified as under caffeine influence when they reported intake less than four hours prior to experiment.

Dependent variables

The dependent variables of the study consisted of participants’ reported sensory perceptions in response to virtual environment and experimental design conditions. Quantitative indices included one general item “Was your experience impressive or compelling?” (VR image quality). Four specific items were introduced with the sentence “How realistic or convincing was your experience, regarding:”. It was followed with descriptions of audiovisual (“Image quality and visual modeling?”, VR video), auditive (“Audio quality and soundscape?”, VR audio), olfactory (“Scent, smell and olfactory sensation?”, VR scent) and bodily (“Other bodily sensations, for example on your skin or in bottom of your stomach?”, VR body) sensations. Items were answered using 1–10 scale (1: “not at all”, 10: ”very much”).

To supplement the quantitative self-report measures, participants answered three open questions: “Briefly describe your sensations during the VR-voyage regarding any possibly experienced scents and smells”, “Was there a scene or an event in the jungle, where the scents and smells were strongest?” and “Were the scents or smells in general even overpowering or distracting? In what way?”. These three items were introduced in the questionnaire page as being randomly selected to give the impression that the participant had been assigned to report experiences regarding olfactory sensations (and not to some other sense-modality group). The provided answers were coded according to certainty/uncertainty of olfactory sensation or certainty of no olfactory sensation. The original answers along with their English translations and their assigned codes are included in Appendix 1.

Results

There were significant associations between manipulation condition with original certainty classes (Fisher’s exact test; p = 0.004), exclusive certainty (not including uncertain answers as positive, p = 0.003) and inclusive certainty (including uncertain answers as positive, p = 0.011). In a quasibinomial generalized linear model predicting exclusive certainty with manipulation condition together with the four sum scales (MCSDS, IO-Q, CEQ, JIS), only the effect of condition was significant (OR = 23.66, p = 0.010, model R² = 0.389). The effect was milder but still significant when using inclusive certainty as a dependent variable (OR = 10.81, p = 0.034, model R² = 0.246). The distributions are depicted in Table 2, along with a word cloud (Fig. 2) further detailing the reported sensations.

Table 2.

The distributions of olfactory sensations between control and manipulation groups.

	Certainty of no sensation	Uncertainty of sensation	Certainty of sensation
Control group	11	1	2
Manipulation group	3	3	10

Fig. 2.

Word cloud representing the described olfactory sensations.

In continuous assessments of VR experience’s quality (Fig. 3), there was a significant difference between control [M(SD) = 1.86(1.29)] and manipulation [M(SD) = 3.56(2.42)] groups in olfactory modality [t (23.49) = 2.45, p = 0.022, d = 0.89], with participants in manipulation group reporting higher olfactory realism.

Fig. 3.

Trumpet plot indicating distributions of continuous assessments of VR experience’s quality in general (A) and per four sense modalities (B-E) in control and manipulation groups.

There were no differences between groups regarding experiment in total [t (27.21) = 0.16, p = 0.870] or with regards to visual [t (27.96) = 0.12, p = 0.903], auditory [t (25.35) = 0.87, p = 0.394] sensation quality. The assessments of quality regarding bodily sensation were similarly low compared to olfactory sensation qualities, but there were no differences between the groups [t (25.34) = 0.07, p = 0.943].

To control for effects due to potential confounding variables, a generalized linear model predicting continuous olfactory sensation quality was tested, containing the continuous sum scales (MCSDS, IO-Q, CEQ, JIS) along with binary variables of condition, gender, background using VR devices and being under caffeine influence. Manipulation condition [β(SE) = 0.814(0.310), t (21) = 2.62, p = 0.016] and IO-Q [β(SE) = 0.439(0.206), t (21) = 2.13, p = 0.045] had a significant positive effects on perception of olfactory quality. Gender and other continuous predictors were not observed to have effects, while having background in VR [β(SE) = −0.564(0.321), t (21) = −1.76, p = 0.094] and being under influence of caffeine [β(SE) = −0.808(0.395), t (21) = −2.04, p = 0.054] exhibited negative trending effects. In total, the model explained just over half of the variance (R² = 0.520). In comparison, the model with the same variables was tested predicting total VR experience quality. In this model, there were no significant effects due manipulation condition or other variables, apart from MCSDS [β(SE) = 0.443(0.183), t (21) = 2.41, p = 0.025, model R² = 0.459], and JIS [β(SE) = 0.410(0.183), t (21) = 2.25, p = 0.036].

