Rosenholtz proposes that what has traditionally been attributed to “attention” may largely be explained by task demands and summary statistics in peripheral vision. I argue that this perspective not only helps reconcile longstanding inconsistencies in ensemble perception research but also offers a novel way to interpret phenomena like the attentional blink (AB). In turn, this reframing has also significant implications for the philosophy of action, particularly debates about whether attention is necessary for or constitutive of action.

Ensemble perception – the ability to extract statistical properties from groups of objects – has generated conflicting findings regarding its dependence on attention. On one side, experiments such as Li et al. (Reference Li, VanRullen, Koch and Perona2002) and Joo et al. (Reference Joo, Shin, Chong and Blake2009) conclude that ensemble perception operates automatically, while on the other side, Cohen, et al. (Reference Cohen, Alvarez and Nakayama2011) and Jackson-Nielsen et al. (Reference Jackson-Nielsen, Cohen and Pitts2017) find results more consistent with attentional requirements. Reinterpreting these apparent contradictions through Rosenholtz’s framework makes them more comprehensible.

Consider the paradigms in these conflicting studies. Li et al. asked participants to discriminate letters at fixation while briefly categorizing whether a peripheral scene contained an animal. High scene-categorization performance led to the conclusion that ensemble perception does not require attention. However, this setup strongly suggests that participants relied on peripheral summary statistics rather than detailed object-level recognition. One could in fact well predict that while they could categorize the scene, they might fail if asked to identify specific objects in it. By contrast, Cohen et al. dramatically increased task demands and complexity by employing multiple-object tracking in front of a scene, adding occlusion and elevated demands for fine-grained tracking. As predicted by Rosenholtz’s approach, categorization performance dropped – likely because visual perception was disrupted by greater (and central) task constraints and occlusion. Thus, what traditionally appears to be “attentional” failure might instead reflect either how increased complexity limits visual perception of how the experimental setup itself limits the extraction of peripheral summary statistics.

These insights become even more striking when applied to the AB paradigm (Raymond, Shapiro & Arnell, Reference Raymond, Shapiro and Arnell1992). People often fail to report a second target (T2) if it appears roughly 200–500 ms after a first target (T1). Surprisingly, some studies find that observers can still extract summary statistics about T2 even if they cannot explicitly report it (Dell’Acqua et al., Reference Dell’Acqua, Sessa, Jolicoeur and Robitaille2006; Joo et al., Reference Joo, Shin, Chong and Blake2009; McNair et al., Reference McNair, Goodbourn, Shone and Harris2017), whereas others show no such capacity (Huang, Reference Huang2015; Jackson-Nielsen, Cohen & Pitts, Reference Jackson-Nielsen, Cohen and Pitts2017). On closer inspection, this apparent contradiction maps almost perfectly onto where T2 appears: those demonstrating preserved summary-statistic extraction generally position T2 in parafoveal or peripheral regions, whereas those reporting no extraction present it centrally. Indeed, many AB studies also use lateralized or peripheral stimuli – either to measure N2pc and SPCN in EEG (Dell’Acqua et al., Reference Dell’Acqua, Sessa, Jolicoeur and Robitaille2006; Zivony & Lamy, Reference Zivony and Lamy2022) or simply due to the type and size of T2 (Joo et al., Reference Joo, Shin, Chong and Blake2009; Mcnair et al., Reference McNair, Goodbourn, Shone and Harris2017) – yet still interpret the resulting “blink” purely as an attentional failure, overlooking known limits of peripheral vision. A valuable next step is to incorporate eye-tracking, ensuring that fixation is accurately monitored and clarifying whether AB effects are better attributed to failures of attention or to the inherent difficulty of extracting precise information in the visual periphery.

Rosenholtz’s framework also provides a compelling way to reinterpret the AB: rather than a strict limitation of selective attention, it may reflect “temporal crowding” (Tkacz-Domb & Yeshurun, Reference Tkacz-Domb and Yeshurun2021), the temporal analog of spatial crowding. Since ensemble perception is an iterative process that updates on the order of ∼100ms and reaches its peak after ∼500ms (Epstein et al., Reference Epstein, Quilty-Dunn, Mandelbaum and Emmanouil2020), presenting T2 within 200–500ms of T1 can easily overwhelm this process. This perspective also helps to explain “lag-1 sparing” (Hommel & Akyürek, Reference Hommel and Akyürek2005), where T2 is recognized if it immediately follows T1 and even interferes with T1 – akin to backward masking. In this view, the AB need not hinge on top-down selection mechanisms but instead reflects the limited temporal resolution of ensemble coding. By reframing the blink as a manifestation of these timing constraints, we gain a clearer understanding of why summary-statistic extraction sometimes persists – especially in peripheral viewing – while at other times T2 is simply lost.

Rosenholtz’s reconceptualization also bears on the debate over whether attention is constitutive of action (Allport, Reference Allport1987; Wu, Reference Wu, Mole, Smithies and Wu2011, Reference Wu2014, Reference Wu2023b) or if action can proceed without it (Dayer & Jennings, Reference Dayer and Jennings2021; Jennings & Nanay, Reference Jennings and Nanay2016). Traditionally, robust ensemble perception in the face of the AB has been taken as evidence that perception – and thus action – can operate independently of attention (Buehler, Reference Buehler2018, Reference Buehler2023). Conversely, cases of diminished ensemble perception have been interpreted as evidence that attention is necessary (Wu, Reference Wu2023a). Rosenholtz’s model suggests both conclusions rest on a flawed premise: the “selective bottleneck” may emerge not because attention is a resource that runs out but because task demands limit visual perception and how well summary statistics can be extracted. In this view, perception is structured by the requirements of the current task, reversing the usual assumption that attention is a discrete selection mechanism in service of action. If attention is reconceived in this way, philosophers of action must grapple with a new explanatory foundation. Task complexity, rather than an internal attentional “gate,” determines which statistical regularities are available for guiding behavior. A failure to extract summary statistics in the AB no longer implies that attention is mandatory; a success in extracting them no longer proves attention is optional. The real driver is the interplay between task-set requirements and the perceptual system’s capacity for summary-statistic extraction.

In sum, Rosenholtz’s emphasis on ensemble perception bridges many gaps in visual cognition research. Rather than attributing the attentional blink and other perceptual limits to a lack of attention, we can understand them as consequences of both spatial and temporal crowding. This shift not only clarifies contradictory findings in ensemble perception studies but also reframes the longstanding “crisis” in visual attention: instead of signaling a failure in resource allocation, these phenomena may reflect deeper perceptual constraints shaped by task complexity. By embracing this perspective – and by employing methods such as eye-tracking to tease apart peripheral from genuinely attentional factors – researchers can develop a more coherent account of how, when, and why perception meets its limits.