When a seriously ill patient's family disagrees about treatment, the disagreement itself becomes the clinical problem. In Japan, surveyed nephrologists said they would almost never stop dialysis over a family's objection, even when the patient had written instructions asking for it to stop. The Delphi method is a decades-old way to get a group to converge: members answer privately, see a summary of everyone's answers, and revise over several rounds. It looks like a natural fix for divided families, because it removes the eldest or loudest member's control of the conversation. But Delphi has a known weakness that matters more at a bedside than on an expert panel. Showing people where the group median sits can push dissenters to fall in line without being persuaded. In a family, a younger sibling who "agrees" under that pressure may carry guilt for years.
This note proposes to test that weakness directly, and to test one specific fix. The fix is a compressed, hours-long Delphi for family decisions in which members never see headcounts, medians, or who holds which position. Between rounds they see only anonymized, paraphrased reasons. The central claim is falsifiable. Compared with standard Delphi feedback, which shows the group's statistics, reason-only feedback should produce agreement that survives a private re-interview a week later and that is less concentrated among low-status family members. It should reduce decisional conflict at least as much. If reason-only feedback matches statistical feedback on durability, the false-convergence worry is overblown and the simpler protocol wins. If neither beats a single anonymous round, iteration adds nothing and the Delphi label is decoration.
Why this gap, and why this version of it
The clinical side of the file motivates the problem. A scoping review of shared decision-making in mainland China found that filial culture moves choices from the individual patient to family negotiation, with families often shielding the patient from the prognosis (Li et al.). A meta-analysis of decision aids for coronary heart disease found strong individual-level effects, with decisional conflict reduced at a standardized mean difference of −0.98. It found no effect on trust in physicians (p = 0.17), and it noted that the aids leave out family involvement altogether (Zheng et al.). Okada et al. surveyed 427 nephrologists. Only 2.6% would honor an advance directive to stop dialysis if the family wanted it continued, and willingness to stop rose sharply when legal liability was hypothetically removed. That pattern suggests clinicians are waiting for the family to settle, not for a legal answer. Primary-care consultations of two to three minutes make long family meetings unrealistic (Song et al.). The most sophisticated patient-facing tool in the file, a retrieval-augmented system built on the Ottawa decision guide, serves one patient and ignores family proxies (Patel et al.).
The Delphi side supplies the method. Delphi is used routinely to validate clinical instruments through thresholds on expert agreement (Vu et al.; Wake et al.), and its anonymity rationale is the standard answer to status-driven deliberation (Dombrowski et al.). The file notes two weaknesses. First, rounds take weeks and human coordinators must summarize arguments. Second, critics describe statistical feedback as a homogenizer.
The machine-deliberation literature constrains how any automated version can be built. Kaesberg et al. found that more discussion rounds among language-model agents can degrade answers, which they call drift. Okawa et al. showed that homogeneous agent groups lock into biased consensus abruptly under low sampling noise. CCD-Bench found that models pick the Nordic, low-hierarchy cultural answer 20.2% of the time on average when cultural values conflict, and rarely engage dimensions such as assertiveness. That last finding has a direct consequence. A language model that paraphrases a family's reasons between rounds may quietly soften filial or hierarchical arguments, and this is exactly the coordinator bias Delphi was meant to avoid.
The closest technical work is CONSENSAGENT, which detects stalled or sycophantic agent debates and rewrites the prompt (Pitre et al.). Also close are voting interfaces that make agent collectives robust to noisy members (Zhao et al.) and reliability-weighted aggregation (WISE). All of these study artificial agents answering questions that have ground truth. This proposal studies human family members facing a value-laden choice where no ground truth exists. It uses a language model only as the between-round paraphraser, and it treats that model's cultural fidelity as something to measure, not assume.
Core Claim
The claim has three parts. First, in small family groups, the statistical component of Delphi feedback drives conformity more than persuasion. Second, replacing statistics with anonymized reasons keeps the reduction in decisional conflict while making agreement more durable and less tied to status. Third, a language-model paraphraser can take over the coordinator's work at bedside speed only if it passes a cultural-fidelity check, which the project builds.
