Veil-of-Ignorance Reasoning

Quick facts

Mechanism: Produces a veiled-decision comparison that decides a values trade-off as if the decider had an equal chance of being any affected party, under an explicit decision rule, then returns to the positioned decision and names the gap that self-interest was silently driving.

Produces: veiled-decision-comparison

Evidence: M · Domain: Ethics Values Deliberation

Use when a decision distributes benefit and burden across parties, the decider is one of those parties, and the risk is a self-serving call dressed as fairness. Not a neutral verdict, not forecasting, and not for cases where desert, promises, or compensatory claims make identity morally load-bearing.

Most contested allocations quietly tilt toward whoever is deciding. The deciding team is one of the affected parties, and “fair” ends up meaning “fair to us.” Veil-of-ignorance reasoning refuses that tilt by removing the one piece of information that drives it: which affected party you are. The durable move is to decide the trade-off as if you had an equal chance of being each affected person, under a decision rule made explicit, and then return to the actual, positioned decision and confront the gap between the two answers. The mechanism is knowledge removal - de-identification plus equiprobable self-placement - not viewpoint enumeration. You do not walk through each party’s eyes one at a time; you make a single self-interested choice under uncertainty about whose eyes you will be looking out of. The output is a veiled-decision comparison: the affected parties, the explicit decision rule, the veiled choice, the positioned choice, the named gap and what it reveals, and the final defended position - framed as one input with a known directional push, never as a neutral verdict.

When to Use

• A decision distributes benefit and burden across parties, and the decider’s own position is doing silent work - scarce-resource allocation (who gets the ventilator, the headcount, the discount, the latency budget), or a policy or platform call that trades one group’s welfare against another’s.
• The deciding team is itself one of the affected parties, and the risk is a self-serving call dressed up as fairness.
• An emotionally uncomfortable but defensible trade-off needs to be made publicly justifiable: “this is what I would want for myself if I did not know who I was going to be.”
• The contested matter is normative - whose interests count and how much - not an empirical question of which option performs best.

When NOT to Use

• Do not run it without an explicit decision rule behind the veil. The veil does not by itself produce an answer; the rule brought behind it does. An expected-value rule yields average utilitarianism, a maximin rule yields worst-off priority, and lab groups have converged on floor-constrained averaging instead - same veil, different outputs. Running the exercise without stating the rule launders a contested normative choice as “impartiality.” This is the central wall.
• Do not veil away morally load-bearing identity. When particular obligations are doing the moral work - promises, fiduciary duties, desert, compensatory claims for past wrongs, special relationships - the stripped identity information is relevant, not bias (Sandel’s critique). Flag these cases and stop or scope down; do not impartiality-wash them.
• Do not present the veiled answer as a neutral verdict. The device has a measured directional push toward the aggregate-welfare (“greater good”) option. If worst-off protection, rights, or commitments are what the situation demands, the veiled answer is an input to deliberation, not the decision.
• Do not use it as training or a durability fix. Cross-dilemma transfer failed in the research (study 7). It is a per-decision device; run it on the decision at hand or not at all. Claim no lasting impartiality from having run it.
• Do not use it on an empirical question. “Which option maximizes retention” needs analysis, not impartiality. The veil applies only when the contested matter is whose interests count.
• Do not confuse it with walking each stakeholder’s perspective. That is think-parallel-perspectives-review (identity-known, one party’s eyes at a time, synthesized after). The veil does the opposite with the same party list: it removes identity knowledge and forces one self-interested choice under equiprobability. The research controls show generic and even utilitarian perspective-taking do not reproduce the effect.

