Perception Test

Pătrăucean et al., Google DeepMind, 2023

What it measures

A published test we don't mirror: it supplies a reference figure, not test items

The Perception Test checks how well models understand short real-world videos: about 11,600 clips, 23 seconds long on average, filmed by around 100 people worldwide.

Questions cover memory, abstraction, physics and semantics, and ask what happened, why, what comes next, and what would have happened otherwise.

Example question

Written for this page in the style of the test, not one of its questions, so the test set stays unspoiled

Follow the ballspecimen · not scored

Watch the video, then pick an answer. Described here in words; the real items are videos with sound.

(a) Video, 20 s

0–6 s. A hand puts a red ball under the left of three upturned cups.

6–13 s. The left and middle cups swap places.

13–20 s. The middle and right cups swap places.

(b) Question

Where is the ball now?

(a) Under the left cup
(b) Under the middle cup
(c) Under the right cup

Figure 1. A cup-shuffle item: keeping track of a hidden object as things move. In the style of Perception Test. Answer: (c). The first swap carries the ball to the middle, the second to the right.

Human vs AI

The faculty-level figures we take from it, 0–100

PerceptionGap +13.0

Humans

91.4

Perception Test, untrained adults

Best published AI

78.4

Perception Test, Gemini 2.5 Pro

These are third-party results on different questions from ours. Where the numbers come from

All reference tests →