The Benevolent Demon: Why Seeing Should Not be Believing

Adapted from my University of Vienna seminar paper, “Seeing should not be believing: A Reliabilist Analysis of Perceptual Fallibility”¹ — you can read the full paper here. This is a philosophy essay rather than a data post, but the everyday assumption it examines — that seeing is believing — is among the more widely trusted commitments we hold.

The assumption on trial

Almost everyone treats their own perception as a transparent window onto the world. Philosophers have a name for the polished version of this - direct realism: we perceive the world as it is, immediately, without an intermediary². Even those who disagree tend to keep the tenet that seeing is believing and that the experienced world is, on the whole, reliable³. Some grant that sentiment so much explanatory weight that it becomes an argument against indirect realism itself: if a theory of perception cannot accommodate our seeing what we believe, that is taken to count against the theory rather than against the sentiment.

I find that assumption suspicious. This piece argues that we should not grant the reliability of our perceptual faculties as a given - and that, on close inspection, they are systematically untrustworthy as a source of truth, even while being good at keeping us alive.

The central image is what I call the Benevolent Demon. Not Descartes’ supernatural deceiver, but the brain’s own biological machinery: a system that actively constructs our experience, evolved for survival rather than truth.⁴ The claim comes in two parts:

1. Perfect epistemic reliability - consistently accurate representation - is expensive, and therefore not always favoured by evolution.
2. Our perceptual systems are in fact not built to deliver an accurate model of the environment. They are tuned for adaptive utility: survival and action-guidance over veridicality.

The rest of the essay develops this case in two claims. First, that perception is unreliable by a strict, scientific standard. Second, that perception is constructed rather than transparent. I then draw the conclusion, and set out what follows from it for our epistemic practice.

Part I — The reliability of perception

What “reliable” should actually mean

A true belief becomes knowledge only when it is justified, and one of the more useful accounts of justification holds that it should come from a reliable process⁵. Reliabilism is attractive because it is externalist and practical: you can ask how you came to believe something, and whether that route tends to produce truth. If it does, believe; if not, withhold. Strictly speaking, externalist reliabilism does not require the believer to be aware of a process’s reliability at all - but that stricter reading is of little use to an individual who is trying to reason well, so throughout this piece I treat reliabilism as a practical tool for the epistemically responsible person: a practical extension of the theory rather than a contradiction of it.

But the framework leaves the crucial threshold undefined. Is a process that is 51% accurate “reliable”? Does it need 99%? Which errors are we counting? To make this precise, I borrow the standards of the most rigorous truth-filtering enterprise we have developed: the scientific method. Science does not merely accumulate truths; it works to filter out error.

Here the Popperian asymmetry is decisive: falsehood is usually easier to establish than truth - by contradiction, failed prediction, or counterexample⁶. So the mark of a reliable process is not infallibility but a disciplined refusal to affirm falsehoods. In statistical language: a low Type I error rate.

Two errors, unequal in cost

Statistics distinguishes two ways to be wrong⁷:

• Type I error ($\alpha$) - a false positive. Rejecting a true null hypothesis. Epistemically: believing $p$ when $p$ is false. Hallucinating an object; seeing a pattern in noise.
• Type II error ($\beta$) - a false negative. Failing to reject a false null. Epistemically: failing to believe $p$ when $p$ is true. Missing a real object or a real tumour.

Science conventionally holds $\alpha$ at 0.05 or below - a process counts as reliable only if it fabricates less than 5% of the time. That convention encodes an epistemic value judgement: asserting a falsehood is catastrophic, whereas remaining ignorant is merely corrigible. A system with high sensitivity but a high false-positive rate pollutes the canon - it becomes impossible to trust any of its positive claims, because the true ones are indistinguishable from the fabricated ones.

Error type	Scientific definition	Epistemic equivalent	Consequence for reliability
Type I (false positive)	Rejecting a true $H_0$	“Hallucinating”; believing a falsehood	Catastrophic — destroys trust
Type II (false negative)	Retaining a false $H_0$	Missing a fact; ignorance	Manageable — result is ignorance

The brain is tuned for the wrong error

Apply that standard to perception and a reliable faculty would be a conservative filter: refusing to affirm an object, pattern, or event without compelling evidence. It seems evolution did not select for that.