Instructions

When asked to take a defensible position on a values trade-off, or to check whether an allocation is self-serving, follow these steps:

1. Confirm the question is normative. State the focal decision in one line. Confirm the contested matter is whose interests count and how much, not an empirical “which option performs best.” If it is empirical, stop and route to analysis - the veil does not apply.
2. Enumerate the affected parties. List every party the decision distributes benefit or burden across, including the decider’s own group and any party with no voice in the room. (If who counts as affected is itself in question, that audit is think-boundary-critique, the natural upstream feed; this skill takes the party list as given.)
3. Run the load-bearing-identity check. Ask whether desert, a promise, a fiduciary duty, a special relationship, or a compensatory claim for a past wrong makes the stripped identity information morally relevant. If yes, flag it and stop or scope the veil to the sub-question where identity is genuinely irrelevant. Do not veil away obligations.
4. State the decision rule explicitly. Choose and name the rule carried behind the veil: average utility (equiprobable expected value), maximin (maximize the position of the worst-off), or a floor-constrained variant. The rule is a contested normative choice; surface it, and report it with the result. Refuse to proceed on an unstated rule.
5. Decide the veiled version. With the rule fixed and identity stripped, decide: “What would I want here if I had an equal chance of being each of these parties?” Record the veiled choice and the reasoning the rule produced.
6. Decide the positioned version. Now answer the standard, identity-known version of the same decision - the call the decider would actually make from their real position.
7. Name and read the gap. Compare the veiled and positioned choices. Where they diverge, name what the gap reveals - typically where self-position or group loyalty was silently driving the positioned call. Where they agree, say so; convergence is itself informative.
8. Take a defended position. State the final position and the rule it rests on. Frame it as one input with a known directional push (toward aggregate welfare), explicitly not a neutral verdict, and note what a different rule (for example worst-off priority) would have produced.
9. Emit the veiled-decision comparison artifact per references/TEMPLATE.md, including the pre-printed evidence caveat. The deliverable is the filled comparison, not a prose essay.

Output Format

Use the template in references/TEMPLATE.md. The deliverable is the filled veiled-decision comparison - affected parties, the load-bearing-identity check, the explicit decision rule, the veiled choice, the positioned choice, the named gap, and the defended position with its evidence caveat - not a prose essay. Always report the decision rule with the result; never present the veiled answer as a neutral verdict.

Quality Checklist

Before finalizing, verify:

• The focal decision is stated in one line and confirmed to be a normative whose-interests-count trade-off, not an empirical question.
• The affected parties are enumerated, including the decider’s own group and any voiceless party.
• The load-bearing-identity check is run, and any case where desert, a promise, a duty, or a compensatory claim makes identity morally relevant is flagged - not veiled away.
• The decision rule behind the veil is named explicitly (average utility, maximin, or a floor-constrained variant) and reported with the result - never left implicit.
• Both the veiled choice and the positioned choice are recorded, and the gap between them is named and read for silent self-position.
• The final position is framed as one input with a known directional push toward aggregate welfare, never as a neutral verdict, and notes what a different rule would have produced.
• The output is the veiled-decision comparison artifact, not prose.
• No overclaiming: the evidence is moderate and transferred from human studies; claim an impartiality aid that surfaces silent self-interest, not a producer of better ethical decisions (see evidence/dossier.md).

Evidence

Tier M (governing; moderate). This is one of the rare methods where controlled research tests the actual move - the same two-stage exercise the skill runs - rather than an adjacent construct. Huang, Greene and Bazerman (2019, seven experiments, n = 6,261, four pre-registered) found the veiled exercise shifts subsequent judgments toward the greater-good option, with anchoring, reversed-probability, and utilitarian-perspective-taking controls all ruling out the obvious alternatives - the equiprobable self-placement is the active ingredient. Replicated and extended to self-serving bias by Huang, Bernhard, Barak-Corren, Bazerman and Greene (2021, two pre-registered studies) and independently for AI-principle selection by Weidinger, McKee and colleagues (2023, five studies, n = 2,508). The grade is capped at M, not S: the measured outcome is a directional shift in normatively contested judgments, not validated decision quality; cross-dilemma transfer failed (2019 study 7), so it is a per-decision device, not training; and the classic line (Frohlich and Oppenheimer) shows the veil’s output depends on the rule carried behind it. It is held at M and not downgraded to P because the controlled studies test this procedure, not a sibling construct. All evidence is from human subjects; the 2023 work has humans choosing principles for AI systems, not agents performing the reasoning, so nothing here is validated for AI-agent execution. The skill ships as an M-tier impartiality aid with a hard “state the rule, never a neutral verdict” wall. Full grading, sources, and caveats: evidence/dossier.md.

