When the signal speaks, who signs?

In September the case of the third participant in Neuralink’s VOICE study was reported: a man with ALS who speaks through a synthetic voice decoded from the speech motor cortex. The device remains investigational, not FDA-approved. One of the physicians involved said: “The impact of giving him his speech back has been remarkable.”

It is. Before any legal reflection, this must be said: giving someone back their voice is one of the finest things technology can do. That is exactly why it deserves careful thought.

Who speaks when the machine speaks?

Until now law has always had a fixed point: a declaration of will passes through a gesture of the person. The signature, the voice, the click, even a nod before a notary. The gesture is the border between what I think and what I say.

In an interface that decodes speech, that border moves inside the system. The decoder chooses, among many possible readings of the signal, the most probable one. It often completes, corrects, anticipates. Is the sentence that comes out the person’s? Almost always. But not always, and not entirely.

Now imagine that voice saying “I consent” to processing, “I accept” a contract, “I revoke” a will.

Three problems no rule solves

1. Intention. How do we tell a sentence thought from a sentence willed? In speech, the act of uttering it makes the difference. A decoder needs an equivalent: a separate confirmation signal, designed on purpose.
2. Error. If the decoder gets it wrong, who is liable? The manufacturer, the physician, the user? The study participation agreement, I would bet, does not say so in a way the patient can understand.
3. Consent to consent. To use the device, the person signs documents. In my study of brain-computer interface consent documents, none was below the experimental accessibility threshold: syntactic density 72 out of 100, semantic distance 68, threshold 45. Those who most need to understand receive the hardest texts.

What I learned by being wrong

This site’s log has an entry marked “wrong”: “One sensor is enough”. I thought a single signal could prove understanding. It cannot. That gave me a rule I always apply: convergence between different sources, never a single indicator.

The same rule should apply to decoded will. A declaration with legal effect produced by a neural interface should require at least two independent confirmation channels, and a simple, always-available way to cancel it. In the Habeas Mentem Charter: “Withdrawing costs as much as consenting.”

The question that remains

Floridi, in his foreword to Mente e Tecnologia, writes that designing a neural-data notice a non-specialist cannot understand “is a way of taking what was not given”. I would add: designing a decoder with no way to say “that’s not what I meant” is a way of making it say what was not said.