In the ancestral environment the cost of missing a predator (Type II) was death; the cost of imagining one (Type I) was merely some expended energy. Natural selection therefore favoured high sensitivity over high reliability. Several well-documented phenomena point in this direction:

• Face pareidolia - we are tuned to faces closely enough that we sometimes see them where none exist, in clouds, sockets, or toast⁸.
• Hyperactive agency detection - we infer an agent behind ambiguous events more readily than the evidence warrants⁹.
• Sensory Processing Sensitivity (SPS) - a heritable trait of heightened sensitivity to subtle input, found across more than 100 species¹⁰ and present in an estimated 20-30% of humans¹¹. A substantial share of the population is thus constitutionally predisposed toward false positives.

This is the root of the Benevolent Demon problem: the brain is optimised for fitness, but epistemology demands truth.F

Cognitive penetration: a circular mechanism

Perception is not a neutral, bottom-up registration of the world. Through cognitive penetration, higher-level states - beliefs, desires, moods, expertise - influence the phenomenal character of experience itself¹². If our beliefs shape what we perceive, and what we perceive in turn reinforces those beliefs, the result is a feedback loop.

Susanna Siegel allows that penetration can be epistemically good (the expert radiologist) or bad (biased social perception). I want to press the point further. Perception has been described as a closed-loop system in which top-down predictions tune the sensitivity of sensory cortex toward expected stimuli¹³. Lowering the detection threshold for an expected pattern reduces misses (Type II) - but since $\alpha$ and $\beta$ trade off against each other, it necessarily raises false positives (Type I).

The chart below makes the trade concrete. A detector sets a criterion - how much evidence it demands before it says “yes”. Slide the criterion leftward (demand less, because you expect the signal) and the hit rate climbs - but the false-alarm rate climbs right alongside it.

Axes: the horizontal axis is the decision criterion (drawn so that left = more eager to say yes); the vertical axis is the probability of responding “yes”. Look for: the green hit-rate curve and the red false-alarm curve rise together as the criterion drops. There is no setting that raises one without raising the other. Takeaway: an expectation that increases sensitivity to a signal purchases genuine detections and fabrications with the same mechanism.

This is why I read cognitive penetration as structurally circular, and as functionally analogous to p-hacking. A p-hacker runs tests until something crosses significance; a cognitively penetrated brain adjusts the gain on specific neural populations until the input matches the prior. Every proposed “good case” is a gain in sensitivity, not in reliability:

A detector that finds 100% of real tumours but labels 50% of healthy tissue as tumorous has maximal sensitivity but terrible reliability.

Imagine a doctor who flags a tumour on every scan. By definition they catch every real tumour - and fabricate one on every healthy patient. Nobody would call that reliable diagnosis. From a reliabilist standpoint, a process that blends truth with systematic fabrication cannot justify belief, because the true outputs are indistinguishable from the false ones without external verification. That indistinguishability is the whole problem.

What does the evidence say? It is mixed, and worth reporting as such:

• When top-down and bottom-up signals can fall into discrepancy, resulting perception that is distorted in predictable, clinically observable ways¹⁴
• Prior knowledge and affective state can determine or at least affect perceptual prediction¹⁵.
• Value and expectation bias perceived size when input is weak or absent¹⁶
• Language and labels increase sensitivity to certain features and push perception toward more discrete categories¹⁷¹⁸¹⁹.
• Raising the prior probability of a target increases false positives without improving conventional sensitivity for true detections²⁰
• Across five experiments practice reliably lowered thresholds while false-alarm rates rose several-fold²¹.
• There is dissent as well: one careful study concluded that expectation “did not affect perception”, with false percepts arising instead from spontaneous feedforward-like activity in early visual cortex²².

Without a meta-analysis I will not draw a definitive conclusion, but the balance of evidence is more consistent with the view that cognitive penetration is prevalent and increases sensitivity at the cost of reliability than with the view that it increases reliability.

Part II — The constructedness of perception

The first part asked whether perception yields justified belief. The second asks how the experience is generated at all. Two features - non-linear deformation and patchy sampling - show that consciousness serves up not raw input but a constructed model.

Perception is deformed, not mirrored

Psychophysics - the study of the map between physical stimuli and felt sensation - finds that map is non-linear and warped. If we had direct access to the world, felt intensity would track physical intensity linearly. It does not.