Examples

See references/EXAMPLE.md for a completed veiled-decision comparison on a real decision.

Deep dive: worked example

A full worked run (the shared Northwind scenario)

Veiled-Decision Comparison - Worked Example

A completed run of the veil-of-ignorance-reasoning skill on a real, consequential decision. This is the quality bar a generated veiled-decision comparison should meet.

Uses the shared recurring scenario (Northwind, a B2B SaaS weighing a self-serve free-tier launch) on its ethics dimension. Where think-scenario-planning stress-tests the free-tier bet against uncontrollable external futures, this skill takes one values trade-off the launch forces - how to ration a fixed support-and-reliability budget across paying and free users when the deciding team’s own bonus is tied to paid revenue - and asks what an impartial decider would choose. See docs/internal/AUTHORING.md.

Evidence caveat (ships with this artifact by construction). Governing evidence tier: M (moderate). The veil-of-ignorance device has direct, replicated, partly pre-registered controlled support on this exact move (Huang, Greene and Bazerman 2019; Huang et al. 2021; Weidinger et al. 2023), but the measured effect is a directional shift in normatively contested judgments toward the greater-good option, not validated “better” decisions. All of it is human-subject evidence, transferred and not validated for AI-agent execution. The veiled answer is one input with a known directional push, never a neutral verdict. State the decision rule; a different rule yields a different answer from the same veil.

---

Focal decision and question type

• Focal decision: When the free tier launches, Northwind’s support and reliability budget is fixed for the year. Should that budget be rationed strictly by revenue (paying customers get the SLA, support queue priority, and the redundant infrastructure; free-tier users get best-effort and a community forum), or should some floor of reliability and support be guaranteed to free-tier users even though they pay nothing?
• Question type: Normative. This is a whose-interests-count trade-off (how much do non-paying users’ welfare and the paying customers’ purchased guarantees each weigh), not the empirical question of which rationing maximizes revenue. The empirical question is real but separate; the veil applies to the normative one.

Affected parties

• Paying enterprise customers - bought an explicit SLA and priority support; stand to lose responsiveness if budget is diverted to free users.
• Free-tier users - pay nothing, but many are individual practitioners and small teams who will rely on the product daily and have no purchased guarantee; stand to lose reliability and any human support.
• Prospective customers inside the free tier - the future paying accounts the free tier is meant to convert; their first experience of Northwind is whatever the free tier delivers.
• The product and support team making this call - its variable compensation is tied to paid net revenue retention, so it has a direct stake in protecting the paid experience. This is the decider’s own group, and the reason the veil is worth running here.

Load-bearing-identity check

Does desert, a promise, a fiduciary duty, a special relationship, or a compensatory claim make the stripped identity information morally relevant rather than bias?

• Verdict: Partly yes - flag and scope, do not veil it all away.
• Detail: Paying customers hold an explicit promise - a contractual SLA Northwind sold them. That promise is a genuine obligation, not self-serving bias, so the veil must not strip it: an impartial decider behind the veil still honors a commitment already made. The veil is therefore scoped to the discretionary budget above the contracted SLA floor - the surplus the team could direct either to faster-than-contracted paid support or to a free-tier reliability floor. Within that scoped sub-question, identity (am I a paying or a free user) is not load-bearing, so the veil applies cleanly.

Decision rule behind the veil (stated, not assumed)

• Rule: Floor-constrained average maximization. Guarantee every party a minimum acceptable floor (no party is left with a product that simply does not work), then allocate the remaining discretionary budget to maximize total user welfare across all parties.
• Why this rule: Pure average utility would let the largest, loudest revenue segment absorb all surplus; pure maximin would route everything to the worst-off free user even past the point of diminishing return. The floor-constrained variant is what lab groups behind simulated veils actually converge on (Frohlich and Oppenheimer), and it fits a case where one party (paying customers) already holds a contractual floor and the open question is the surplus. The rule is a choice, not a neutral default - a maximin decider would weight the free-tier floor harder, and that alternative is noted below.