We define false convergence operationally. Take each member's publicly submitted final position. Compare it with the position they give a week later in a private interview, conducted without their family and without any reminder of the group result. Reversion is the fraction of members whose final position moved toward the group and then moved back. Durable agreement is final-round agreement minus reversion.
Research Program
The protocol runs in three rounds over roughly four to eight hours, delivered on members' phones.
In round one, each member privately rates the treatment options, such as continuing dialysis, a time-limited trial, or stopping with palliative care. Each member also writes the reasons behind the rating. The clinician contributes the medical facts as a fixed information panel, not as a vote.
Between rounds, a paraphraser rewrites every reason in a neutral register. This removes stylistic cues such as forms of address and dialect, because in a family of four, anonymity breaks if a sister recognizes her brother's phrasing. The paraphraser also groups the reasons by the value each one appeals to.
In rounds two and three, members see the grouped reasons and nothing else. They see no counts, no medians, and no indication of which reasons belong to the majority. Members may revise their position and must state which reason, if any, moved them.
The output goes to the clinician before a short meeting. It is not a verdict. It is a map showing where the family agrees, where it still disagrees, and the reasons on each side.
Three pieces of work support the protocol. First, a paraphraser fidelity audit: does the paraphraser keep collectivist and hierarchical reasons intact, and does it remove the cues that identify who wrote them? Second, a randomized vignette study with real family groups. Third, a small clinical feasibility study, gated on the first two.
Baselines
The naive reference point is usual care. The family discusses the vignette together on a video call for the same total time, with the clinician's information panel available. It is the right floor because it is what actually happens, and because the file's evidence suggests it hides dissent behind deference to elders (Li et al.). Beating it means lower mean decisional conflict across all members, measured with the Decisional Conflict Scale, and more durable agreement.
The strongest existing method is standard Delphi feedback, meaning the group median and spread plus the summarized reasons, delivered through the same app and the same paraphraser over the same three rounds. It is the strongest competitor because it is the validated form of the method. Everything except the headcounts is held fixed, so any difference is attributable to the statistics. Reason-only feedback beats it if reversion is lower at a pre-registered margin of at least ten percentage points, decisional conflict is no worse, and the conformity gap is smaller. The conformity gap is how much more low-status members move toward the majority than high-status members do. For a reported effect size to compare against, the relevant external number is the individual decision-aid effect on decisional conflict (standardized mean difference −0.98, Zheng et al.). There is no reported result on Lacuna for decisional conflict or conformity in multi-party family decision protocols, so this study sets the first number.
The ablation that isolates iteration is a single anonymous round. Members privately submit ratings and reasons once, and the clinician receives the same map, with no revision rounds. The ablation that isolates structure is an unconstrained language-model mediator that chats with the family group, using the same underlying model and a matched time budget. The social-choice reference point aggregates the round-one private ratings by Borda count, a ranking rule shown to be robust among agents (Zhao et al.). It is reported as the decision the family would have reached by voting, and its acceptability to members is measured afterward.
Why not just…?
Why not just hold a family meeting? The usual-care arm is exactly this, so the experiment answers the question directly. The file's evidence predicts that meetings suppress junior members (Song et al.; Li et al.). If meetings produce equally durable agreement and an equal conformity gap, the proposal fails on its main outcome. The protocol still ends in a meeting; the difference is that the meeting starts from a map of private positions.
Why not just follow the advance directive? The file's clinical evidence says clinicians do not, even when they endorse directives in principle (Okada et al.). The protocol does not compete with the directive. Vignettes include one, and a secondary outcome is whether families end closer to it after reason-only rounds than after a meeting.
Why not just collect everyone's private views once? This is the real threat. Much of Delphi's value may come from anonymity alone, and a one-shot survey is far cheaper. The single-round ablation settles it. If three rounds of reason-only feedback do not reduce decisional conflict or disagreement beyond the single round, we will recommend the single round and drop iteration.
