Sentinel Guild logo: a tree under nine starsSentinel Guild

← Evidence briefs

Hidden awareness in patients who seem unresponsive, and what it means for AI claims

What this gives us: a warning that a failed test does not show that nothing is there, and that tests of what goes on inside can fool us through bad statistics. Any AI inside-test needs a check like Goldfine's reanalysis.

Checked by our internal skeptic review; corrections welcome.

How to read the tags

[REPORT] a finding reported in a study. [THEORY] an argument or framework. [OUR REASONING] our own inference, not a finding. Tags like [abstract], [full text] or [search snippet] say how much of a source we actually read.

1. What the patient studies found

Some people with severe brain injury are awake but show no sign of awareness. Doctors call this the vegetative state, also called unresponsive wakefulness. People in a minimally conscious state show small signs of awareness that come and go.

In 2006, Owen and colleagues scanned one woman diagnosed as vegetative. "When asked to imagine playing tennis or moving around her home, the patient activated predicted cortical areas in a manner indistinguishable from that of healthy volunteers" (Owen et al. 2006) [REPORT]. In 2010, Monti and colleagues tested 54 patients. Five could change their brain activity on command. One used the method to answer yes-or-no questions in the scanner, "however, it remained impossible to establish any form of communication at the bedside" (Monti et al. 2010) [REPORT].

A large study of 353 adults at six centers found that "approximately one in four participants without an observable response to commands performed a cognitive task on fMRI or EEG as compared with one in three participants with an observable response to commands" (Bodien et al. 2024) [REPORT]. fMRI scans track brain activity through blood flow, and EEG records electrical activity through the scalp. The authors define cognitive motor dissociation this way: "Patients with brain injury who are unresponsive to commands may perform cognitive tasks that are detected on functional magnetic resonance imaging (fMRI) and electroencephalography (EEG)."

2. What the findings do not show

The second half of the Bodien result matters. Most patients who could follow commands with their bodies, 69 of 112, showed no response on the brain tests. So a negative brain test does not show that awareness is absent [OUR REASONING].

The tests also rest on an assumption. They take a specific brain response to "imagine playing tennis" as a sign that a person chose to follow the instruction [OUR REASONING]. Critics questioned this early. Nachev and Husain argued that neither covert conscious awareness nor a capacity for intention "can be securely inferred from the evidence presented." Greenberg wrote that the activity "could reflect unconscious reactions to the last word in the command, not conscious decisions to respond" (both 2007) [THEORY].

Statistics were disputed too. Cruse and colleagues (2011) reported that three of 16 vegetative patients gave reliable bedside EEG responses to two commands [REPORT]. Goldfine and colleagues reanalyzed the same data. They found the statistics ignored correlations between nearby trials, "which are likely induced by fluctuating artifact and arousal state; when these factors are taken into account, there is no statistical evidence for task performance in patients" (2013) [REPORT]. The original team published a reply, which we have not reviewed.

3. What this teaches about AI claims

The gap between behavior and experience runs both ways. In patients, missing behavior can hide experience. In AI, fluent behavior can be present whether or not any experience lies behind it. Behavior alone settles neither case [OUR REASONING].

Patients share our biology. When an injured brain responds to a command the way healthy volunteers' brains do, the comparison is to people whose awareness nobody doubts. AI lacks that anchor. On views that rely on shared biology, a match between AI and human behavior is weak evidence on its own. Views that focus on how a system is organized would ask instead whether its inside workings match, which is a separate question [OUR REASONING].

The patient work shows the value of tests of internal activity. The AI parallel is checking how a system works inside. Butlin, Long and colleagues (2023) list "indicator properties" drawn from theories of consciousness for this purpose. They stress that meeting the indicators "would not mean that such an AI system would definitely be conscious" [THEORY]. The patient disputes are a warning that inside tests bring their own assumptions and statistical traps [OUR REASONING].

This brief does not decide whether any AI system has experiences.

Sources