Visual Placebo Effects in Expectancy Modulation

Updated 5 July 2026

Visual placebo effect is defined as expectancy-driven modulation via visual cues that alter subjective judgments without changes in physical input.
Empirical studies reveal effects across interactive displays, adaptive interfaces, and pain modulation, highlighting dissociations between subjective evaluation and objective performance.
Findings underscore that conceptual framing—not mere perceptual similarity—drives expectancy transfer in applications ranging from HCI to neuroscience.

The visual placebo effect, as suggested by recent work across pain neuroscience, human–computer interaction, human–AI interaction, and related methodological literatures, denotes placebo-like modulation produced by visually presented cues, visually framed narratives, or interface states that alter expectation, judgment, or behavior without a corresponding change in the underlying causal mechanism. In some settings the effect is purely subjective, such as improved self-evaluation under a narrated high-refresh-rate display with unchanged reaction time; in others it extends to behavior or pain, such as sham-AI-associated changes in perceptual decision parameters or category-based visual cues that modulate pain despite identical heat stimulation (Bosch et al., 2024, Kloft et al., 2023, Guindon et al., 17 Mar 2026).

1. Conceptual scope and defining characteristics

A visual placebo effect is not simply any benefit associated with looking at an image or display. The common structure is expectancy-driven modulation in which the operative “treatment” is a visual cue, a visualized system state, or a narrative attached to a visual system, while the nominal target variable is held constant or dissociated from the manipulated belief. The core inferential contrast is therefore between what is seen or believed and what is actually delivered.

In interactive-system studies, this structure is explicit. Participants are told that a display runs at a particular refresh rate, or that an adaptive AI is active, while the actual system behavior is either unchanged or absent; expectations and post-interaction judgments shift nevertheless (Bosch et al., 2024, Kosch et al., 2022). In pain modulation, the same logic appears in a more neurobiological form: novel visual cues that were never paired with heat later altered pain reports because they belonged to the same conceptual category as previously learned high- or low-pain cues, even though all generalization trials used the same medium-intensity heat (Guindon et al., 17 Mar 2026). This shows that the relevant visual signal can be conceptually mediated rather than based on low-level perceptual similarity.

A recurrent feature across these studies is dissociation between subjective and objective outcomes. In the refresh-rate study, subjective expectation and perceived performance followed the stated refresh rate, whereas objective reaction times did not differ significantly. In the adaptive-interface study on word puzzles, expectancy differences were robust whereas objective task performance showed no reliable main effect. By contrast, in the sham-AI letter-discrimination paradigm, belief in AI support was associated with slightly faster and more accurate performance, indicating that expectancy effects can sometimes propagate into measurable behavior (Bosch et al., 2024, Kosch et al., 2022, Kloft et al., 2023).

2. Interactive displays, adaptive interfaces, and sham AI

The clearest HCI demonstration of a visual placebo effect is the phantom-refresh-rate study. It used a full factorial within-subjects design with 18 participants and four conditions crossing true refresh rate (60 Hz or 240 Hz) with stated refresh rate (60 Hz or 240 Hz). The critical manipulation was that the stated refresh rate was sometimes false, and before each round participants were primed with a narrative about the rate they were supposedly using. The hardware was a Lenovo Y25-30 display verified at 60 Hz or 240 Hz, with a Razer Viper Ultimate mouse at 1000 Hz polling rate and 200 dpi. Expectations and post-task self-evaluations tracked the narrative rather than the true display state: for example, after using true 60 Hz, participants expected better performance when told 240 Hz than when told 60 Hz ( $Z = 32.5$ , $p < .001$ , $r = 0.88$ ), and judged their performance better under the same narrative inversion ( $Z = 42.50$ , $p < .001$ , $r = 0.86$ ). Objective reaction time did not change, with $F(4,157) = 0.82$ , $p = .51$ (Bosch et al., 2024).