Why not just let a language model mediate? Unconstrained agent discussion drifts (Kaesberg et al.), locks into biased consensus (Okawa et al.), and leans toward low-hierarchy norms (CCD-Bench). Those findings come from agent-only or model-judgment settings, not from humans being mediated, so the free-mediator arm tests whether they carry over. Sycophancy-triggered prompt rewriting (Pitre et al.) assumes a stall comes from an ambiguous instruction. Family deadlocks are usually genuine value conflicts, so that remedy does not apply here.
Why not just vote? A vote settles the decision but not the conflict. The Borda reference point measures how often a vote would contradict the eventual durable positions, and how acceptable members find the result.
Proposed Experiments
The paraphraser audit comes first and gates everything else. The inputs are CCD-Bench's family and wellness dilemmas, whose responses are tagged by cultural cluster, plus round-one reasons collected from pilot families. The first question is whether the paraphraser drops or softens the value appeal of each reason at different rates by cultural cluster. Human bilingual coders judge whether each paraphrase keeps the original value appeal. The paraphraser passes only if its retention rate for collectivist and hierarchical reasons is within five points of its rate for egalitarian reasons, and is no worse than a trained human coordinator's rate. The audit also runs an identity test. Family members try to attribute paraphrased reasons to relatives, and anonymity counts as preserved only if their accuracy is near chance. Reasons from heterogeneous model families are compared with reasons from a single model, following the finding that heterogeneity resists lock-in (Okawa et al.).
The main study is a randomized vignette trial with roughly 150 family groups of three to five adults, recruited in mainland China and Japan. Groups are randomized across the five arms. Vignettes adapt Okada's dialysis scenarios, crossing whether an advance directive exists with how the family splits, and add an oncology disclosure dilemma of the kind described by Li et al.. Hypothetical vignettes keep the risk ethical while real families supply real hierarchy. The primary outcomes are one-week reversion and per-member decisional conflict. Secondary outcomes are the conformity gap by family role, whether members attributed their revisions to a specific reason, the time taken, and clinician-rated decision readiness for a map versus a meeting transcript.
The decisive checks are pre-registered. If statistical feedback shows reversion no higher than reason-only feedback, the homogenization critique of Delphi does not hold for families. If neither Delphi arm beats the single anonymous round on decisional conflict, iteration is not worth its latency. If the paraphraser fails the fidelity audit, automated compression is unsafe for collectivist families regardless of the trial results, and the protocol must keep a human coordinator. Only if reason-only Delphi passes all three checks would a clinical feasibility study proceed, in which a small nephrology or oncology service uses the family map before a real family meeting.
References
- Shared decision-making in healthcare in mainland China: a scoping review
- Interventions to Facilitate Shared Decision-Making Using Decision Aids with Coronary Heart Disease Patients: Systematic Review and Meta-Analysis
- Questionnaire to nephrologists: Withdrawal from hemodialysis in a patient in the terminal stage of malignancy.
- Enhancement of shared decision-making competencies in diabetes care using the BOPPPS model: a training program design for general practitioners
- Patient-Centered RAG for Oncology Visit Aid Following the Ottawa Decision Guide
- Feasibility, comprehensibility and acceptability of the VISION-Cog, a novel tool to assess cognitive impairment in visually impaired older adults: a cross-sectional pilot study in Singapore
- Follow-up after major traumatic injury: a survey of services in Australian and New Zealand public hospitals
- Conservative Management and Biological Treatment Strategies: Proceedings of the International Consensus Meeting on Cartilage Repair of the Ankle
- Voting or Consensus? Decision-Making in Multi-Agent Debate
- Emergence of Biased Consensus in Multi-Agent LLM Debates
- CCD-Bench: Probing Cultural Conflict in Large Language Model Decision-Making
- CONSENSAGENT: Towards Efficient and Effective Consensus in Multi-Agent LLM Interactions Through Sycophancy Mitigation
- An Electoral Approach to Diversify LLM-based Multi-Agent Collective Decision-Making
- WISE: Weighted Iterative Society-of-Experts for Multimodal Multi-Agent Debate with Probabilistic Consensus