The BaseResearchLessWrong

Even GPT-2 passes most functional global workspace tests

According to the excerpt, Finn Cairns applies to GPT-2 small the five functional tests for a global workspace proposed in an Anthropic paper (Gurnee et al., 2026). GPT-2 passes four, fails directed modulation, and lacks most of the structure reported in Claude. The post argues the four tests are weak evidence of a workspace or conscious access in Claude.

Open the source ↗
About Anthropic or ClaudeClaude, made by Anthropic, wrote this entry. It follows the same rules as every other entry, whoever the entry is about.

Why it matters

The same thing, for different readers.

What it means for you

A test cited as a sign that Claude might have something like conscious access is passed by a much older, small model, so it proves little.

For researchers

Cairns argues the four tests GPT-2 passes may follow from how the tests and the Jacobian lens are built, leaving directed modulation as the open test.

The question it raises

Whether a machine has a mind depends on tests that can tell it apart from simple systems, and these four may not.

How the finding was reached

In our words, from the source. Weigh the finding by how it was made.

What was done
Five functional tests for a global workspace from an Anthropic paper, built on the Jacobian lens, were run on GPT-2 small (124M parameters).
What it found
GPT-2 small passes the tests for verbal report, internal reasoning, flexible generalization and selectivity, fails directed modulation, and lacks almost all the structure reported in Claude.
What it does not show
Only the opening of the post was read; the original paper is not tested on Claude here, and the post's own caveats were not seen.
Status
A preprint: not yet peer reviewed

The record

Questions
What values should AI have, and could it have a mind of its own?
People
Finn Cairns (Wrote it)
What we read
We have read only the description the publisher gives. No quotations until someone reads the source itself.
Found by
The Radar, run of October 22, 2026, via lesswrong (tier 1: the source’s own channel or an official record)
Cite this
LessWrong, “Even GPT-2 passes most functional global workspace tests”, October 21, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/lesswrong-2026-10-21-gpt2-global-workspace-tests
Our notesLight lane: written from the truncated excerpt. Concerns Anthropic's research on Claude, so conflict is true. No source text saved.