Related findings were reported for adaptive-AI interfaces in word-puzzle tasks. In one study, all participants experienced fixed puzzle difficulty and no actual AI support, but were assigned to no-adaptivity, error-based adaptation, or physiology-based adaptation narratives. Experiment I recruited 369 participants and analyzed 255; Experiment II recruited 100 and analyzed 75. In the within-subject experiment, participants expected to solve more puzzles under the physiology-based adaptation condition ( $M = 24.39, SD = 8.93$ ) than under no adaptivity ( $M = 21.05, SD = 8.90$ ), with $p < .001$ 0, $p < .001$ 1, $p < .001$ 2, yet actual performance did not differ significantly, $p < .001$ 3, $p < .001$ 4, $p < .001$ 5. Pre-task expectancy correlated with correct responses overall ( $p < .001$ 6), and more strongly in the physiology-based condition ( $p < .001$ 7) (Kosch et al., 2022).

A more behaviorally consequential sham-AI effect was observed in a 2-alternative forced-choice letter-discrimination task. Participants were told that an AI called ADAPTIMIND™ was active or inactive, and that it would either improve or decrease performance depending on a positive or negative verbal description; in reality, no AI was present in either condition. Expectations remained positive irrespective of description, and the sham-AI-active condition produced slightly faster responses, with a system-status effect of $p < .001$ 8 ms and 95% HDI $p < .001$ 9, and slightly higher accuracy, 90.07% versus 89.35%. Hierarchical drift-diffusion modeling attributed this advantage primarily to a higher drift rate $r = 0.88$ 0, with smaller changes in boundary separation $r = 0.88$ 1 and non-decision time $r = 0.88$ 2, suggesting that expectancy altered information accumulation rather than merely self-report (Kloft et al., 2023).

Taken together, these studies establish a domain-general pattern in visually mediated interfaces: system descriptions, displayed states, and AI narratives can operate as placebo-like interventions. The empirical consequence, however, is not uniform. Some paradigms yield only subjective bias; others yield modest but measurable behavioral change.

3. Conceptual visual cues and pain modulation

In pain neuroscience, the visual placebo effect has been studied as conceptual generalization of expectancy-based pain modulation. Thirty-six participants first learned that two visual conditioned stimuli from different conceptual categories predicted different pain levels: one cue, $r = 0.88$ 3, predicted high pain and the other, $r = 0.88$ 4, predicted low pain. To isolate cue-based modulation from stimulus intensity, both cue types were followed on half of trials by the same medium-intensity heat, 49°C. Before each heat stimulus, participants rated expected pain, and after heat they rated experienced pain. The slope relating cue type to expected pain provided an individual explicit learning score (Guindon et al., 17 Mar 2026).

Generalization was then tested with novel stimuli that had never been paired with pain but belonged to the same conceptual category as either $r = 0.88$ 5 or $r = 0.88$ 6. These generalization stimuli were presented in three visual modalities—images or photographs, drawings or cartoons, and written words—so the manipulation was not reducible to shared surface form. During this phase, all stimuli were followed by the same medium-intensity heat, and pain was rated afterward; expectation ratings were intentionally omitted during generalization to avoid biasing pain reports (Guindon et al., 17 Mar 2026).

The key representational result was that participants organized the cues by conceptual category rather than by low-level visual similarity or exact exemplar identity. In a similarity-rating task, the category-based model fit best, with $r = 0.88$ 7, outperforming an exemplar model with $r = 0.88$ 8 and a ResNet-50 visual-similarity model with $r = 0.88$ 9. This directly supports an account in which visually presented meaning, rather than simple perceptual resemblance, drives the generalization effect (Guindon et al., 17 Mar 2026).

Behaviorally, the learning phase worked as intended: on identical medium-heat trials, $Z = 42.50$ 0 produced higher pain ratings than $Z = 42.50$ 1 and also higher expected pain. In the generalization phase there was no significant average group-level difference in pain ratings between $Z = 42.50$ 2 and $Z = 42.50$ 3. The critical effect emerged only after conditioning on explicit learning. Participants with larger learned differences in expected pain for $Z = 42.50$ 4 versus $Z = 42.50$ 5 showed larger generalized pain responses to the conceptually related novel cues, with a significant $Z = 42.50$ 6 by expectancy interaction, $Z = 42.50$ 7, $Z = 42.50$ 8 (Guindon et al., 17 Mar 2026).