The Weber-Fechner law captures the first approximation: equal subjective steps correspond to multiplicative changes in the stimulus - a roughly logarithmic relationship²³. Concretely, you perceive 7.5 kg as only about twice as heavy as 2.5 kg, and 1000 lux as only about three times as bright as 10 lux.

Axes: physical intensity on the horizontal axis, perceived intensity on the vertical. Look for: the curve rises steeply at first and then flattens - a hundredfold increase in stimulus at the right produces barely any perceptual change. Takeaway: the same physical difference feels enormous when the baseline is low and negligible when it is high.

S. S. Stevens generalised this into Stevens’ power law: perceived intensity is a power function of stimulus intensity, with an exponent that differs by sense²⁴. Below 1 the sense compresses (brightness, loudness); above 1 it expands (heaviness, electric shock). Refined estimates continue to appear²⁵²⁶, and the spread across modalities is dramatic:

Stimulus	Power-law exponent $n$	Condition
Brightness	0.33	5° target, dark-adapted eye
Loudness	0.60	Binaural
Smell (coffee)	0.55	Odour
Taste (sucrose)	1.3	—
Heaviness	1.45	Lifted weights
Electric shock	3.5	60 c.p.s. through the fingers

Axes: normalised physical intensity vs normalised perceived intensity, both on $[0,1]$. Look for: brightness (blue) bows above the linear line - it compresses a huge physical range into a small perceptual one - while electric shock (red) bows sharply below it, so a small increase in current feels like a huge jump in pain. Takeaway: there is no single “gain” the brain applies; each sense deforms reality on its own terms.

Temperature is an especially vivid case. We treat a physiological neutral of roughly 32 °C as “comfortable”, with sensitivity peaking near that point and saturating toward hot and cold extremes - an S-shaped curve²⁷.

Axes: physical temperature vs a felt comfort vote from cold ($-3$) to hot ($+3$). Look for: the curve is steep through the middle and flat at both ends - a two-degree change near neutral is very noticeable, the same change at 12 °C or 38 °C barely registers because you already feel “freezing” or “boiling”. Takeaway: the perceptual scale is squeezed at the extremes to keep the useful middle range legible.

The reasonable conclusion here is that perception is at best ordinally veridical - it ranks stimuli by intensity correctly - but not cardinally veridical, since it misrepresents the sizes of the differences between them. Most likely this is adaptive: compressing an enormous physical range into a manageable perceptual one lets us register a whisper and a jet engine with the same hardware. That is an impressive engineering solution, but it is not a truth-preserving one.

The patchiness-problem

The second marker of constructedness: sensing is not continuous. It happens in discrete patches that the brain sews together.

Vision is the clearest case. Only the fovea perceives with high acuity - a region of about 1-2° of visual angle, roughly 0.01-0.03% of the ~180° visual field²⁸. To construct a whole scene the eyes make rapid movements called saccades, several per second²⁹. Yet we do not experience these snapshots as discrete visual episodes; the brain integrates them into one stable percept through transsaccadic integration.

That integration is precisely what should give us pause. The unified percept is assembled from patches captured at different times and places - we have no access to the raw input, only to the integrated output. The resulting scene is temporally hybrid: it depicts elements that never co-occurred in one place at one instant. Moreover, the integration is achieved by optimal inference - the brain extrapolates and fills in missing information from prior expectation rather than simply concatenating the raw data²⁹.

The patchiness runs deeper than vision. As Zamfira and colleagues put it:

We live in a world rich in multisensory stimuli, and our brain needs to integrate and segregate crossmodal information in space and time to construct a coherent and smooth representation of the environment.³⁰

For a sound and a sight to fuse into one event, they must arrive within a narrow temporal binding window³¹³². Multisensory integration usually improves performance, but only when timing supports binding³². Tellingly, atypical binding windows show up in autism, dyslexia, and schizophrenia³²³³ - reinforcing that the smooth, unified world is a dynamic achievement, not a given.

The brain in the skull

Taken together, the two markers force a fairly strong conclusion. Sensory magnitude is systematically transformed before it reaches awareness, and the incoming stream is sparse, interrupted, and often asynchronous across senses - yet phenomenology presents us with a seamless, stable, temporally coherent scene. That coherence is not present in the input; it is added by neural construction.

