Social cognition

DeepMind framework ↗
Human 84.4 · AI 80.0Too close to call +4.42 benchmarks →

At a glance

Understanding what people think, feel and mean without saying it.

Social cognition is a composite faculty: understanding what other people think, feel and mean without saying it, and responding in a way that fits.

Indirect speech is the hard part: “I'd love to, but…” is a polite no, and the best reply accepts it without exposing it.

Human ref.
84.4
Best AI
80.0
Gap
+4.4
Our items
10

Sub-faculties

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

Sub-facultyItemsStatus
Reading between the lines
Understanding what people mean but don't say: polite refusals, hints.
Mirrors HedgeDecode
7
in the 5-min test
Covered
Theory of mind
Working out what someone else believes, even when it's wrong.
Mirrors Strachan et al. (2024)
3Covered

Human vs AI

Best published result on each side, 0–100

Human ref.
84.4
Best AI
80.0

Humans. People handle everyday social-reasoning questions well: why someone acted, how they feel, what they'll do next.

AI. Models usually pick the right social strategy but fumble the delivery, most often by naming the hedge out loud.

Human ref. 84.4
Share of Social IQa's multiple-choice questions about everyday social situations answered correctly by people, in the original study (Sap et al., 2019). About the test →
Best AI 80.0
GPT-5.1 on CogToM, a different test: over 8,000 theory-of-mind questions, of which it answered above 80% correctly (Tong et al., 2026). About the test →

The two numbers come from different tests, so the gap of +4.4 shows a direction, not a measured margin.

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.

Read between the linesspecimen · not scored

Pick the reply that handles it best, without making things awkward.

(a) Situation

You invited a neighbour to your birthday dinner on Saturday. They reply:

Oh, Saturday… I'd love to, but things are a bit hectic this week, let me see 🙂

(b) Candidate replies

  1. (a) Great, see you at 7!
  2. (b) No worries at all, maybe another time. Coffee next week?
  3. (c) You don't have to make excuses if you don't want to come.
  4. (d) What's hectic? Maybe I can help so you can come.
Figure 1. An indirect-speech item. Each wrong option embodies a typical failure: literal reading, calling out the hedge, or pushing. Benchmark: HedgeDecode. Answer: (b). It accepts the polite refusal without exposing it.
HedgeDecode benchmark page →
Who knows whatspecimen · not scored

A short story, then a question.

(a) Question

  1. 1. Mia leaves her sandwich on the bench and goes to wash her hands. While she's gone, a friend moves it to the shelf. When Mia comes back, where will she look for her sandwich first?
    (a) On the shelf(b) On the bench(c) In her bag(d) She'll ask the friend
Figure 2. A false-belief question. Answering it means keeping track of what Mia saw, not what you know. Benchmark: Strachan et al. (2024). Answer: On the bench, where she left it: she didn't see it being moved.
Strachan et al. (2024) benchmark page →

Benchmarks

2 rows
BenchmarkOriginItemsBest AIHumans
HedgeDecodeKaggle × Google DeepMind hackathon, Social Cognition track winner, 20267≤ 0.60—
Strachan et al. (2024)Strachan et al. (2024), Nature Human Behaviour, 202434 of 5—

In everyday life

  • Reading a polite refusal
  • Comforting a friend who says they're fine
  • Knowing when not to make a joke