Attention
DeepMind framework ↗At a glance
Focusing on what matters and filtering out distractors that look relevant.Attention is focusing on what matters for the task and filtering out the rest, especially distractors that look relevant.
The hard case is not noise but a convincing decoy: a number that looks like the answer, or a line that matches the keyword but sits in the wrong group.
- Human ref.
- ~95.0
- Best AI
- 51.0
- Gap
- +44.0
- Our items
- 15
Sub-faculties
The parts of this faculty, what our items cover today, and what is still to come
| Sub-faculty | Mirrors | Items | Status |
|---|---|---|---|
Distractor resistance Reporting what was asked for while a look-alike decoy sits right next to it. Mirrors RIAC | RIAC | 7 in the 5-min test | Covered |
Selective attention (grouping) Attending only to the part of what you see that the task is about, singled out by where it sits (its group, line or region), and tuning out the rest. How it differs: In distractor resistance, a single look-alike sits next to the target. Here what lies outside the selection looks just like what's inside it, and only position tells them apart. Mirrors ABC | ABC | 5 | Covered |
Sustained attention Staying accurate through a long, repetitive stretch. | Sustained attention to response task | 3 | Covered |
Human vs AI
Best published result on each side, 0–100
Humans. Healthy adults almost always get clinical attention tests right, which is why those tests are scored on speed instead. How often a convincing decoy catches people has not been measured.
AI. A single planted lure is enough to throw frontier models, and their misses pick the lure. They also struggle when the answer depends on belonging to the right group.
- Human ref. ~95.0, our estimate
- Our estimate of the share of adults who would report the right number from a short note with a look-alike decoy next to it, without time pressure.
- Why an estimate
- No study reports this. Clinical attention tests are built so that healthy adults almost always answer correctly, and score speed instead. This task is short and untimed, so we put adults near ceiling and allow a few percent for reading slips. It is a reasoned guess, not a measurement.
- Best AI 51.0
- Share of RIAC questions answered correctly when one decoy number is planted next to the target, pooled across the frontier models tested (2026). It is an average over models, not the best one. About the test →
One side is an estimate, so the gap of +44.0 shows a direction, not a measured margin.
Example items
Specimens written for this page. Live items are similar but never published.
Read the note, then type the number. 12 seconds.
(a) Stimulus
(b) Question
How many people can the reading room seat?
Look at the checklist and answer. 25 seconds.
(a) Stimulus
(b) Question
In Garden, how many tasks are done, counting sub-tasks?
Press for every digit except 3.
(a) One digit every 1.15 s
Benchmarks
| Benchmark | Origin | Items | Best AI | Humans |
|---|---|---|---|---|
| RIAC | Kaggle × Google DeepMind hackathon, 2026 | 7 | 51% | — |
| ABC | Kaggle × Google DeepMind hackathon, Attention track winner, 2026 | 5 | weakest | — |
| Sustained attention to response task | Sustained attention to response task (Robertson et al., 1997), 1997 | 3 | n/a | — |
In everyday life
- Finding the right figure in a busy invoice
- Following one conversation in a noisy room
- Ignoring a tempting but irrelevant detail