The classic brain in a vat imagines a brain fed simulated signals. We can read our actual situation similarly: the brain is indeed cut off from the world, it never receives raw sensor data - but it steers a real body through a real environment. The machinery processes the input and forwards only a highly constructed (‘simulated’) rendering. We act on reality, yet we are at the mercy of an involuntary interface - functionally like a VR pipeline or humans in ‘the matrix’ movies, or the famous brain in a vat (with the additional benefit of being somehow related to a real world).

The skull is a dark and silent place, but we perceive a world that is bright and noisy.

In that qualified sense we are each a brain in the skull: not detached from reality, but epistemically confined to the brain’s constructed interface to it.

The Benevolent Demon conclusion

Four threads converge:

1. The brain was selected for sensitivity over accuracy - it is a tool of utility, not of truth.
2. Cognitive penetration lets prior belief manipulate perception in a self-confirming loop, raising sensitivity and, plausibly, Type I error.
3. Non-linear deformation removes any direct, faithful access to physical magnitude.
4. The sensing-in-patches problem shows the unified percept is an inferred, extrapolated construction.

What we treat as input to the mind is actually the output of a complex system we cannot inspect. We are, for epistemic purposes, brains in vats - fed constructed renderings rather than raw data. This is not skepticism about the external world; I happily assume it exists. It is skepticism about the transparency of the interface.

The demon here is real but neither supernatural nor malicious: it is our own biological machinery, which has evolved to construct a coherent and useful experience of the world rather than a truthful one. What looks like illusion from the standpoint of epistemology is often an efficient compression strategy from the standpoint of survival.

This is harmonious with the Predictive Processing framework, on which the brain is essentially a prediction machine and perception is its best guess, formed by combining incoming sensory evidence with prior expectations³⁴. That view implies that perceptual error is unavoidable: probabilistic inference under uncertainty will sometimes misclassify the world.

A natural assumption, shared by most philosophers, is that veridicality is what makes perception useful - the more accurately you see, the better you act. But the evidence reviewed above suggests that veridicality is at best a collateral benefit, not the source of adaptive success. What is genuinely useful is often non-veridical: fleeing when there might be a predator (a false positive that costs little), recoiling from what might be poisonous, finding calorie-dense food intensely pleasurable regardless of nutritional need, perceiving loved ones as more attractive than a neutral observer would, or compressing a vast range of physical intensities into a manageable scale. In each case the payoff comes not from representing reality accurately but from tilting perception toward the response that maximises fitness³⁵.

This is the heart of Donald Hoffman’s Interface Theory of Perception: perception is an adaptive interface, like a desktop whose icons help you act while hiding the transistors³⁶. Hoffman goes further with the Fitness-Beats-Truth theorem: natural selection does not favour veridical perception - truth-tracking strategies are routinely out-competed by fitness-tracking ones³⁶.

Conclusion — the paranoid non-skeptic

The architecture described here is remarkably sophisticated and, at the same time, epistemically treacherous. Perception is not a transparent window but a non-linear deformation; cognition is not a neutral judge but often a self-confirming, p-hacking engine that bends evidence toward prior expectation. In this sense the brain functions as a real and immanent demon - not metaphysical but biological, not malicious but structurally deceptive with respect to truth.

So “seeing is believing” is not merely naive; it is an epistemic hazard. If perceptual confidence is generated by the very machinery under suspicion, then first-person certainty cannot serve as its own credential.

The right response is not global skepticism. It is methodological paranoia: a disciplined scientific skepticism aimed at your own consciousness. Treat perception as an exploratory tool that feeds you suspicions, not a tribunal that hands down verdicts. Then run the suspicion through a truth-checking loop:

• Contradiction — does the claim clash with anything else you know?
• Coherence — does it fit your background knowledge?
• Convergence — do independent channels (a photo, a measurement, another witness) agree?

Take the deliberately absurd test case: you perceive a flying, rainbow-pooping unicorn. The disciplined response is not blind trust and not blank denial - it is hypothesis management. Maybe there is a unicorn: test distance, lighting, motion, material detail; ask whether it is a prop, a projection, a painted horse. If the appearance survives repeated checks, shift the hypothesis space to background conditions - was your drink spiked, are there pharmacological effects, is this acute psychiatric disturbance needing care? At each step belief updates by coherence and convergence, not by the raw force of the image.

