Every Contact Leaves a Trace: Does the way we engage with AI have a computational signature inside the model?
Does the way a person leads an exchange with an AI change the depth and quality of what it produces? I start with the part that can be measured: does it leave a trace inside the model? The model and task were held fixed while the user's side varied. Two communication styles often seen in people’s exchanges with AI were compared. In the Navigator style, a person adds specifics, corrects, and steers the exchange across turns. In the Operator style, they give a directive, self-contained request to be fulfilled as stated. Conversation segments representing the two styles were matched on task, domain, and token length, then traced through one reader model, gemma-2-2b, using Gemma Scope transcoders and circuit tracing. There were 54 matched pairs; the 35 rated higher-confidence before tracing form the primary cohort. Over the model's interior layers (L7–L25 of 26), the activation-weighted mean layer is later in the Navigator segment in 31 of 35 pairs (+0.554 layers; sign p = 3.5×10⁻⁶). Navigator graphs are also smaller and sparser, with fewer nodes, fewer edges, and a lower feature ratio, so the shift is not produced by graph size. The direction holds over feature nodes alone, at every interior cutoff from L5 up, and across all 54 pairs (43/54). Including the earliest layers reverses the sign: Navigator exchanges have more attribution mass near the input and a later interior center. This is an association measured in a reader model; it does not establish a causal mechanism. The human side of an exchange has a measurable internal correlate, which gives a reason to include it as a variable in model evaluation and in teaching people to work with AI.
Authors
- Petya Sarafova
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24287638
- Primary Topic
- Explainable Artificial Intelligence (XAI)
- Type
- preprint