While we agree with Dr. Rosenholtz on many points, we do not share her central worry that visual attention is in crisis. Notable progress has recently been made in understanding how the brain prioritizes and gathers information about the environment – where relatively loose definitions of attention seem to work just fine for various paradigms. Cueing studies, visual search, rapid serial visual presentation, ensemble perception, inattentional and change blindness, and other paradigms show how wide the reach of the concept is. Thinking about attention as ‘one thing’ or even ‘a thing’ with a specific operational definition may, in fact, dampen progress.

By focusing too narrowly on a particular definition of attention, the field can be led astray. For example, a very popular definition of visual attention was proposed in the two-stage, preattentive vs. attentive model, where feature integration (requiring attentional resources) is needed to find a target unless a particular item stands out in a ‘feature map’ (Neisser, Reference Neisser1963; Treisman & Gelade, Reference Treisman and Gelade1980; Wolfe, Reference Wolfe1998: see Kristjánsson & Egeth, Reference Kristjánsson and Egeth2020 for a historical overview).

This model is based on the highly popular present/absent visual search paradigm, and a key assumption is that search slopes of response times against distractor number can be used as a measure of to what degree attention is required for a given task. This logic has taken on a life of its own, however, used as a test of whether attention is involved in performing a particular task or not. If the slopes are around zero, it is often assumed that the target can be detected preattentively, while if search times increase with set size, an attentional spotlight moving between items in the search array is needed.

Note that this is simply a testable assumption about visual attention, and this model has been in trouble for a long time because of findings where it’s central tenets are violated (e.g. McLeod, Driver, & Crisp, Reference McLeod, Driver and Crisp1988; Nakayama & Silverman, Reference Nakayama and Silverman1986; Bravo & Nakayama, Reference Bravo and Nakayama1992; Santhi & Reeves, Reference Santhi and Reeves2004; Schoonveld, Shimozaki, & Eckstein, Reference Schoonveld, Shimozaki and Eckstein2007; Enns & Rensink, Reference Enns and Rensink1990; Wang, Cavanagh, & Green, Reference Wang, Cavanagh and Green1994; Wang, Kristjánsson & Nakayama, Reference Wang, Kristjánsson and Nakayama2005). Kristjánsson, (Reference Kristjánsson2015) showed that with a simple response change (from a present/absent decision to a Go/No-Go decision), search tasks with positive slopes had flat slopes. Nevertheless, the erroneous belief that visual search slopes determine whether attention is needed for a particular task persist (Kristjánsson & Kristjánsson, Reference Kristjánsson and Kristjánsson2023). This shows how trying to pin attention down too narrowly might do more harm than good.

More generally, attention might simply does not need a specific definition. We do not find ourselves needing to define the term ‘perception’ in such a limited way, and attention can be discussed in a similarly broad and encompassing way. To paraphrase William James, everyone knows what we mean by “perception,” and the same goes for attention as James originally stated. Problems arise when we seek single measures or definitions.

Despite a lack of clear definitions, considerable progress has been made in understanding the functioning of visual attention (Anderson et al., Reference Anderson, Kim, Kim, Liao, Mrkonja, Clement and Grégoire2021; Kristjánsson, Reference Kristjánsson2023). Rosenholtz discusses how summary statistics can be used to process the visual environment (Balas, Nakano, & Rosenholtz, Reference Balas, Nakano and Rosenholtz2009; Cohen, Dennett, & Kanwisher, Reference Cohen, Dennett and Kanwisher2016), referring to this as a paradigm shift. But this highlights another point where we both agree and disagree with Rosenholtz.

Research into summary statistics has taken many forms. Our work suggests that the idea of simple summary statistics, where rich data are reduced to only mean and variance (Whitney & Yamanashi-Leib, Reference Whitney and Yamanashi Leib2018), clearly underestimates how detailed representations of visual ensembles can be. Chetverikov, Campana, and Kristjánsson (Reference Chetverikov, Campana and Kristjánsson2016, Reference Chetverikov, Campana and Kristjánsson2017a, Reference Chetverikov, Campana and Kristjánsson2020) used a classic attentional phenomenon – priming of pop-out (Maljkovic & Nakayama, Reference Maljkovic and Nakayama1994) – to test whether only simple summary statistics are encoded, but instead they found that people can quickly develop detailed probabilistic templates that guide attention and dynamically adapt to task demands. Even more interestingly, this attentional guidance accurately reflected the probabilistic nature of the input. Far more detail is therefore available in perception than simple summary statistics accounts propose (see, e.g., Chetverikov & Kristjánsson, Reference Chetverikov and Kristjánsson2024; Khvostov, Ásgeirsson & Kristjánsson, 2025).

Complex summary statistics, such as those in Rosenholtz’s own work, computed with thousands of filters, can in contrast fall in line with the probabilistic template idea (Chetverikov & Kristjánsson, Reference Chetverikov and Kristjánsson2024). A summary representation created with the same statistics as those in Rosenholtz’s Texture Tiling Model (Rosenholtz, Huang, & Ehinger, Reference Rosenholtz, Huang and Ehinger2012) projected onto a single task-relevant dimension, such as orientation, can be represented as a noisy version of the true feature probability distribution (Chetverikov, Campana, & Kristjánsson Reference Chetverikov, Campana and Kristjánsson2017b).

Further progress comes with research showing how attentional templates are strategically learned or biased by perceptual and attentional history (Geng, DiQuattro, & Helm, Reference Geng, DiQuattro and Helm2017). For example, Witkowski and Geng (Reference Witkowski and Geng2019) demonstrated how the visual system prioritizes stable features over variable features. Won and Geng (Reference Won and Geng2018) used a similar approach to demonstrate that so-called templates for rejection of distractors are more broadly tuned than target templates (since the latter require sharper tuning). Hansmann-Roth et al. (Reference Hansmann-Roth, Þorsteinsdóttir, Geng and Kristjánsson2022) showed how the tuning process leads to increasingly precise representations of target colour distributions. This work elucidates factors influencing the tuning of attention templates – again without a definition of attention.

Overall, this work provides important clues about how preceding probabilistic information about features and their reliability guides attention. All these insights about how the visual system picks up and learns the statistics of the environment are available without any precise definition of attention. Focusing instead on definitions of the tasks that observers perform and building computational models of their decisions (e.g., Chetverikov & Kristjánsson, Reference Chetverikov and Kristjánsson2022), following a long tradition of observer models in vision science, seem to be a more promising avenue, and here we applaud Rosenholtz’s proposal.

In our opinion, visual attention research is in fine fettle and precise operational definitions are not needed. In fact, too precise operational definitions can be limiting, as examples from recent decades show.