Perception

DeepMind framework ↗
Human 91.4 · AI 78.4Humans lead +13.02 benchmarks →

At a glance

Taking in information from the world and making sense of it.

Perception is taking in information from the world, such as images, sounds and text, and turning it into something usable: objects, positions, quantities, events.

It looks effortless in people because most of it happens below awareness. Reading an analog clock, for instance, combines recognising hands, estimating angles and knowing that the short hand moves continuously.

Human ref.
91.4
Best AI
78.4
Gap
+13.0
Our items
13

Sub-faculties

The parts of this faculty, what our items cover today, and what is still to come

Sub-facultyItemsStatus
Visual-spatial reading
Reading positions and angles precisely: clock hands, dials, gauges.
Mirrors ClockBench
7
in the 5-min test
Covered
Reading text precisely
Seeing exactly what is written, letter by letter: counts, typos, doubled words.
Mirrors CUTE
6
in the 5-min test
Covered
Hearing
Telling sounds apart: pitch, rhythm, a word in background noise.
0Not covered yet

Human vs AI

Best published result on each side, 0–100

Human ref.
91.4
Best AI
78.4

Humans. Healthy adults find both tasks easy: they follow everyday videos, and read analog clocks to within a few minutes.

AI. Multimodal models describe images fluently, and still trail people on everyday video. Precise spatial reading, such as an analog clock, was a glaring weak spot until 2025; the best models have since nearly caught up.

Human ref. 91.4
Share of the Perception Test's multiple-choice questions about short real-world videos that people with no special training answered correctly, as reported in the test's paper (2023). About the test →
Best AI 78.4
The same questions answered by Gemini 2.5 Pro, from Google's own report: the best published result. About the test →
255075100Perception+13Generation-35Attention+44Learning0Memory+21Reasoning-27Metacognition-55Executivefunctions+17Problemsolving+2Socialcognition+4
Human referenceOur estimateBest published AILabel = human − AI, points

Example items

Specimens written for this page. Live items are similar but never published.

Count the lettersspecimen · not scored

Read the sentence, then type the number. 30 seconds.

(a) Stimulus

Tomorrow the tall statue gets a fresh coat of paint.

(b) Question

How many times does the letter t appear (capital or small)?

Figure 1. A letter-counting item. People read words as wholes, models as tokens; both can lose track of single letters. Benchmark: CUTE. Answer: 8: Tomorrow, the, tall, statue (twice), gets, coat, paint.
CUTE benchmark page →
Read the clockspecimen · not scored

What time does each clock show? Within two minutes counts.

(a) Dials

XIIIIIIIIIVVVIVIIVIIIIXXXI
Figure 2. An analog-clock item. One dial has ticks only, one has Roman numerals. Benchmark: ClockBench. Answer: 2:47 and 11:23.
ClockBench benchmark page →

Benchmarks

2 rows
BenchmarkOriginItemsBest AIHumans
CUTECUTE (Edman, Schmid & Fraser, EMNLP 2024), 2024613.1%—
ClockBenchClockBench (Safar), 2025783.3%90.7%

In everyday life

  • Reading a dial or a chart
  • Judging distance before crossing a road
  • Spotting what changed in a scene