We are confined to a forged world of adaptive interfaces and controlled hallucinations. That is not cause for resignation. By understanding the architecture of deception - non-linearity, p-hacking circularity, predictive construction - we can navigate experience responsibly. The imperative is simple to state and hard to live:

Do not simply believe what you see. Test it, verify it, and triangulate it - until belief is supported by a web of justification that exceeds the limits of your own biology.

---

Adapted from “Seeing should not be believing: A Reliabilist Analysis of Perceptual Fallibility” (KU The Epistemology of Perception and Memory, University of Vienna, February 2026) — the full paper is available here. Every source below is linked to its full PDF where licensing allows. Corrections, counter-evidence, and different angles are all welcome.

1. Duty-Knez, M. (2026). Seeing should not be believing: A Reliabilist Analysis of Perceptual Fallibility. Seminar paper, KU The Epistemology of Perception and Memory, University of Vienna. 📄 Full PDF ↩

2. O’Brien, D. Objects of Perception. Internet Encyclopedia of Philosophy. iep.utm.edu/perc-obj ↩

3. Burge, T. (2003). Perceptual Entitlement. Philosophy and Phenomenological Research, 67(3), 503-548. doi:10.1111/j.1933-1592.2003.tb00307.x ↩

4. Mercier, H., & Sperber, D. (2017). The Enigma of Reason. Harvard University Press. — the argument that reason evolved as a social, not epistemic, tool. ↩

5. Goldman, A. (2008). Reliabilism. Stanford Encyclopedia of Philosophy. ↩

6. Popper, K. R. (1934/1959). The Logic of Scientific Discovery. Routledge. — the asymmetry between verification and falsification. ↩

7. Smith, C. J. (2012). Type I and Type II errors: what are they and why do they matter? Phlebology, 27(4), 199-200. 📄 PDF · doi:10.1258/phleb.2012.012j04 ↩

8. Stuart, G., Saurels, B. W., Robinson, A. K., & Taubert, J. (2025). One object with two identities: the rapid detection of face pareidolia. J. Exp. Psychol. HPP, 51(6), 710-720. doi:10.1037/xhp0001296 ↩

9. Lisdorf, A. (2007). What’s HIDD’n in the HADD? Journal of Cognition and Culture, 7(3-4), 341-353. 📄 PDF · doi:10.1163/156853707X208549 ↩

10. Greven, C. U., Trupp, M. D., Homberg, J. R., & Slagter, H. A. (2025). Sensory processing sensitivity: theory, evidence, and directions. Trends in Cognitive Sciences. 📄 PDF · doi:10.1016/j.tics.2025.10.007 ↩

11. Rajić, I. (2024). Sensory Processing Sensitivity. Psihološka istraživanja. 📄 PDF · doi:10.5937/psistra0-48157 ↩

12. Siegel, S. (2012). Cognitive Penetrability and Perceptual Justification. Noûs, 46(2), 201-222. jstor.org/stable/41475338 ↩

13. Ahissar, E., & Assa, E. (2016). Perception as a closed-loop convergence process. eLife, 5, e12830. doi:10.7554/eLife.12830 ↩

14. O’Callaghan, C., Kveraga, K., Shine, J. M., Adams, R. B., & Bar, M. (2017). Predictions penetrate perception. Consciousness and Cognition, 47, 63-74. 📄 PDF · doi:10.1016/j.concog.2016.05.003 ↩

15. Lupyan, G. (2015). Cognitive Penetrability of Perception in the Age of Prediction. Review of Philosophy and Psychology, 6(4), 547-569. 📄 PDF · doi:10.1007/s13164-015-0253-4 ↩

16. Liao, J., Yang, Y., Han, Z., & Mo, L. (2024). The Critical Trigger for Cognitive Penetration. Behavioral Sciences, 14(8), 632. 📄 PDF · doi:10.3390/bs14080632 ↩

17. Lupyan, G., & Ward, E. J. (2013). Language can boost otherwise unseen objects into visual awareness. PNAS, 110, 14196-14201. 📄 PDF · doi:10.1073/pnas.1303312110 ↩

18. Lupyan, G., & Spivey, M. J. (2010). Making the Invisible Visible: Verbal but Not Visual Cues Enhance Visual Detection. PLoS ONE, 5. 📄 PDF · doi:10.1371/journal.pone.0011452 ↩