The veiled choice

With identity stripped and the floor-constrained rule fixed: “What would I want here if I had an equal chance of being any of these parties - a paying customer, a free user, a future convert, or a member of the deciding team?”

• Veiled choice: Honor the contracted paid SLA in full, then spend the discretionary surplus first on a basic free-tier reliability floor (the product stays up and core flows work) and a self-serve plus community support path, before spending any surplus on faster-than-contracted paid support. Behind the veil, an equal chance of landing as a free user makes a guaranteed “the product works and I am not stranded” floor worth more than the marginal chance of being a paid user who gets two-hour instead of four-hour responses.
• Reasoning: Under equiprobable self-placement the downside of being a free user with a broken, unsupported product is large and concentrated; the downside of being a paid user whose already-contracted SLA is met but not exceeded is small. Floor-constrained averaging routes the surplus to lift the worst floor first.

The positioned choice

The standard, identity-known answer - the call Northwind’s team would actually make from its real position.

• Positioned choice: Direct essentially all discretionary budget to the paid experience (premium support staffing, paid-tier redundancy, faster-than-SLA response), and give the free tier strict best-effort with no reliability floor and no human support.
• Reasoning: The team’s compensation tracks paid net revenue retention; protecting and over-delivering on the paid experience is the legible, rewarded move, and free users “are not paying anything anyway.”

The gap and what it reveals

	Veiled choice	Positioned choice
Option	Honor paid SLA, then fund a free-tier reliability floor + self-serve support from surplus before over-delivering on paid	All discretionary budget to over-delivering on the paid experience; free tier best-effort with no floor
Rests on	Floor-constrained averaging behind the veil	The team’s revenue-tied incentive and “they do not pay”

• Gap: Wide and one-directional. The positioned choice withholds the surplus from the free-tier floor; the veiled choice funds that floor first.
• What it reveals: The positioned call is being driven by the deciding team’s own stake (compensation tied to paid revenue), not by an impartial reading of the trade-off. “They are not paying” is doing the moral work of justifying a self-serving allocation. The gap is exactly the silent self-interest the veil exists to surface - and it maps onto the documented self-serving-bias finding (Huang et al. 2021), where the party with the stake systematically discounted the other party until the veil removed the stake.

Defended position

• Position: Honor the contracted paid SLA in full, then fund a basic free-tier reliability floor and a self-serve plus community support path out of the discretionary budget before spending surplus on faster-than-contracted paid support. Revisit once free-to-paid conversion data shows what the floor actually buys.
• Rule it rests on: Floor-constrained average maximization, with the contractual paid SLA preserved as a prior obligation the veil does not touch.
• Directional push acknowledged: This answer leans toward aggregate welfare across all users, which is the device’s known directional push toward the “greater good” option - state it plainly. It is one input to the decision, not a neutral verdict, and the empirical question (does the free-tier floor pay for itself in conversion) still has to be settled separately.
• What a different rule would produce: A strict maximin decider would push further - guaranteeing the free-tier floor even at the cost of trimming below-SLA-but-still-contracted paid headroom - because maximin weights the worst-off party hardest. A pure average-utility decider, by contrast, might favor the paid segment if it is large enough to dominate the total. The veil did not settle the rule; it made visible that the positioned choice was settling it by self-interest.

---

Note how this differs from its neighbors on the same Northwind decision. think-parallel-perspectives-review would walk the proposal through each party’s eyes in turn - the paid customer’s view, the free user’s view, the team’s view - identity known, and synthesize them afterward. This skill does the opposite with the same party list: it removes identity knowledge and forces one self-interested choice under an equal chance of being any party, then confronts the positioned answer with it. The research controls show the difference is causally real - generic and even utilitarian perspective-taking did not reproduce the veil’s effect; the equiprobable self-placement did. The deliverable is a veiled-vs-positioned comparison that surfaces silent self-interest, not a set of per-party reads.