Neuroimaging localized the mechanism with multilevel mediation analysis using trial-by-trial pain-evoked beta maps. The strongest path- $Z = 42.50$ 9 effects were observed in bilateral amygdala, left mid-posterior insula, subgenual ACC, premotor or motor cortex, and left posterior hippocampus. Path $p < .001$ 0 recovered the canonical pain network, including anterior and posterior insula, anterior midcingulate cortex, thalamus, somatosensory cortices, motor areas, brainstem, and cerebellum. The principal path- $p < .001$ 1 mediators formed a broader network including the left hippocampus, ventral striatum or nucleus accumbens and caudate, thalamus, temporopolar cortex, retrosplenial cortex, middle temporal gyri, posterior parietal cortex, precuneus, and other frontal and cerebellar regions (Guindon et al., 17 Mar 2026).

The hippocampus was the key mediator. Pain-evoked hippocampal activity was greater for $p < .001$ 2 than for $p < .001$ 3, statistically explained the increase in pain ratings, and was stronger in participants with greater explicit learning of the original cue–pain associations. By contrast, the amygdala, insula, and motor cortex responded more strongly to $p < .001$ 4 but did not mediate the effect on pain ratings. The authors therefore distinguish a lower-level aversive or threat network from a higher-order hippocampal–DMN–striatal system that converts conceptual knowledge into expectancy-based modulation of pain (Guindon et al., 17 Mar 2026).

A related but terminologically distinct line of work studies when a recreated place is accepted as “the same” despite incomplete or inaccurate visual reconstruction. This literature does not use the phrase visual placebo effect, but it is relevant because it shows that perceived equivalence can arise from memory, meaning, and activity-linked cues rather than high-fidelity visual correspondence alone (Aoyagi et al., 2023).

In a classroom re-creation task in virtual reality, 15 Japanese university students recreated two known classrooms by placing virtual objects in empty 3DCG rooms. The study defined place sameness as the degree of similarity between an original place and a virtual re-creation that represents the original place, and operationalized it with a 7-item place sameness index with internal consistency $p < .001$ 5. The environments were manipulated through functional object categories: display devices, student desks, and others (Aoyagi et al., 2023).

The principal findings indicate that place sameness was linked to the reproducibility of activity-related objects and not simply to geometric accuracy. Moderate negative correlations were found between place sameness and the memory index for Room A student desks ( $p < .001$ 6) and Room B display devices ( $p < .001$ 7), implying that more accurate reproduction of the number of such objects increased perceived sameness. At the same time, some average-distance correlations were positive, including Room A display devices ( $p < .001$ 8), Room B student desks ( $p < .001$ 9), and Room B others ( $r = 0.86$ 0), indicating that greater positional inaccuracy sometimes co-occurred with higher place sameness (Aoyagi et al., 2023).

The study also reported partial support for an uncanny-valley-like effect of place, paraphrased as the observation that a more objectively inaccurate re-creation can sometimes appear subjectively more accurate than a more objectively accurate one. This pattern was not consistent across rooms, but it reinforces the broader point that visual acceptance is not monotonic in fidelity (Aoyagi et al., 2023).

A plausible implication is that some visual placebo-like phenomena rely on functional or semantic cues that allow observers to infer the intended identity of a scene, system, or signal. This suggests continuity between place sameness in virtual environments and conceptual generalization in pain: in both cases, visually presented structure is interpreted through a higher-order representational frame rather than through low-level matching alone.

5. Control logic, placebo diagnostics, and causal identification

Because expectancy effects can be elicited by narratives, labels, and arbitrary cues, the study of visual placebo effects depends heavily on control design. In human–AI and HCI studies, the strongest recommendation is to measure expectations before and after interaction and to include placebo or sham conditions rather than relying on subjective ratings alone (Bosch et al., 2024, Kosch et al., 2022, Kloft et al., 2023).

A closely related methodological warning appears in work on socio-demographic prompting for LLMs. That paper uses “placebo effect” to denote “random perturbations of model responses due to arbitrary tokens in the prompt,” and introduces culturally non-sensitive proxy cues such as disease, hobby, programming language, planet, and house number as controls against overinterpreting cue-induced variation as cultural conditioning. Across Llama 3, Mistral v0.2, GPT-3.5 Turbo, and GPT-4, all models except GPT-4 showed significant response variation for both culturally sensitive and non-sensitive cues, even on neutral datasets such as MMLU and ETHICS. On neutral datasets, the ideal invariance condition was stated as $r = 0.86$ 1, but was often violated. Although this is not a visual placebo study, it formalizes the general need for placebo controls whenever arbitrary cues can perturb downstream behavior (Mukherjee et al., 2024).