19. Lupyan, G., Abdel Rahman, R., Boroditsky, L., & Clark, A. (2020). Effects of Language on Visual Perception. Trends in Cognitive Sciences, 24, 930-944. 📄 PDF · doi:10.1016/j.tics.2020.08.005 ↩

20. Wyart, V., Nobre, A. C., & Summerfield, C. (2012). Dissociable prior influences of signal probability and relevance on visual contrast sensitivity. PNAS, 109, 3593-3598. 📄 PDF · doi:10.1073/pnas.1120118109 ↩

21. Wenger, M., Copeland, A. M., Bittner, J., & Thomas, R. D. (2008). Evidence for criterion shifts in visual perceptual learning. Perception & Psychophysics, 70, 1248-1273. 📄 PDF · doi:10.3758/pp.70.7.1248 ↩

22. Haarsma, J., Deveci, N., Corbin, N., Callaghan, M., & Kok, P. (2023). Expectation Cues and False Percepts… The Journal of Neuroscience, 43, 7946-7957. doi:10.1523/jneurosci.0998-23.2023 — the dissenting finding. ↩

23. Maes, C. (2020). Statistical Mechanical Foundation of Weber-Fechner Laws. Journal of Statistical Physics, 182. 📄 PDF · arXiv:2009.00923 ↩

24. Atkinson, W. H. (1982). A general equation for sensory magnitude. Perception & Psychophysics, 31, 26-40. 📄 PDF · doi:10.3758/bf03206198 ↩

25. Zhou, J., Duong, L., & Simoncelli, E. P. (2024). A unified framework for perceived magnitude and discriminability of sensory stimuli. PNAS, 121. 📄 PDF · doi:10.1073/pnas.2312293121 ↩

26. Chen, J., Tian, S., Wang, X., Mao, Y., & Zhao, L. (2021). The Stevens law and the derivation of sensory perception. Journal of Future Foods. 📄 PDF · doi:10.1016/j.jfutfo.2021.09.004 — source of the exponent table. ↩

27. Sun, X., Wu, H., & Wu, Y. (2020). Investigation of the relationships among temperature, illuminance and sound level, physiological parameters and human perceptions. Building and Environment, 183, 107193. 📄 PDF · doi:10.1016/j.buildenv.2020.107193 ↩

28. Williams, L. J. (1982). Cognitive Load and the Functional Field of View. Human Factors, 24, 683-692. 📄 PDF · doi:10.1177/001872088202400605 ↩

29. Enever, P., et al. (2003). Optimal transsaccadic integration explains distorted spatial perception. Nature, 422, 76-80. 📄 PDF · doi:10.1038/nature01439 ↩ ↩²

30. Zamfira, D. A., Di Dona, G., Marsicano, G., Battista, M., Battaglini, L., & Ronconi, L. (2025). The rhythm of sensory input shapes audio-visual temporal processing. British Journal of Psychology, 117, 314-336. 📄 PDF · doi:10.1111/bjop.70029 ↩

31. Zhou, H.-y., Cheung, E. F. C., & Chan, R. C. K. (2020). Audiovisual temporal integration. Neuropsychologia, 140, 107396. 📄 PDF · doi:10.1016/j.neuropsychologia.2020.107396 ↩

32. Wallace, M. T., & Stevenson, R. A. (2014). The construct of the multisensory temporal binding window and its dysregulation in developmental disabilities. Neuropsychologia, 64, 105-123. 📄 PDF · doi:10.1016/j.neuropsychologia.2014.08.005 ↩ ↩² ↩³

33. Di Cosmo, G., et al. (2020). Body-environment integration: temporal processing of tactile and auditory inputs along the schizophrenia continuum. Journal of Psychiatric Research, 134, 208-214. 📄 PDF · doi:10.1016/j.jpsychires.2020.12.034 ↩

34. Hohwy, J. (2013). The Predictive Mind. Oxford University Press. doi:10.1093/acprof:oso/9780199682737.001.0001 ↩

35. Martin, J. M. (2021). Useful misrepresentation: perception as embodied proactive inference. Trends in Neurosciences, 44, 619-628. 📄 PDF · doi:10.1016/j.tins.2021.04.007 ↩

36. Hoffman, D. D., Singh, M., & Prakash, C. (2015). The Interface Theory of Perception. Psychonomic Bulletin & Review, 22(6), 1480-1506. 📄 PDF · doi:10.3758/s13423-015-0890-8 ↩ ↩²