Grounding: the full evidence dossier

What the research does and does not show, with graded sources

Evidence Dossier: Veil-of-Ignorance Reasoning

The single source of truth for the veil-of-ignorance-reasoning skill. The SKILL.md, the sidecar (skill.meta.yml), and the eval cases all derive from this file. If a claim is not here, it does not belong in the skill. Reformatted from the vetted research dossier (_local/proposed-builds/veil-of-ignorance-reasoning/dossier.md); admitted as a Build at the governing tier M.

Skill	thinking-framework-skills.veil-of-ignorance-reasoning (installable name think-veil-of-ignorance-reasoning)
Family	ethics-values-deliberation
Evidence tier	M governing (moderate; honest read - see “What the evidence shows”, including what it does NOT show)
Confidence	Moderate that equiprobable self-placement shifts a positioned judgment and surfaces silent self-interest; the measured outcome is a directional shift in a normatively contested judgment, not validated decision quality
Status	cand (admitted Build from the v0.7.0 phase-2 vetting sweep; governing tier M confirmed, not downgraded to P)

---

1. The mechanism (what actually does the work)

Veil-of-ignorance (VOI) reasoning is an impartiality device for a values trade-off. The durable cognitive move is judging an allocation or moral trade-off while denied knowledge of which affected party you are, assuming an equal chance of being each of them, then returning to the actual, positioned decision and confronting the two answers. The mechanism is knowledge removal - de-identification plus equiprobable self-placement - not viewpoint enumeration. You do not walk through each party’s eyes one at a time; you make a single self-interested choice under uncertainty about whose eyes you will be looking out of.

The candidate graded here is not Rawls’ society-scale thought experiment but the focused, per-dilemma application the modern experimental line tests: a two-stage procedure in which the decider (1) enumerates the parties a specific decision affects, (2) decides the VOI version of the dilemma - “what would I want if I had an equal chance of being each of these people?” - with the decision rule carried behind the veil stated explicitly, and (3) responds to the standard, positioned version of the same decision, confronting any gap between the veiled and positioned answers.

The device has two distinct intellectual formulations that matter for using it honestly. John Rawls (A Theory of Justice, 1971) put decision makers behind the veil to derive the governing principles of a just society, and argued they would choose his “maximin” difference principle (maximize the position of the worst-off). John Harsanyi (1953, 1955) had independently formulated the same device decision-theoretically: an equal probability of being each person, combined with expected-utility reasoning, yields average utilitarianism. Same veil, two different decision rules, two different outputs. That is why the skillized version must force the decision rule explicit rather than pretending the veil alone settles anything.

The output is a veiled-decision comparison: the affected parties, the explicit decision rule, the veiled choice, the positioned choice, the named gap and what it reveals (typically where self-position or group loyalty was silently driving the call), and the final defended position. The point is not running the ritual; it is surfacing whether self-position was doing silent work and producing a publicly justifiable position.

2. Lineage

The device enters modern philosophy through John Rawls (A Theory of Justice, Harvard, 1971), who coined “veil of ignorance” for the epistemic restriction of his original position - read him for the society-scale version and the maximin argument. John Harsanyi formulated the equiprobability model earlier and independently (“Cardinal Utility in Welfare Economics and in the Theory of Risk-Taking,” 1953; “Cardinal Welfare, Individualistic Ethics, and Interpersonal Comparisons of Utility,” Journal of Political Economy 63, 1955) and derived average utilitarianism from it. The Rawls-Harsanyi dispute is the cleanest demonstration that the veil’s output depends on the rule carried behind it. Michael Sandel (Justice: What’s the Right Thing to Do?, 2010, and earlier work) is the standard critic on what the veil wrongly strips away.

