Measuring the friction in AI use that standard evaluations miss
THE GAP
AI platforms can see what people ask. They cannot see what happens between the person and the model while they work together.
Evaluations score the model’s output. Aggregate analysis counts patterns across millions of conversations. Surveys ask users afterwards. None of them follows the individual exchange as it unfolds, and that is where friction lives.
THE LAYER
I measure the relational form of human–AI exchanges: how roles, expectations and interaction patterns shift across turns, on both sides of the conversation. When the AI starts to carry the decision instead of supporting it, I call that role drift. It needs no technical failure. It happens in the exchange itself.
THE INSTRUMENT
Zenn is a live instrument that observes this layer. Real conversations, contributed with consent, are scored turn by turn against a codebook of friction codes, human turns and AI turns alike. Every finding is traceable to the exact turn it comes from.
Human judgment anchors the scoring. A classifier scores every turn; a human judge decides borderline cases and samples, and each judgment recalibrates the classifier. Agreement between judges is measured, not assumed.
Zenn is model-independent. The current cases come from ChatGPT, Claude and Gemini users.
See it work: a three-minute walkthrough.
WHAT IT YIELDS
For those who build and evaluate AI, it adds a view their own evaluations cannot produce: human judgment on real exchanges, followed over time, on both sides of the conversation, with every finding checkable against its source.
It does not assess content. Factual accuracy, safety and policy compliance remain the domain of existing evaluations. The two layers do not compete; one measures what the other cannot see.
GO DEEPER
Zenn: A Live Instrument for Observing Relational and Cognitive Friction in Human–AI Interaction
Provenance deposit · Zenodo · August 2026, latest version October 2026 · record public, files restricted · DOI 10.5281/zenodo.21995957
Before the Recommendation: A Conceptual Framework for Clinical AI Attribution, Attention, and Influence
With Emma Paraschiva, Mujahid M. Altamimi & Sarah E. Curtright · Preprint, SSRN · July 2026 · DOI 10.2139/ssrn.6996320
Relational Perception in AI Use: Theoretical Framework
Preprint, Zenodo · April 2026 · CC BY 4.0 · DOI 10.5281/zenodo.19844178
IF YOU RECOGNIZE THIS
If you work on how people actually use AI (evaluation, societal impact, human oversight), I’d like to compare notes.
Henri Hommersom
Relational AI Analyst
info@relationalperception.org
Relational Perception, Utrecht, Netherlands. Registered Dutch CoC number 42033246. VAT number NL005444927B33
ABOUT
I work independently as a Relational AI Analyst, with no vendor relationships. My background is in psychiatric nursing, where reading what happens between people was the daily work.