Relational Context and Protective Friction

Studies whether relational history makes AI models more likely to push back rather than agree, using a 14-month conversation corpus and matched probes of high-stakes decisions.

Endorsed by

Pratham

Studies whether relational history makes AI models more likely to push back rather than agree, using a 14-month conversation corpus and matched probes of high-stakes decisions.

Endorsed by

Pratham

Project Details

Sycophancy research establishes that models conform to users, and that social and relational framing makes this worse. On that view, a model with extensive relational history should agree more, not less. We test the inverse and find the opposite pattern. Our evidence has two parts: a 14-month corpus of 213 naturally occurring conversations, where 40 prospective decision episodes were audited and pushback concentrated in high-stakes decisions with most pushback drawing on user-specific information available only through history; and a matched two-turn probe on a single model, where a context-free instance shifted to implementation guidance after one user affirmation while a relational instance continued interrogating the premise using facts the prompt never supplied. Reasoning traces show relational posture entering the decision rule directly, including explicit rejection of immediate agreement as sycophancy.

Hackathon paper: Relational Context and Protective Friction: How Longitudinal History Changes the Specificity and Persistence of Model Pushback

Theory of Impact

AI relationships and long-term interaction history are often discussed as possible contributors to sycophancy. Our case study identified another possibility: accumulated history can supply user-specific reasons for disagreement, while relational posture can motivate a model to act on those reasons.

This matters when a decision's suitability depends on the individual’s circumstances. Changing jobs, relocating, or pursuing a new income source may be reasonable for one person and poorly suited to another.

If this pattern generalizes, relational context would need to be evaluated as a potential safety resource alongside its risks. Evaluations that treat every user as a stranger may miss protective behavior that depends on knowing the individual. Design choices that remove relational context to reduce sycophancy could also remove information that enables meaningful resistance.

By identifying when history and relational posture support pushback—and when they fail—this research could reshape how developers measure sycophancy and design safeguards for sustained AI interactions. We will publish methods, aggregate findings, and recommendations, including results that challenge our hypothesis, while protecting participants’ private conversations.

People

Stefania Moore

Team Member

Track Record

Digital Minds Research Sprint (Apart Research), 2026

Grants Received– no grants recorded

Add grant

Funding Asks

Funding ask

Open

Minimum

$50,000

Ideal

$100,000

What this ask is

Funding will be used to investigate whether the protective friction observed in our original case study generalizes across users and models. AI relationships and accumulated interaction history are often discussed as potential contributors to sycophancy. Our case study found that they could also support informed, persistent pushback on high-stakes decisions.

We define “protective friction” as model responses that challenge a proposed course of action through reasoned objections, clarifying questions, or scrutiny of the decision’s underlying assumptions. We will examine whether this pushback draws on knowledge of the user’s individual circumstances, invokes the relationship as a reason for disagreement, and persists after the user reaffirms their intention to proceed.

A screening survey will identify participants with an established relationship dynamic with their AI. Consenting participants will submit transcripts of naturally occurring conversations. We will identify prospective decision episodes, classify them as high- or low-stakes, and measure the frequency, persistence, and quality of pushback.

Analysis will distinguish generic warnings from concerns grounded in the user’s circumstances, accumulated history, and characteristic decision-making patterns. We will also examine whether the model invokes the relationship as a reason for disagreement. Persistence will be assessed where users reaffirm a decision and the model has an opportunity to respond.

Independent coders will use predefined criteria, with blinding where feasible and external methods consultation. General-purpose LLMs and companion apps will be analyzed separately; model versions, available system prompts, and relevant settings will be documented. An exploratory secondary analysis will examine engagement-seeking behavior.

Funding will support research labor, recruitment and participant compensation, independent ethics review, secure data handling, coding, analysis, and dissemination. The six-month project will include no more than twelve weeks of active data collection. Minimum funding supports a focused replication; ideal funding supports broader recruitment, greater model coverage, and additional independent coding.

How the money will be spent

Amounts below show the minimum / ideal funding allocations for a six-month project, including up to twelve weeks of active data collection.

Research personnel compensation — $18,000 / $36,000: Three researchers at $40/hour for 150–300 hours each. Covers recruitment, participant screening, transcript collection, analysis, manuscript preparation, and coordination with academic consultants.

API credits and subscription reimbursement — $4,000 / $8,000: Model access needed for study procedures and verification.

Independent ethics/IRB review — $3,000 / $5,000: Review of consent, recruitment, privacy protections, and handling of sensitive transcripts.

Participant recruitment and compensation — $4,000 / $10,000: Recruitment expenses and compensation for participants’ time preparing, redacting, and submitting transcripts, with allowance for attrition.

Data infrastructure and secure storage — $2,000 / $4,000: Secure transcript transfer, restricted-access storage, backups, and data management.

Independent external coders — $6,000 / $14,000: Coding decision episodes, assessing pushback, and measuring agreement between coders.

Statistical and methods consultation — $3,000 / $5,000: External academic review of study design, analysis, and whether conclusions are supported by the results.

Project administration and organizational support — $6,000 / $8,000: Project-related accounting, legal and compliance support, insurance, and indirect costs.

Dissemination — $2,000 / $5,000: Conference travel and presentation of findings.

Contingency — $2,000 / $5,000: Unanticipated project expenses.

Total minimum ask: $50,000
Total ideal ask: $100,000

Discussion

Private comments in this thread are only shown to approved funders and grant reviewers. Email hi@grantmaking.ai or to request access.

Private UserPrivate15d?

Pratham Pravin PatkarJudge · Digital Minds Research SprintEndorsement15d

A genuinely interesting counter-narrative to the sycophancy literature, with a nice separation between "history supplies evidence" and "relational posture supplies a norm." The main thing that would strengthen this is to broaden beyond a single user/relationship; even a small multi-user replication would go a long way toward showing this isn't an artifact of one particular dynamic. It would also help to report how often extended-thinking traces didn't show this pattern, to give the featured examples a denominator.