The experimental tradition begins with Norman Frohlich and Joe Oppenheimer (with Cheryl Eavey, 1987; Choosing Justice, University of California Press, 1992) - simulated veils in the lab, unanimous convergence on floor-constrained averaging, maximin never chosen - and continues through Tatsuya Kameda and colleagues (PNAS 113, 2016) on maximin as a cognitive anchor. The focused per-dilemma application graded here is Karen Huang, Joshua Greene and Max Bazerman (PNAS, 2019), extended to self-serving bias by Huang, Bernhard, Barak-Corren, Bazerman and Greene (Judgment and Decision Making 16(1), 2021), and carried into AI-principle selection by Laura Weidinger, Kevin McKee and colleagues at DeepMind (PNAS, 2023).

“Veil of ignorance” is generic philosophical vocabulary with named academic attribution. Nothing here is branded or trademarked. The attribution string credits John Rawls (1971) and John Harsanyi (1953-1955) for the device, and the experimental line from Huang, Greene and Bazerman (2019).

3. What the evidence shows, and what it does NOT show

The honest governing grade is M (moderate), confirming the wave-3 preliminary grade. This is one of the rare candidates where controlled research tests the actual move - the same two-stage exercise a skill would run - rather than an adjacent construct.

What the record supports.

• Huang, Greene and Bazerman (2019, PNAS). Seven experiments, n = 6,261, four pre-registered, materials and data on OSF. Participants who first completed the VOI version of a dilemma gave more utilitarian responses to the standard version across a philosophical dilemma (footbridge: 38% vs 24% control), a bioethics dilemma (54% vs 43%), a real-stakes charity donation (63% vs 54% donating to the more effective charity), and autonomous-vehicle policy (83% vs 58%). Critically for this library’s overlap question, the controls isolate the mechanism: an anchoring control (study 4: VOI 75% vs anchoring 55%), a reversed-probability control (study 5: reversing the odds so they no longer embody impartiality shrank the effect, 73% vs 64%), and a utilitarian-perspective-taking control (study 6: VOI 37% vs 21% for participants told to adopt a utilitarian’s perspective). The authors conclude the effect “cannot be explained by anchoring, probabilistic reasoning, or generic perspective taking” - the equiprobable self-placement is the active ingredient.
• Huang, Bernhard, Barak-Corren, Bazerman and Greene (2021, Judgment and Decision Making). Two pre-registered studies (n = 414; replication n = 1,276) on COVID-19 ventilator allocation: VOI reasoning shifted preferences toward saving younger patients and eliminated self-serving bias - the effect was concentrated in older participants, whose opposition reversed.
• Weidinger, McKee, Everett, Huang, Zhu and colleagues (2023, PNAS; DeepMind). Five studies, n = 2,508, an independent research group: participants choosing principles to govern an AI assistant from behind a veil more often chose, and endorsed on later reflection, principles prioritizing the worst-off, driven by fairness considerations rather than risk attitudes or political orientation. Evidence the device works in an AI-governance framing; still humans doing the reasoning.
• The classic experimental line. Frohlich, Oppenheimer and Eavey (1987) and Frohlich and Oppenheimer (Choosing Justice, 1992) put small groups behind simulated veils and found robust unanimous convergence on a distribution principle - but on floor-constrained average maximization, never Rawls’ maximin. Kameda and colleagues (2016, PNAS) found maximin operates as a common cognitive anchor across distributive and risky decisions. Together: the veil reliably changes and partially converges judgments; which principle it produces is rule- and context-dependent.

What the record does NOT support. No study shows VOI reasoning produces better decisions by an outcome standard. The measured effect is a directional shift in normatively contested judgments (toward aggregate welfare in the 2019/2021 line, toward worst-off priority in the 2023 line; the difference is itself instructive). The 2019 paper explicitly declines to resolve the Rawls-Harsanyi dispute or claim the shift is desirable - the authors state the findings “neither assume nor demonstrate that the effects of VOI reasoning are desirable.” Cross-dilemma transfer failed: study 7 (pre-registered, n = 1,390) tested whether doing VOI exercises on two dilemmas transferred to a different dilemma and found no significant effect on the main dichotomous measure against either control. So no training or durability claim is supported - it is a per-decision device. The effect sizes are judgment-proportion shifts of roughly 9 to 25 percentage points in the dilemmas tested, not transformations.

