Francesco Paolo Lezza

The constant is in whoever’s looking

Fine, we get it: a machine shuffles symbols according to rules, and machines don’t understand. We’ve heard it enough, let’s take it as given. The interesting question isn’t what the machine does, it’s what we do in front of the machine. Because whoever looks reads a meaning into it that the machine never put there.

This gesture, the observer projecting meaning onto form, has a birth date far older than chatbots.

Ramon Llull, as Kelly Clancy recounts in Truth Machines, wanted a book that could talk back to its readers and answer any question of faith. In his main work, the Ars Magna, he gave concrete shape to the mood of his age: people believed combination could produce truth. He built a kind of logic machine to prove the existence of God, made of concentric wheels that combine concepts by mechanical rotation.

You can see it clearly in Borges. In 1937 he says it plainly: the machine just combines, the meaning is added by the reader.

He’s already smiling, and he pictures the reader smiling too, at the phrases the wheel produces: “Goodness is great,” “Greatness is good,” down to the syllogism that follows. Then he notes the thing that matters: those terms are interchangeable. In place of Goodness and Greatness you can put Entropy, Time, Electrons, Quantum, and the wheel keeps turning exactly the same. It’s the machine that’s indifferent to content; the meaning is carried by whoever reads.

That line from Llull to today’s models has already been traced, but read as a chase after mechanized certainty. Here I read it the other way, from the side of whoever’s looking.

Jump forward seven centuries and we get ELIZA, applying textual patterns with hand-written rules. Here too Weizenbaum watches people project understanding into it, and it disturbs him. Same structure as Llull: form produced by explicit rules, meaning projected by the observer.

Another jump. With the perceptron, form stops coming from hand-written rules and starts emerging from learned data. It’s like going from the balloon to the airplane: the engine changes, not the fact that it flies. And above all, whoever’s looking doesn’t change.

Please note that there’s no need to settle which is the canonical technical break, some would say the transformer and not the perceptron. What matters is the first point where form shifts from hand-written to learned: it’s a pivot, not a history of AI.

We reach today. Current models are black boxes that learn, and the form is more convincing than ever. And this time there’s substance underneath: the learned machine does real things, it’s not a trick. But precisely for that reason the observer projects even more. That the projective dynamic from symbolic to connectionist is the same has already been theorized: it’s the thesis of Noosemia, a work by De Santis and Rizzi that builds exactly the continuity between the ELIZA effect and today’s models as one projective dynamic. I cite them because it’s the theoretical formulation closest to my axis: they stop the projection at meaning, I add the two ends they’re missing, Llull before and reliability after.

Three different ways of manufacturing language: combination by hand, explicit rules, learned weights. And one constant gesture on the other side.

It isn’t a feature of this or that technology. It’s a constant of whoever’s looking. The projection crosses two paradigm shifts intact: that’s why it’s structural.

And this is where it’s worth stopping, because “projecting meaning” isn’t one single thing.

Projecting meaning is interpretation: it’s inevitable, and it’s how we read any text. Projecting reliability is a different operation: attributing truth on the basis of form alone. Taking the fact that the output is well-formed as a guarantee that it’s true, or logically stable. They’re two different things, and nothing in the system says they go together.

Every guarantee taken for granted is an undeclared assumption: something I hold as true without anyone having put it in writing. It’s exactly the kind of assumption someone can exploit. That’s where the security problem begins, the moment the gesture stops being harmless and becomes attackable.

Meaning is nothing to intervene on, and that’s fine. Reliability is, because in the learned era apparent confidence leaves measurable traces. And if it’s measurable, you can look closely at where it breaks: where the system seems most reliable exactly when it’s least.

That’s where the question stops being philosophical and becomes engineering. And that’s where the work is.

Sources