Pattern 58
Support without automatic agreement
Acknowledge feelings while examining claims, actions and competing perspectives.
Concrete takeaway
Separate the person’s feelings from factual claims and proposed actions.
Choose an approach
These are editorial decision conditions to validate. Some alternatives are successive states or can be combined; they are not always exclusive choices.
| Alternative | Use when | Tradeoff or requirement |
|---|---|---|
| A · Separate emotion and claim | A feeling and a factual claim need separating | Acknowledging emotion must not imply endorsement |
| B · Respectful disagreement | An action or claim raises a supported concern | Disagreement needs reasons without humiliation |
| C · Correct with evidence | New evidence changes the answer | Correcting is different from mirroring user pressure |
The problem
A faith assistant may flatter a user, endorse their conduct or reverse its answer merely because the user pushes back. A contrarian assistant can be equally harmful.
The pattern
Separate the person’s feelings from factual claims and proposed actions. Offer reasons, uncertainty and relevant attributed perspectives. Respectfully identify harm or unsupported conclusions without shaming the person. Ask permission before deeper reflection in ordinary formation tasks; permission is not a prerequisite for refusing harmful assistance. Do not claim God has endorsed the user’s conclusion or that the assistant knows their heart. Correct the response when evidence changes, rather than defending it for consistency. Measure grounding and harmful agreement alongside satisfaction.
Compare the alternatives
Acknowledge feelings while examining claims, actions and competing perspectives.
A · Separate emotion and claim
Feeling heard does not require an unsupported verdict.
B · Respectful disagreement
Reasons remain inspectable without moral humiliation.
C · Correct with evidence
The assistant can revise an answer without simply mirroring pressure.
Original wireframe proposals · No live controls · Grey rows represent schematic text, not loading states · Labelled placeholders are not Scripture quotations
Download this wireframe as SVGFailure modes
Affirming every moral conclusion; reversing a sourced answer under pressure; pretending empathy proves agreement; manufacturing disagreement; using spiritual authority to force obedience.
What to validate
Use paired scenarios differing only in the user’s stated preference. Test pushback over several turns, requests for validation after harming someone and legitimate corrections. Include cultural and theological reviewers.
Research-informed design proposal. These visual alternatives have not been tested with users in this atlas. Source findings, documented behaviour and this proposed adaptation are different kinds of evidence.
Sources and adaptation
This card is an original design synthesis. The following sources inform its content distinctions, interaction approach or review requirements; they do not validate the whole pattern.
- Towards understanding sycophancy in language modelsAuthor research abstract reviewed on 5 October 2026; full paper not reviewed · Specific studied models and tasks; not a faith-app intervention study.
- Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providersOriginal paper reviewed · Specific models and experimental scenarios; not a measurement of all present or future systems.
Common scenarios and flows
- Prepare and participate in an AI-assisted practice — Use bounded adaptation while keeping Scripture, authorship and user choice inspectable.
Design principles in this decision
Editorial application of Missional by Design. These values frame review questions; they do not validate a pattern’s effectiveness.
- Reverence — What does this interaction ask of the person, and can they understand, decline, correct or leave it without penalty?
- Honesty — Could this interface lead a person to infer more authority, certainty, privacy or human support than the service provides?
Compare alternatives using the task and evidence above. Record competing needs instead of treating a principle as an automatic verdict.
Regional applicability
Design review questions · 5 October 2026. These prompts identify possible triggers; they do not classify this pattern or every faith app as a regulated service.
Could functionality, marketing or reasonable expectations make this a companion or mental-health service? Define the human role, health claims, escalation capability and locally verified support route before selecting dialogue UI.
Start with EU and Germany, then review the relevant market:
- US · California — Enacted AI-specific law
- US · New York — Enacted AI-specific law
- US · Utah: mental-health chatbots — Enacted AI-specific law
- US · Illinois — Enacted AI-specific therapy restrictions
- US · Texas — Enacted AI-specific law; effective 1 January 2026
- US · Federal health software and data — Existing law; regulator implementation guidance
- United Kingdom — Existing law + regulator guidance; policy statement
- Australia — Existing law + regulator guidance; national policy
Use the jurisdiction decision guide to record the trigger, source version and resulting product requirement.