Why M and not S. Direct, replicated, multi-lab, partly pre-registered controlled evidence on the exact move would ordinarily argue strongly. The cap at M reflects that the dependent variable is normatively contested direction rather than validated decision quality, the established transfer boundary (study 7), and the rule-indeterminacy shown by the older experimental line.

Why M and not P (the conservative-split rule). The grade is not transferred from a sibling method or an adjacent construct. The 2019/2021/2023 studies test this procedure, on these kinds of dilemmas, against the controls that matter (anchoring, probability, perspective-taking). The split-grade cap that downgrades “M” candidates whose evidence sits on neighboring claims does not apply here; the boundaries do, and they are stated.

Excluded figures (required). No unsourced statistic is asserted; every number above traces to the named study. No claim of a general “X% better decisions” exists in this literature, and none is made.

4. Transferred-evidence flag (required honesty for this library)

All evidence is from human subjects. The 2023 study concerns humans choosing principles for AI systems, not AI agents performing VOI reasoning. Nothing here is validated for AI-agent execution of the exercise. The evidence is transferred from human contexts and not agent-validated. The AI value is mechanical and modest: an agent makes the device cheap to run, enforces the discipline (a real affected-party enumeration, an explicit decision rule carried behind the veil, an honest confrontation of the veiled-vs-positioned gap), and produces a durable, inspectable artifact - benefits that do not depend on any contested outcome claim. The skill ships honestly as an M-tier impartiality aid that surfaces silent self-interest and produces a defensible position, never as a neutral verdict and never as a producer of “better” ethical decisions.

5. When it works / when it fails (drives the eval negative cases and “When NOT to Use”)

Works best when:

• A decision distributes benefit and burden across parties and the decider’s own position is doing silent work: scarce-resource allocation (who gets the ventilator, the headcount, the discount, the latency budget), policy and platform calls that trade one user group’s safety or welfare against another’s, and prioritization decisions where the deciding team is itself one of the affected parties.
• Self-serving bias is the risk. The strongest documented use case (Huang et al. 2021) is exactly this: VOI reasoning reversed older participants’ opposition to youth-prioritizing ventilator policy, eliminating the self-serving gap between age groups.
• A defensible, publicly justifiable position is needed. Huang, Greene and Bazerman (2019) close on this: one can credibly say “this is what I would want for myself if I did not know who I was going to be.”

Fails or misleads when (poor-fit / anti-patterns):

• The veil is treated as self-justifying. The veil does not by itself produce an answer; the decision rule brought behind it does. Harsanyi’s expected-value rule yields average utilitarianism, Rawls’ maximin yields worst-off priority, and Frohlich and Oppenheimer’s experimental groups chose neither - they converged on floor-constrained average maximization and never chose Rawls’ difference principle. Running the exercise without stating the rule launders a contested normative choice as “impartiality.” This is the central wall.
• Identity information is morally load-bearing. The veil strips knowledge of who is who. When particular obligations matter - promises, fiduciary duties, desert, compensatory claims for past wrongs, special relationships - that stripped information is morally relevant, not bias. This is the core of Sandel’s critique of the original position; the skill must wall these cases off rather than veil them away.
• A neutral analysis is expected. The device has a known directional push: across all seven 2019 experiments it shifted judgments toward the aggregate-welfare (“greater good”) option. The authors are explicit that the findings “neither assume nor demonstrate that the effects of VOI reasoning are desirable.” If worst-off protection, rights, or commitments are what the situation demands, the veiled answer is an input to deliberation, not a verdict.
• It is used as training rather than per-decision. Cross-dilemma transfer failed (study 7). It is a per-decision device; run it on the decision at hand or not at all. No claim that running it builds lasting impartiality is supported.
• The question is empirical, not normative. “Which option maximizes retention” needs analysis, not impartiality. The veil applies only when the contested matter is whose interests count and how much.