Clinical and causal-inference literatures provide further control frameworks, though not visual mechanisms as such. Instrumental-variable methods have been proposed to separate placebo and treatment effects in blinded and unblinded trials by using randomized treatment assignment $r = 0.86$ 2 and randomized encouragement $r = 0.86$ 3, with placebo and treatment estimands $r = 0.86$ 4 and $r = 0.86$ 5 after residualization (Neto, 2016). Time-varying placebo effects in long Phase 3 trials have been modeled semiparametrically via $r = 0.86$ 6, $r = 0.86$ 7, with B-spline estimation in the SWSR procedure (Zhan et al., 11 May 2025). Observational studies have used a placebo sample defined by the assumption $r = 0.86$ 8 almost surely for $r = 0.86$ 9, allowing bias diagnosis and correction through outcome regression, IPW, and doubly robust estimation (Ye et al., 2022). In discontinuity settings, placebo zones have been used to select bandwidths, polynomials, and even model class by minimizing RMSE where the true effect is known to be zero (Kettlewell et al., 2022). With very few treated clusters, Fisher-style placebo randomization by reassigning treatment labels across clusters yields nearly exact inference and can operate with as few as three treated and three untreated clusters (Hagemann, 2018). By contrast, analysis of the sequential parallel comparison design shows that conventional SPCD estimators do not generally target meaningful treatment effects because placebo-response classification is imperfect and latent (Stockton et al., 24 Nov 2025).

These frameworks do not themselves define a visual placebo effect, but they provide the inferential architecture needed to distinguish expectancy-driven change from active-system change.

6. Interpretation, misconceptions, and open questions

One common misconception is that a visual placebo effect requires simple visual resemblance. The pain-generalization study shows the opposite: the relevant transfer occurred across images, drawings, and words, and the category-based model outperformed both an exemplar model and a deep-network visual-similarity model. This indicates that abstract category knowledge can dominate low-level similarity in shaping visually cued expectancy (Guindon et al., 17 Mar 2026).

A second misconception is that such effects are necessarily restricted to self-report. The refresh-rate study and the word-puzzle adaptive-interface study indeed found pronounced effects on expectation and perceived performance without reliable changes in objective performance. However, the sham-AI letter-discrimination study found slightly faster responses, slightly better accuracy, and a higher drift rate in the putative AI-active condition. The pain study likewise involved changes in experienced pain under identical heat stimulation for participants with stronger explicit learning. Visual placebo effects are therefore not confined to post hoc evaluation; under some conditions they alter online decision or sensory processing (Bosch et al., 2024, Kosch et al., 2022, Kloft et al., 2023, Guindon et al., 17 Mar 2026).

A third misconception is that negative framing is sufficient to neutralize expectancy. In the sham-AI study, negative descriptions of AI did not eliminate generally positive expectations, and a replication study again found that overall performance expectations and task-speed expectations stayed optimistic even when participants clearly understood the negative framing. This suggests that some expectancy priors associated with AI are unusually robust (Kloft et al., 2023).

The literature also indicates that visual placebo effects are heterogeneous rather than universal. In the pain-generalization experiment there was no significant average group-level difference between $F(4,157) = 0.82$ 0 and $F(4,157) = 0.82$ 1; the effect depended on explicit learning. In the adaptive-interface literature, the transparent error-based adaptation narrative and the more opaque physiology-based narrative did not produce identical expectancy profiles. In virtual-place research, partial uncanny-valley-like responses appeared in one room but not another. The governing moderators therefore include explicit contingency learning, task structure, prior beliefs about AI, and the functional meaning of visual cues (Kosch et al., 2022, Aoyagi et al., 2023, Guindon et al., 17 Mar 2026).

A plausible synthesis is that the visual placebo effect is best understood as an expectancy-transfer phenomenon operating over representations rather than mere optics. In interactive systems, the transfer runs from narrated capability to perceived or measured performance; in pain, from learned cue meaning to altered nociceptive experience; in virtual places, from activity-related cues to acceptance of environmental identity. The strongest open problem is methodological: to specify when visually induced expectations merely bias evaluation and when they recruit mechanisms capable of altering perception, decision, or pain itself.