In sum, there was a significant difference in reported olfactory sensation between the manipulation and control conditions. Participants in the manipulation group reported significantly higher olfactory realism in their VR experience compared to the control group, while no group differences were observed in overall VR experience quality or in visual, auditory, or bodily sensation quality. A generalized linear model controlling for potential confounders (gender, VR experience background, caffeine influence, and continuous sum scales) identified significant positive effects of the manipulation condition and IO-Q on perceived olfactory quality. Gender and other predictors showed no effects, but VR experience background and caffeine influence indicated negative trending effects. The model explained just over half of the variance in olfactory quality. In contrast, a similar model predicting overall VR experience quality found no significant effects for the manipulation condition or most other variables, except for positive effects due to MCSDS and JIS scores.

As for relationships between dependent variables (Fig. 4), VR image quality and VR audio were strongly positively correlated with the overall experience (r = 0.76 and r = 0.66, respectively), indicating that participants who rated these sensory aspects highly also tended to report a better overall VR experience. This suggests that visual and auditory realism are critical factors in shaping users’ general impressions of the VR environment. In contrast, VR scent and VR body engagement showed weaker correlations with both overall experience and each other, suggesting that these elements are perceived more independently. VR scent had moderate correlations with VR audio (r = 0.37) and overall experience (r = 0.23), indicating potential enhancement for the experience to some extent, but not as central as visual and auditory elements. VR body engagement had near-zero correlations with other components, which implies that physical engagement (e.g., body tracking) is seen as a separate dimension and does not strongly influence participants’ perception of VR quality.

Fig. 4.

Pearson correlations between VR experience’s realism assessments, used as dependent variables.

Discussion

This study investigated the role of predictive processing (PP) in shaping olfactory perception within a virtual reality (VR) setting by manipulating participants’ short term prior expectations of olfactory cues. The results revealed a significant difference between the manipulation and control groups regarding reported olfactory sensations, with participants in the manipulation group rating olfactory realism significantly higher than those in the control group. Those in the manipulation group described detailed scents related to the VR experience in the 360° amazonian jungle, after implemented with the believe that the experiment condition had actually modelled olfaction. These findings provide empirical support for the PP framework in the olfactory modality, demonstrating that short-term expectation manipulations can influence sensory experiences. The strong effect of manipulation condition in predicting olfactory quality suggests that PP’s top-down processes can be effectively primed through external cues, thus enhancing the perceived realism of sensory input.

Our findings align with previous research on predictive processing. However, this study extends the PP model into the underexplored methodological domain of olfactory perception in VR, while also supporting earlier hypotheses that olfactory experiences may be shaped by anticipatory mechanisms^29,42. The observed effect of prior expectation manipulation is consistent with the concept of “controlled hallucination”⁶, wherein sensory experiences are guided by the brain’s predictive models. From a neuroscientific perspective, our findings also align well with prior studies that have shown the brain’s tendency to form pre-stimulus templates in sensory perception, suggesting that similar cognitive mechanisms may operate within the human olfactory system.

Our choice of using VR as an ecologically valid method for testing PP hypotheses highlights the potential of immersive environments to examine sensory-cognitive integration processes in naturalistic settings. VR allowed the dynamic presentation of stimuli in a controlled yet lifelike environment, enabling a realistic examination of how expectations shape sensory experiences. Our findings stipulate that while visual and auditory realism are important drivers of the overall VR experience, olfactory cues add distinct layers to the immersive experience. These results suggest that VR could serve as a powerful tool for investigating multisensory integration and PP in various other sensory modalities, potentially advancing our understanding of how different senses contribute to the holistic perception in virtual environments.

