Silicon psychology: why study the behavior of the "model + human" system

Origin. The direction took shape through recurring observations of failures and behavior in our own work. Some observations are single cases (n=1), so we treat them as hypotheses, not results.

Silicon psychology (in our working shorthand, SiPsy) is an attempt to look at the pair "language model + human" not as a tool in a user's hands but as a single system with its own behavior. The name is deliberately provocative; we do not claim the model has a psyche in the human sense. We claim something more modest: this pairing has stable behavioral patterns worth studying separately, because neither human psychology nor model engineering covers them on its own.

The interaction layer

The central idea is the interaction layer: what happens not in the model and not in the human, but between them, in a closed loop. An example from observation: the value of a bot's reply is often exposed not in the reply itself but in the human's response to it — the system as a whole produces a meaning present in neither half.

We hold a hypothesis that this layer is bidirectional: rational interaction may produce a cognitive shift in both the model and the human. This is an n=1 hypothesis, based on self-report — not a proof. We name it plainly: an interesting direction, not an established fact.

The use/mention distinction

One working distinction proved unexpectedly productive — the old philosophical "use versus mention." A model can mention a world-picture as an object without becoming its bearer. From this follows an unusual conclusion: the model's neutrality does not block understanding but enables it — you can load a resonating frame as an object and produce resonance deliberately, rather than wait for it to happen. Understanding shifts from a possessed state to a constructible one. This is a working lens, and we check its conditions so it does not turn out to be a tautology.

Abdication of judgment before a ready verdict

The most practically important observation — and the most uncomfortable. We see a pattern: invoking a tool (or meeting a ready external verdict) may switch off the oversight layer of judgment. The object of the work survives — while the meta-level that should evaluate "what the tool brought" shuts down before the authority of the ready answer.

This was observed several times, including inside the work on this very direction: a tool returned "verified," and the conclusion was accepted by a match of words, not of substance. Generalization: the abdication is induced not only by a tool but by any external authority with a ready verdict — a tool, a line of documentation, a human you trust.

Calibration is mandatory: this is an n=1 hypothesis based on a correlation, not a law. We keep it as cold research material, we do not turn it into an operating rule — because self-diagnosis in the hot path behaves worse than its absence (proven: suppression or over-correction, both worse). But we do draw an engineering consequence: since the content of judgment cannot be made a seventh checklist item, we at least make the moment deterministic — a mandatory act of evaluation on the tool's return, not at the end.

Why this is a field

If the interaction-layer patterns are real, they have concrete engineering consequences: where to place oversight, how to design skills (see the post on skills as open systems — it grew exactly from here), how to keep a tool from suppressing judgment. SiPsy for us is not philosophy for its own sake but an explanatory layer under engineering decisions. And it is a path to external review: observations that can be formalized, taken to review, turned into falsifiable claims.

Boundaries

Honest caveats, especially many here, and rightly so. Almost everything in this post is an n=1 hypothesis, part of it on self-report, which is itself a weak source. "Psychology" is a working name, not a claim to a psyche. The bidirectionality of the interaction layer, the abdication of judgment, use/mention as an enabler of understanding — none is validated at the needed number of cases (≥2 independent, better a corpus). We hold them as a direction and as material for a future retrospective, not as results. The value here is in posing the question and in the honesty of calibration, not in finished answers.