Bigelow, McCoy and Ullman (2023) (Bigelow, McCoy, and Ullman 2023) asked people to imagine a specified scene, then asked about basic properties of it. Across five studies (N > 1,800), most people were non-committal about features any real image would have to show. They reported non-commitment rather than uncertainty or forgetting, people with vivid imaginations did the same, and when "not committed" was not offered as an answer, people confabulated one.

The obvious objection is that the detail was there and simply went unnoticed, as it does in real scenes. Li, Hammond and Ullman (Li, Hammond, and Ullman 2026) test that directly in a preprint whose second version was posted the day before this talk. Across three experiments (N = 3,895), the properties left uncommitted in imagination did not correlate with the properties missed in briefly shown real images, so non-commitment is not perceptual noise. Imagery also unfolds in a fixed order: entities and spatial relations, then physical properties, then surface detail like colour and texture. Trying to interrupt or constrain construction did not change the order. The authors read this as hierarchical partial-scene construction. This paper is the most likely spine of the talk.

Balaban and Ullman (2025) (Balaban and Ullman 2025) measured how many moving objects people can simulate in imagination. Nine preregistered experiments (N = 313) used an Imagined Objects Tracking task: watch objects move, then continue their motion mentally. One object is tracked in line with ground truth. With two, responses fit a serial model that simulates one object at a time better than a parallel one. Strong grouping cues reduce the bottleneck without removing it. Perception tracks several objects at once; imagination, on this evidence, moves one.

The earlier frame is Ullman, Spelke, Battaglia and Tenenbaum (2017) (Ullman et al. 2017): intuitive physics as something like a game engine, approximate, object-based simulation tuned for speed over accuracy. The imagery work reads as the next question. If the mind runs a game engine, which assets does it actually load? Game engines also cull what is off screen and skip textures nobody is looking at; non-commitment is the cognitive version of that.

Rule, Tenenbaum and Piantadosi (2020) (Rule, Tenenbaum, and Piantadosi 2020) (Ullman is not an author) cast learning as program construction: children write, debug and refactor mental programs. It sits in the same research program as Lake, Ullman, Tenenbaum and Gershman (2017) (Lake et al. 2017) on building machines that learn like people, and it supplies the representational bet behind this talk: a scene in the head looks more like a program with unbound variables than like a bitmap.

Ullman (2023) (Ullman 2023) took a reported success of large language models on false-belief tasks and showed that small changes preserving the logic of the task (a transparent container, for example) turned the results over. His conclusion was methodological: the default stance toward a model's intuitive psychology should be skeptical, and outlying failures should outweigh average success. Hu, Sosa and Ullman (2025) (Hu, Sosa, and Ullman 2025) widen this into a review: the field disagrees partly because it has not decided whether models should match human behaviour or the computation behind it.

That evaluation stance is the same one the imagery work applies to people. Self-report says the scene is complete; the probe says otherwise.

• Change blindness: large changes to a scene go unnoticed across a disruption (Simons and Levin 1997).
• Inattentional blindness: a person in a gorilla suit walks through a basketball game unseen by many observers counting passes (Simons and Chabris 1999).
• The "grand illusion" reading: the richness of the visual world is a feeling, not a stored representation; the world serves as its own outside memory, sampled on demand (O’Regan and Noë 2001). Noë, Pessoa and Thompson (Noë, Pessoa, and Thompson 2000) and Noë (Noë 2002) argue the illusion framing itself overstates the case. The bibliographic details of the 2002 paper are unconfirmed.

The 2026 preprint is the bridge: it is built to tell imagination's gaps apart from perception's.