These results have implications for potential therapeutic help to individuals with psychiatric and neurological disorders associated with olfactory hallucinations. By understanding how hallucinations can be induced through VR and short-term expectation manipulation, researchers may gain insights into the mechanisms underlying persistent sensory hallucinations in conditions like schizophrenia. Given the recent reports of smell loss (hyposmia and anosmia) related to COVID-19, VR settings that manipulate expectations to stimulate olfactory experiences could potentially aid in rehabilitation or sensory retraining for those with temporal smell impairments. By helping patients’ brains expect and thus potentially “experience” lost smells again, there is a potential theoretical path toward gradual restoration or compensation for sensory loss. Since multisensory hallucinations are also common in so-far untreatable disorders like Lewy body dementia and Parkinson’s disease, VR could be also used to safely explore these conditions. If expectations can induce olfactory hallucinations, this approach might help patients manage distress associated with multisensory hallucinations, as these are often more troubling and difficult to differentiate from reality.

Limitations and future directions

Limitations to this study should be considered when interpreting the findings. First, the relatively small sample size (N = 30) limits the immediate generalizability of the results. Although statistical power was deemed sufficient for addressing the preregistered main hypothesis, future studies will provide more robust evidence for the observed effects and their relation to further background variables and extend the reliability of findings to general population. Second, the study’s reliance on self-reported measures of sensory realism may introduce subjective biases. The design did not allow for differentiation between experiencing hallucinations and merely reporting them. While the effects persisted together with the included confounders, such as socially desirable answering style, future research should incorporate physiological, such as neural or olfactory bulb activity, or behavioral measures to corroborate self-reported perceptions. In supplement, a design contrasting induced hallucinations with actual olfactory perception conditions could be used to determine the magnitude of the effects. A limitation of this study is also the lack of blinding during the coding of responses to open-ended survey questions. The experiments who applied the coding scheme were aware of whether a participant belonged to the manipulation or control condition. This awareness introduces a risk of experimenter bias, where coders might be biased toward encoding answers in a manner consistent with the hypothesized effect of the manipulation. Future research should employ independent coders who are fully blind to the experimental condition to mitigate this potential threat to the internal validity of the findings. Also, future studies should take into account that the actual olfactory abilities of the participants were not controlled for, which leads to the lack of a baseline. However, the robust finding that the control group lacked any hallucinated olfactory sensations weakens the necessity of a separate baseline measurement in the context of this study. Finally, while the VR environment provided an ecologically valid setting, the use of a single olfactory manipulation might limit the scope of inference. Future research could explore a wider range of manipulated expectations and sensory modalities to examine the generalizability of PP effects across different sensory and cultural domains.

Conclusion

This study contributes to the growing body of research on predictive processing (PP) by demonstrating that short-term manipulations of prior expectations can significantly shape olfactory experiences in a virtual reality setting. The study found a significant association between the manipulation and control conditions. Participants in the manipulation group reported higher olfactory realism in VR, while no group differences emerged in total VR experience quality or in visual, auditory, or bodily sensation quality. These findings underscore the importance of top-down cognitive processes in sensory perception and highlight the potential of VR as a tool for exploring PP across different sensory modalities. Findings of this research further elucidate the role of predictive processing in olfactory perception and expand the application of VR in cognition- and neuroscience research.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

A.B had the initial idea for the experiment, designed the experiment, collected the data, run the statistical analyses and participated in writing the article. A.B also did 95% of all the revision work. P.H partake in designing the experiment, running the statistical analyses and participated in writing the methods, discussion section of the article and making some revisions.

Declarations

Competing interests

The authors declare no competing interests.