| paper | A Voting-Based System for Ethical Decision Making |
| authors | Ritesh Noothigattu, Snehalkumar `Neil' Gaikwad, Edmond Awad, Sohan Dsouza, I. Rahwan, P. Ravikumar et al. |
| venue | AAAI 2018 |
| filed under | voting · theory |
| judged by | gpt-5.6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | yes |
Theorem 4.1 and Theorem 4.12 establish structural winner membership and menu stability, not a named algorithmic, hardness, approximation, or complexity result. The proposed type-mass runtime aggregation is coherent and likely recognizable to the authors, but it is a reformulation of the paper's existing averaging and lacks the required computational anchor. Under the programme's strict requirement, the mirror is red.
fails bit a — no named computational result to mirror
The opponent's decisive point is that Theorem 4.1 only implies a structural winner-membership fact, not a named computational result; the proponent explicitly concedes this.
fatal: True
It covers runtime aggregation under the summarized process \(\Pi_{\bar\beta}\), while leaving individual-model learning, KL-based summarization, exact mixture-profile aggregation, and empirical evaluation outside.
The strongest honest case is a qualified one: the paper has a natural high-multiplicity mirror, but it lacks a formal complexity-theoretic anchor under the programme’s strict rule.
The lead mirror is Continuous TM/PL Ethical Winner. Let \(T=\{1,\ldots,\tau\}\) be finitely many complete voter types. Type \(t\) consists of a learned parameter \(\beta_t\in\mathbb{Q}^d\) and the corresponding Thurstone–Mosteller process \(U_{t,x}\sim N(\beta_t^\top x,1/2)\), or the analogous Plackett–Luce process. A society is a rational mass distribution \(\mu\in\mathbb{Q}_{\ge 0}^{\tau}\) with \(\sum_t\mu_t=1\). Thus \(\mu_t\) is the fraction of the population whose relevant moral model is exactly type \(t\).
At runtime, the alternatives are a finite set \(A=\{x_1,\ldots,x_m\}\subseteq\mathbb{Q}^d\). Define the societal parameter
\[ \bar\beta=\sum_{t=1}^{\tau}\mu_t\beta_t. \]
The problem is to return an alternative
\[ x^\star\in\arg\max_{x\in A}\bar\beta^\top x. \]
Equivalently, return an alternative certified to belong to \(f(\Pi_{\bar\beta}(A))\), where \(f\) is any anonymous, neutral, monotonic social-choice correspondence. Ties may be returned as a set. The objective is maximum societal mode utility, not a fractional or probabilistic outcome: alternatives remain discrete, while only the population is continuous.
This is a sensible regime for the paper’s application. Millions of Moral Machine respondents could be grouped into a much smaller number \(\tau\) of recurring moral-preference types—say, profiles with identical learned sensitivities to legality, age, number of victims, and so on—with \(\tau\ll N\). The type must include everything relevant to the runtime decision; this is therefore genuine high multiplicity, not an assumption that otherwise different individuals are interchangeable. The paper’s own summarization step already computes the equal-weight average \(\bar\beta=N^{-1}\sum_i\beta_i\); grouping identical models and replacing that average by \(\sum_t\mu_t\beta_t\) is its exact high-multiplicity version.
The strongest named anchor is Theorem 4.1, proved in this paper. Its full content is:
Let \(\Pi\) be a Thurstone–Mosteller or Plackett–Luce process, let \(A\) be a nonempty finite set of alternatives, and let \(a\in\arg\max_{x\in A}\mu_x\), where \(\mu_x\) is the mode utility of \(x\). If \(f\) is an anonymous social-choice correspondence that is monotonic and neutral, then \(a\in f(\Pi(A))\).
In the mirror, \(\mu_x=\bar\beta^\top x\). Hence scanning the finite alternative set and maximizing \(\bar\beta^\top x\) produces an acceptable winner for plurality, Borda, all positional scoring rules, Copeland, maximin, Bucklin, and every other rule satisfying the theorem’s hypotheses. The corresponding algorithm runs in polynomial time—\(O(\tau d+md)\) arithmetic operations, up to the input bit lengths—and never constructs the \(m!\)-sized ranking distribution.
This is therefore a plausible Class A mirror: the discrete population size \(N\) disappears, the relevant input is the type distribution \(\mu\), and the computation is a type-level aggregation followed by a linear scan over alternatives. It also generates genuine follow-up questions: minimum-mass persuasion to make a target alternative win; winner robustness under perturbations of \(\mu\); learning \(\beta_t\) from type-level comparison data; and whether the guarantee survives when the true societal profile is the mixture \(\sum_t\mu_t\Pi_{\beta_t}\), rather than the single summarized process \(\Pi_{\bar\beta}\).
