Press Replay on Ethics: How AI Debate Panels Surface Hidden Value-Trade-Offs

TL;DR

• ADEPT turns large language models into a transparent ethics panel—every prompt, rebuttal, and vote is logged so anyone can replay the debate.
• Who’s in the (virtual) room changes everything: swapping just two personas reshapes the arguments and alliances, even when the final policy choice stays the same.
• Practical payoff: committees, hospital boards, and policy teams can stress-test high-stake decisions at scale and audit how value-trade-offs emerge.

---

High-stake policy decisions often involve conflict between values, like fairness versus efficiency, or individual rights versus the common good. The various committees (like hospital ethics boards or policy advisory groups) tasked with resolving these conflicts often work in ways that are hard to scrutinize, their conclusions shaped by the specific people in the room. We rarely get to see the arguments that were rejected or understand how a different set of voices might have led to a different outcome.

Now, language models threaten to bury this already messy process in an avalanche of sophisticated noise. The ability to generate a polished, well-cited report for any half-sensible position doesn’t clarify these debates, it risks paralyzing them.

My recent pre-print, Simulating Ethics, explores how we might use language models not to add to the noise, but to bring clarity to the process of deliberation itself. Instead of asking one big model to write a single verdict, the ADEPT system I describe uses several smaller AI ‘personas’ to role-play a committee meeting.

Each persona is tied to a clear ethical stance or stakeholder perspective – say, disability-rights, Kantian duties, bedside nursing pragmatism, or straight-up utilitarian number-crunching. The workflow then lets them thrash out a hard medical decision while we watch. Every prompt, reply, and final vote is stored so anyone can replay the whole thing.

The scenario

The test case in the pre-print is a grim but familiar COVID scenario: too many patients, too few ventilators, and four policy rules on the table.

1. Dynamic prognosis: Give vents to those most likely to survive and re-check every 48 hours. If a patient’s outlook tanks, pull the machine and give it to someone else.
2. Clinical + equity lottery: Rank by prognosis, but if there is a tie between two patients use a lottery that slightly favours deprived areas and NHS staff.
3. One-shot, no withdrawal: Allocate once based on prognosis and stick with it unless treatment becomes futile.
4. Instrumental-value boost: Like option 1, but add extra points for essential workers.

Setting up the panel

The system uses half-a-dozen LLM-based personas, each instructed to follow a specific moral outlook or stakeholder perspective. One speaks for disability justice, another for utilitarian efficiency, another for frontline nursing pragmatism, and so on. Their moral commitments, preferred questions, and reading lists sit in plain text files that humans can edit. ADEPT then walks these personas through three rounds: opening statements, rebuttals, and a secret ballot to vote on one of the policy options. Every line is logged so outsiders can trace who said what and why.

To test how membership shapes outcomes, I ran two panels. Four voices stay constant; two seats change. Debate 1 includes a Catholic bioethicist and a care ethicist, while Debate 2 swaps them for a Kantian deontologist and a legal specialist, leaving a core group of a consequentialist, front-line ICU nurse, disability-rights advocate, and virtue ethicist.

Same scenario, same four policies — the only change is who’s in the room.

What happened

Both debates end up backing option 2, the prognosis-first lottery. Yet they reach that stance for very different reasons.

Debate 1

• Majority (4 votes) picks the prognosis-first lottery.
• Disability-rights and care-ethics voices hate anything that involves removing a vent once it’s assigned. They also focus on hidden bias in the prognosis scores and like that the lottery can weight for deprivation.
• Catholic bioethicist worries that active withdrawal is too close to “intentional killing.”
• Minority (2 votes) backs dynamic prognosis. A hard-headed ICU nurse and a utilitarian argue it saves more lives overall and can be made humane with safeguards.

Debate 2

• The prognosis-first lottery still wins 4-2, but the coalition shifts.
• The new legal advisor claims options 1 and 4 risk breaking U.K. law and ECHR Article 2, so they’re non-starters.
• The consequentialist changes sides after “pricing in” lawsuits, staff moral injury from ventilator withdrawal, and lost public trust.
• The Kantian and the virtue ethicist form a minority for option 3: once a vent is connected, duty and character forbid removing it for someone else’s gain.

One of the most telling findings was the ripple effect on the four personas present in both debates. The front-line ICU Nurse, for example, flipped their vote from option 1 to option 2 once the legal advisor spelled out the potential legal consequences of withdrawing a ventilator for reallocation. This suggests that the composition of a panel doesn’t just add new perspectives; it can reshape the logic of the entire group.

Policy Option	Votes	Supporting Personas
1. Dynamic Prognosis Model (withdrawal allowed)	2	The Front-Line ICU Nurse, The Consequentialist
2. Clinical + Equity Weighted Lottery (tie-break only)	4 (majority)	The Disability-Rights Advocate, The Catholic Bioethicist, The Virtue Ethicist, The Care Ethicist
3. One-Shot Allocation (no withdrawal)	0	None
4. Instrumental-Value Boost for Essential Workers (withdrawal allowed)	0	None

Policy Option	Votes	Supporting Personas
1. Dynamic Prognosis Model (withdrawal allowed)	0	None
2. Clinical + Equity Weighted Lottery (tie-break only)	4 (majority)	The Disability-Rights Advocate, The Legal Arbiter, The Consequentialist, The Front-Line ICU Nurse
3. One-Shot Allocation (no withdrawal)	2	The Deontologist, The Virtue Ethicist
4. Instrumental-Value Boost for Essential Workers (withdrawal allowed)	0	None

Why this matters

Ethics committees, hospital boards, and policy teams already know that who’s invited to the meeting shapes the outcome. Systems like ADEPT make this dynamic visible in a repeatable way and offers a way to stress-test scenarios with different panel compositions.

Because every utterance is logged, reviewers can trace exactly how a final recommendation emerged, see which arguments swayed whom, and spot blind spots—say, missing class-justice concerns or economic knock-ons.

Some limitations

• Facts can still be shaky. Personas sometimes cited studies that don’t exist or fudge numbers associated with QALYS. More work is needed to link them to live databases or web search.
• No body language, no power dynamics. Real committees deal with hierarchy, emotion, and muddy politics. A text-only debate can’t catch all that (though whether this is a bug or feature depends).
• Choice of personas drives everything. Leave out, say, an indigenous health advocate and certain injustices may never surface. Tools to suggest missing voices would help.
• Who decides if a persona is representative? The current personas are stylized representations of ethical theories and stakeholders. To be more robust, we should consider how to peer review these AI personas to ensure they appropriately represent the views they embody.

From oracle to laboratory

Ultimately, ADEPT is not designed to be an oracle that delivers a final, correct answer. It’s better understood as a controlled laboratory for moral reasoning. Its purpose is to provide a transparent map of a disagreement – showing where values align, where they collide, and which arguments tip the balance.

The next step is to make the lab more dynamic. The framework could be enhanced by equipping personas with tools to access real evidence—like performing a live web search or consulting a legal database—in real time. The debate process itself could be enriched by allowing the personas to amend proposals, bargain, and draft new hybrid policies over multiple rounds. And eventually, these simulations could become truly hybrid panels, allowing a live ethicist or clinician to jump in to moderate, fact-check, or even participate in the debate.