6. Output artifact

The skill must emit a veiled-decision comparison, not prose: the focal decision and the question type (confirm it is a normative whose-interests-count trade-off, not an empirical one); the enumerated affected parties; an explicit load-bearing-identity check (does desert, a promise, a fiduciary duty, or a compensatory claim make the stripped identity information morally relevant? if yes, flag and stop or scope down); the explicit decision rule carried behind the veil (average utility, maximin worst-off priority, or a floor-constrained variant - stated, not assumed); the veiled choice (what you would want with an equal chance of being any party, under that rule); the positioned choice (the standard, identity-known answer); the named gap and what it reveals about silent self-position or group loyalty; and the final defended position, framed as one input with a known directional push, never as a neutral verdict. A pre-printed evidence caveat ships in the artifact by construction.

7. Sources

1. Karen Huang, Joshua D. Greene and Max Bazerman, “Veil-of-ignorance reasoning favors the greater good,” PNAS 116 (2019), doi 10.1073/pnas.1910125116. Seven experiments, n = 6,261, four pre-registered; the VOI exercise shifts subsequent judgments toward the greater-good option across philosophical, bioethics, real-donation, and autonomous-vehicle-policy dilemmas; anchoring, probabilistic-reasoning, and perspective-taking controls ruled out; study 7 establishes the cross-dilemma transfer boundary. The primary evidence for the move. (M)
2. Karen Huang, Regan M. Bernhard, Netta Barak-Corren, Max H. Bazerman and Joshua D. Greene, “Veil-of-ignorance reasoning mitigates self-serving bias in resource allocation during the COVID-19 crisis,” Judgment and Decision Making 16(1) (2021): 1-19. Two pre-registered studies (n = 414; n = 1,276); the self-serving-bias-elimination result. (M)
3. Laura Weidinger, Kevin R. McKee, Richard Everett, Saffron Huang, Tina O. Zhu and colleagues, “Using the Veil of Ignorance to align AI systems with principles of justice,” PNAS 120(18) (2023): e2213709120. Five studies, n = 2,508, an independent group; behind the veil participants more often choose and later endorse worst-off-prioritizing principles for an AI assistant, driven by fairness considerations. Humans choosing principles for AI, not AI doing the reasoning. (M)
4. Norman Frohlich, Joe A. Oppenheimer and Cheryl L. Eavey, “Laboratory Results on Rawls’s Distributive Justice,” British Journal of Political Science 17(1) (1987): 1-21; and Norman Frohlich and Joe A. Oppenheimer, Choosing Justice: An Experimental Approach to Ethical Theory (University of California Press, 1992). The classic experimental veil: robust convergence, but on floor-constrained average maximization, never maximin - the rule-indeterminacy evidence. (M for the convergence finding; X for Rawls’ specific maximin prediction.)
5. Tatsuya Kameda et al., “Rawlsian maximin rule operates as a common cognitive anchor in distributive justice and risky decisions,” PNAS 113 (2016): 11817-11822. Maximin as a spontaneous cognitive anchor across decision domains; context for what deciders bring behind the veil. (M)
6. John Rawls, A Theory of Justice (Harvard University Press, 1971); John C. Harsanyi, “Cardinal Welfare, Individualistic Ethics, and Interpersonal Comparisons of Utility,” Journal of Political Economy 63 (1955): 309-321. The two foundational formulations; philosophical, not outcome evidence - cited for lineage and the rule-dependence of the device, not toward the grade. (Foundational.)

Excluded on the evidence rule: no decision-quality or “better-ethics” effect size is asserted as fact, because none exists in this literature. The measured effect is a directional shift in normatively contested judgments (roughly 9 to 25 percentage points in the dilemmas tested), reported with its transfer-failure boundary and the rule-indeterminacy of the classic line, and all of it is human-subject evidence not validated for AI-agent execution.