RESOURCE · COGNITIVE TESTS AT SCALE · Results · October 2026 · work in progress

Which senses does a text-only AI hear in a word?

A text-only model matches people’s ratings of touch words best and inner-body words least. That is a pattern in text judgments, not evidence that it senses either one.

How to read this. We sampled 600 words from a public mirror of the Lancaster Sensorimotor Norms. Jev rated each word on eight dimensions, in two wordings. Bars show correlation with people’s ratings; whiskers are bootstrap 95% intervals. The probe calls are in the cost ledger but not in the analysis.
Bar chart of Jev-human correlations for eight sensory and body dimensions.
Touch is highest. The highlighted inner-body dimension, interoception, is lowest.

Text tracks touch best, inner body least

Jev’s highest correlation was touch: 0.783 (95% interval 0.745 to 0.818). Its lowest was interoception, the inner-body sense: 0.494 (0.431 to 0.557). This does not mean it feels touch or body states. It means its text judgments align with the human ratings to different degrees.

The surprise was taste and smell

We expected smell and taste to be weaker than vision. They were not: taste was 0.747 and smell was 0.705, above vision at 0.683. Hearing was 0.679; hand/arm was 0.694; mouth was 0.645.

Can it name the main sense?

Across the six perceptual senses, Jev and people chose the same highest-rated sense for 379 of 600 words: 63.2%. A uniform six-way chance baseline is 16.7%. That is useful agreement, but below the pre-registered 70% target.

What it cost

All logged L1 calls, including the probe, were 10,080 calls and $0.1548. Source: the `calls*.jsonl` ledger read by this build script.

Sources

  1. Lynott, Connell, Brysbaert, Brand and Carney (2020), The Lancaster Sensorimotor Norms (abstract read through a search result)

The file used here is the `seantrott/cs_norms` GitHub mirror; where it came from was not checked against the authors’ source. The human side is Mechanical Turk ratings. One model version: jev-1.13.0. Independent work using Jev from TypeSafe AI; not affiliated with or endorsed by TypeSafe.