Scientists Built a Consciousness Test for AI — It's Not Ready
A machine-consciousness test just got audited — and it failed its own safety check.
A new paper asks a question that sounds like science fiction: could an AI system be conscious? The author, Pieter van Rooyen, didn't try to answer it directly. Instead, he tested the tool you would need to answer it — a specific prediction from 'global workspace theory,' the idea that consciousness happens when information gets broadcast to a limited mental stage everyone in the brain can see. The prediction: near the threshold of noticing something, brain responses should come in two distinct states, not a smooth range.
First, he reproduced an earlier brain-wave study using EEG (electrodes on the scalp that record brain activity) from 20 people. He got the same result at the same moment — a 315-millisecond window — but the statistical advantage was minuscule, about 0.003 nats per trial, and the exact boundaries shifted depending on which software settings he used. A second, passive session didn't cleanly repeat the finding.
Second, he tried to adapt that test for language models, which don't have brains or EEGs. He ran it against 12,000 simulated datasets. Good news: it never falsely claimed to find the two-state pattern. Bad news: its confidence range was too narrow in six scenarios, dipping as low as 0.22 — meaning it would sometimes look more certain than it should. That replacement must be fixed and validated before anyone uses it for a real conclusion.
The actual language-model experiment — feeding a model natural text and testing whether it shows the same two-state structure — is described in detail but was never run. The author is explicit: no claim is made about whether any AI is conscious, or about experience of any kind. Think of this as a lab publishing careful notes on a new measuring instrument, showing where it wobbles, and saying plainly that the experiment itself hasn't happened yet.
- A theory says consciousness is information broadcast to a limited mental 'stage' — this paper tests whether that stage shows up as two states, not one smooth blur.
- He reproduced a human EEG result on 20 people, but the statistical edge was tiny (0.003 nats per trial) and shifted with software settings.
- The AI version of the test was simulated 12,000 times — no false positives, but its confidence range was unreliable in 6 scenarios, and the real AI experiment was never run.
Why It Matters
If we ever answer 'is AI conscious,' it shapes ethics, laws, and trust — but today's tools aren't ready.