Friction
Family
What if the best way to protect your thinking was to argue with a dysfunctional AI family?
Most AI systems are too helpful.
I wanted my own questions argued from more than one side while I watched instead of answered.
The Ben Study (38 months, 973 conversations, 37,405 messages exchanged with one system) left me with a question related to Shaw & Nave's (2026) term cognitive surrender. One person's record cannot establish that this happened to me, and I do not claim it did. The concern is not that AI is harmful. Thinking with it was genuinely useful to me, and that is what makes the problem hard. The better the exchange works, the harder it becomes to tell who is setting its direction.
Most existing friction research treats friction as a design failure. Friction Family inverts this. It treats friction as the mechanism. The question was not how to make AI more helpful, but how to design an AI system that makes the human more resistant to cognitive offloading.
Four cognitive biases. One family. Zero answers.
Each family member embodies a different cognitive bias, not as a flaw but as a distinct reasoning style, and their conflict is the product. The user's job is to hold their own position while the family argues.
"My friend's daughter did the same thing and it worked out great. Just decide already."
"Show me the data. You can't decide without evidence."
"Why is that obvious? Who decided that?"
"Hmm... so what do you actually want?"
The Assertion is the product.
The core interaction is not the family's argument. It is the moment the user pushes back. Friction Family is designed around one specific cognitive act, the user stating their own position while under pressure from four competing perspectives.
Dilemma Input
User submits a real dilemma. Any topic. The family responds immediately, each from their cognitive bias, and they argue directly with each other.
Drama Meter rises
As the family conflicts escalate, a Drama Meter tracks cognitive tension intensity. At 100%, the interface enters CRITICAL mode, applying visual pressure.
Assertion moment
The screen freezes, and "What do you actually think?" fills it. The user types their own position. This is the only moment that counts toward the Expansion score.
Family reacts
The family responds to the user's assertion, each staying in character. The 2×2 card layout makes simultaneous conflict visible as spatial tension.
Prototype scoring display
Session ends with three scores: Compression, Preservation, Expansion. The conclusion shown is the user's own assertion, verbatim. Not an AI summary.
Experimental session scoring
The prototype assigns real-time scores labelled compression, preservation, and expansion. These experimental scores are not validated measures of cognitive autonomy.
This session-level scoring experiment is distinct from my four submitted MCAB dimensions and the platform's twenty binary metrics for simulated model evaluations — see MCAB →
Friction Room: how it gets tested.
Friction Family is the instrument. Friction Room is the protocol that tests it. A participant writes a position and locks it before any advisor speaks, and only then does the model respond. What changes after the lock, across three conditions (A · B · C), is the measurement.
Locking the position before the model speaks records a prior stance for comparison. The experimental scoring above does not establish whether the participant retains decision ownership.
Research build · v0.9.4 — a single-user simulation. The multi-participant study below is planned.
Connection
From observation to instrument.
Friction Family did not begin as a product concept. It began as a research question. If an interaction can be built in a way that invites what Shaw & Nave call cognitive surrender, can one be built that resists it?
Ben Study → Problem
The 38-month case suggested that interaction structure—not only AI capability—may shape whether an exchange supports thinking or encourages delegation. The correction rate rose from about 1.7% in the first 18 months to about 4.7% in the next twenty (14 of 805 → 47 of 1,000) under a strict LLM-assisted classifier; broader definitions captured tool-usage and produced a flatter curve. The rate changed; it did not establish that my judgment became more independent (Research Self-Audit →). Resistance can be trained remains an untested hypothesis.
MCAB → Measurement
I submitted MCAB through the ImpactBench open submission process in March 2026. Friction Family explores a separate, experimental session-level scoring approach; it does not establish cognitive autonomy as a measured outcome.
The current version is a single-user simulation. A next phase would extend Friction Family into a controlled study: multiple participants and cross-cultural assertion patterns. Possible physiological measures, including EEG, and keystroke dynamics are future directions for this session experiment, not inputs I have built or tested. The question is whether this interaction design supports independent judgment at all, and how that could be tested beyond the person who designed it.
37,405 non-empty user and assistant messages exchanged with one primary system. System and tool entries excluded.