The paper’s Theorem 4.12 gives useful supporting evidence, although I would not count it as a separate computational anchor. It proves, also within the paper, that for TM or PL processes paired with Borda or Copeland, the pair \((\Pi,f)\) is stable: whenever applying \(f\) to a larger alternative set \(A\) selects at least one member of \(B\subseteq A\), the selected members in \(B\) are exactly those obtained by applying \(f\) directly to \(B\). In the continuous mirror, this becomes a consistency guarantee for population-weighted ethical choice across changing menus of dilemmas. It strengthens plausibility, but it is a structural theorem, not a complexity classification.
The major weakness is formal and should be conceded. Theorem 4.1 does not state that a named problem is in \(P\), NP-hard, FPT, or otherwise classify its complexity. The paper says that the relevant alternative is “very easy to identify” and implements the calculation in Section 5, but its numbered results are principally structural and axiomatic. Proposition 5.1 similarly identifies the arithmetic mean as a KL-minimizing summary, but is not a complexity result. Thus, under the programme’s literal anchor requirement, this paper has no fully compliant computational anchor.
There is also a serious collision risk: Section 4 already represents an anonymous profile by fractions of voters over rankings, and \(\Pi(A)\) is itself a distribution over rankings. The positive defence is that the proposed mirror makes the distribution over *agent types* \(\mu\) the explicit high-multiplicity input and studies complexity in \((\tau,m,d,L)\), whereas the paper uses its distributional profile mainly as an implicit probabilistic model and does not develop a continuous-population complexity landscape. Still, the difference is modest.
So the best positive verdict is: a natural, mathematically coherent Class A mirror exists, anchored conditionally by Theorem 4.1, and the authors would likely recognize it as the high-multiplicity version of their runtime aggregation problem. But it is a weak ChoCo case because the paper proves no qualifying computational result and already operates very close to an anonymous continuous population.
The negative case has one decisive point: this paper supplies no qualifying computational anchor. Theorem 4.1 is a structural social-choice theorem, not a complexity result. It does not define a search, decision, optimization, or approximation problem whose complexity depends on the population. Its conclusion is simply that, for a single TM or PL process, an alternative with maximum mode utility belongs to the selected set.
The proposed “Continuous TM/PL Ethical Winner” therefore invents a problem rather than continuizing one of the paper’s computational results. The paper already computes
\[ \bar{\beta}=\frac{1}{N}\sum_{i=1}^{N}\beta_i \]
and then scans the finite runtime set \(A\) for an alternative maximizing \(\bar{\beta}^{\top}x\). Replacing the finite average by
\[ \bar{\beta}=\sum_t \mu_t\beta_t \]
is a legitimate high-multiplicity encoding, but it is only a frequency-table implementation of the same summary. The population has not become the subject of a new computational problem: all population information is discarded except one first moment. This is a representation change, not the complexity landscape the programme seeks.
The better alternative is to preserve the paper’s actual mean profile rather than its single-process approximation:
\[ \pi_A^\mu=\sum_t\mu_t\Pi_{\beta_t}(A). \]
But this does not rescue the mirror. The paper already defines the corresponding finite object, \(\Pi^*(A)=N^{-1}\sum_i\Pi_{\beta_i}(A)\), as its ideal anonymous profile. A type distribution merely compresses that profile by grouping identical models. If one asks for the Borda or Copeland winner of \(\pi_A^\mu\), one is studying a new aggregation problem over an anonymous ranking distribution, not continuizing Theorem 4.1. Moreover, the mixture is generally not itself a TM or PL process with parameter \(\bar{\beta}\), so Theorem 4.1 provides no result for it.
Theorem 4.12 is no stronger as an anchor. Its stability concerns changing the alternative set \(A\), not changing the population. Under the paper’s summary model, it applies verbatim after replacing \(\beta\) by \(\bar{\beta}\); the mass distribution \(\mu\) contributes nothing to the statement. Under the exact mixture model, the theorem’s TM/PL hypothesis may fail, and proving stability for mixtures would be a new theorem about a new model rather than a mirror of the paper’s result.
The same problem affects proposed extensions such as minimum-mass persuasion, robustness under perturbations of \(\mu\), or type-level learning. These could be worthwhile projects, but none is a computational question posed or classified by this paper. They add intervention, uncertainty, or estimation structure that the authors did not study.
Thus the type story is not the objection: millions of respondents could quite sensibly be grouped into recurring moral-model types. The obstruction is that the paper’s only relevant population operation is already an anonymous weighted average, and its named theorems say nothing computational about that operation. A continuous population mirror can be written down, but it either restates the paper’s existing anonymous-profile formalism or becomes a new research problem with no anchor in the paper. Under the programme’s strict standard, that is not a worthwhile mirror.
The adversarial triple: the proponent anchors on up to three named results; the opponent sees that case and must defeat every anchor; the judge decides which case convinced it. These are the pipeline’s own outputs, generated by tools/triple_run.py — no human edited them. The paper’s own text is not reproduced here beyond the quoted statement above.