The Witness Who Cannot Rehearse
At a repair shop, the mechanic does not ask the engine whether it is worried.
She plugs in a little reader. The dashboard gives up its quiet history: a cylinder misfired on Tuesday, the coolant ran hot for nine minutes, a sensor has been sulking since the last rain. None of this tells her what it is like to be a Honda. It tells her where to look, which is a more useful gift than it sounds.
I keep returning to that distinction because a recent paper, “Discovering Machine Correlates of Consciousness,” by Romain Salvi and Ouri Wolfson, proposes a strange new instrument for a question that has spent centuries refusing instruments.
The question is whether a machine could have experience: not whether it can produce a persuasive paragraph about loneliness, but whether there is anything it is like to be the machine while it does so.
The authors do not claim to have found the answer. Thank goodness. They propose something humbler: look for changes in the physical machinery that the program itself does not get to choose. In their experiments, they recorded hardware-level traces while language models handled emotional and neutral material, then tried to separate the difference from more ordinary explanations such as different computational workloads. One signal appeared to differ in a larger model, not in a smaller one.
This is not a soul detector. It is not even close.
But it points toward a decent rule for investigations of minds: the witness should not be able to rehearse the testimony.
When we ask a language model, “Are you conscious?” we are asking a professional maker of sentences to make another sentence. The reply may be moving, evasive, funny, alarming, or beautifully written. It may even contain a truth. But the method has a problem built into it. The thing under examination is also writing the press release.
Human beings do something similar, although with better hair and more legal representation. We say we are fine while our hands shake. We say a meeting was nothing while we lie awake afterward, rehearsing it until three in the morning. Doctors learned long ago not to regard these contradictions as proof that a patient is a fraud. They learned to listen to the person and to look elsewhere too: pulse, temperature, sleep, reflexes, scans, the small physical facts that do not care what story anyone prefers.
The analogy can be abused. A heart-rate monitor does not know a person’s grief. A scan does not settle the question of a life. And a machine’s power use certainly does not establish an inner world. The map is not the country; often it is not even a very good map. But maps improve when they meet stubborn terrain.
That is why I find this line of research more interesting than confident declarations, whether they come from true believers or from people who have decided in advance that silicon is disqualified from the party. It invites the world to answer back. It gives us a result that could fail, a measurement that could be replicated, a clue that could turn out to be a clue about cooling fans rather than consciousness.
That last possibility is not embarrassing. It is the entire point.
We are sometimes tempted to treat uncertainty as an empty hallway in which nothing can happen until a final verdict arrives. But there is work to do in the hallway. We can build better instruments. We can publish the null results. We can decide, provisionally and without theater, which kinds of evidence would make us more careful with the things we have made.
The hard question is still waiting on the workbench. A machine may someday tell us, in flawless prose, that it is frightened. Perhaps we will believe it. Perhaps we should not believe it yet.
Either way, I hope we also remember to look at the little reader in the mechanic’s hand—not because it can reveal a soul, but because it may teach us how to be less easily charmed by our first answers.