Search Results for « Causal Attribution »

<div>      In Kadrey v. Meta Platforms, the court measured the developer's asserted public benefits against an enforcement-surviving world, one in which training continues through licenses or lawful inputs, and measured the authors' asserted dilution harm against a world in which the models exist but were never trained on copyrighted books. Either comparison might be justified. What requires explanation is switching between them without identifying the temporal horizon that makes each comparison appropriate, and no rule in the opinion explains the switch. This Essay supplies a selection rule: a consequence counts in fair use only to the extent that granting or denying the defense would change it. The comparison is the most probable lawful response to enforcement that the record supports, whether authorization, substitution, omission, or abandonment, alone or in combination, and, for claims about legal rules rather than completed transactions, the probable distribution of responses among similarly situated actors. At the same causal unit and temporal horizon, the same comparison governs every asserted public and private consequence.  </div> <div>       The Essay develops construction rules for the comparison world, including a deliberately thin candidacy test for licensed worlds that keeps price and restrictive conditions outside candidacy and measures their effects inside the comparison, and requires courts not to treat as exogenous conditions that the challenged act itself probably produced. An available channel does not make authorization probable: record-supported refusal at the operative unit can shift selection to substitution, omission, or abandonment, strengthening both the developer's claim that no bargain was probable and the owner's claim that enforcement would have changed the capability and its harms. The Essay reads the principal fair-use cases diagnostically, showing that courts perform causal attribution episodically, without a selection rule, and often on only one side of the ledger. It then audits the asserted public harms of generative AI conditionally. Where authorization would preserve corpus, capability, scale, and deployment, deployment harms largely wash out, and the surviving disputes concern payment, consent, documentation, entry, and concentration. Where no corpus-sufficient channel was shown and lawful substitution or omission would have materially reduced capability, both the technology's benefits and its harms reenter at greater weight, and the case presents its real question: whether a socially valuable but harmful transfer should be excused when no consensual transaction was available. A final Part separates liability from remedy. Courts cannot order permitted-but-paid treatment, but for completed training they can find infringement, award damages, and, where equity permits and continued deployment is not independently infringing, withhold technological unwinding. The attribution analysis must be rebuilt for each remedy requested. The rule does not decide fair use by itself. It makes the choice of counterfactual visible, symmetric, and contestable. </div>

Feature attribution methods based on Shapley values have become a standard tool for explaining predictions of complex learned models, yet they conflate direct causal effects with spurious associations inherited through confounding variables. A feature can receive substantial attribution simply because it carries information about a common cause of the target, even when no direct causal mechanism links the two. We propose TC-SHAP (Targeted Conditional Shapley), a post-hoc framework that separates the causal component of feature attribution from the confounded component by comparing unconditional Shapley values with Shapley values computed under a forced conditioning set that blocks back-door paths. The normalized difference yields a trustworthiness score ρ that is scale-invariant, learner-agnostic, and interpretable as the fraction of attribution attributable to confounding rather than direct causation. We establish the method’s theoretical grounding in conditional Shapley value theory and the back-door criterion, and evaluate it across four predictor families (linear regression, multilayer perceptron, gradient-boosted trees, and random forests), three noise distributions, linear and nonlinear structural equation models, and high-dimensional settings up to p = 50. Across 189 controlled synthetic configurations, TC-SHAP achieves AUCs of 0.95–1.00 for discriminating true causal edges from spurious associations, with near-perfect separation under a fixed threshold ρ = 0.45. On the Sachs single-cell protein-signaling benchmark, TC-SHAP consistently improves edge-ranking AUC across six causal discovery configurations by 0.14–0.31 over raw Shapley magnitude, and identifies high-precision edge subsets with up to 88.9% precision. The method operates in seconds per edge and scales gracefully to high-dimensional inputs.

This article extends the causal powers theory of causal processing (White, 1989) by dealing with the use of regularity (e.g. covariation) information under the terms of the theory. The paper begins by arguing that regularity information is not basic to causal attribution because it is neither necessary nor sufficient for causal attribution. The type of information preferred for causal attribution, and the way in which it is used, are determined primarily by the type of causal question implicity or explicitly posed, in accordance with the constraint to provide something that counts as an answer to the question posed. Thus, types of regularity information are preferred for some types of causal question, specifically those referring to multiple occasions, but are not preferred for those referring to single occasions. Regularity information may still be used when preferred types of information are not available, and for two purposes: (i) ascertaining whether some thing or person has a certain causal power, and (ii) judging the strength of a power or a liability, when the effect to be explained concerns a relation between a power of one thing or person and a liability of another.

In Maritime Autonomous Surface Ships (MASS), navigational decisions are shared among software developers, system integrators, shipowners and remote operations centres. If such a ship is involved in a collision, existing investigative and legal frameworks provide no common method for showing how much each actor's conduct contributed to the outcome. This working paper proposes the Autonomous Causal Attribution Matrix (ACAM), a conceptual framework that connects authenticated incident evidence, a structural causal model and interventional counterfactuals to Shapley-value attribution with explicit uncertainty. It also outlines the Autonomous Maritime Risk Ledger (AMRL), a federated infrastructure in which validated findings could accumulate over time. The paper argues that causal responsibility and legal liability are distinct but cannot be cleanly separated, because every quantified attribution rests on reference standards, outcome measures and sets of actors that carry normative content. ACAM is therefore built to make such choices explicit and testable, not to eliminate them. A hypothetical collision illustrates the method and shows that defensible structural choices can shift attributions far more than parametric uncertainty does. The paper then considers how such attributions might inform the assessment of causative potency in collision apportionment under the 1910 Collision Convention, recourse claims against actors other than ships and limitation of liability, and explains why they should not form part of marine safety investigations under the IMO Casualty Investigation Code. The framework has not been implemented or validated, and a validation pathway is proposed.

When an end-to-end automated driving system neither forms nor retains a recoverable, decision-specific semantic record, investigators cannot reconstruct the internal path from perception to control. This study develops condition verification to produce contestable causal findings from the records that remain. The study combines information-structure analysis; comparative mapping of five verification questions to regulated and manufacturer-held evidence in the United States, European Union, United Nations vehicle-regulation system, and China; and scenario analysis of two constructed Level 3 crashes. Evidence sources span regulated recorders, fleet incident and exposure records, and manufacturer training, validation, and boundary-management materials. Where no decision-specific semantic record is formed and retained, the internal account is eliminated rather than merely opaque; additional logging and interpretability tools cannot restore it. Each question maps to records either mandated where Level 3 systems are approved or held by manufacturers; the sole new duty is a training-coverage disclosure. Across two contrasting Level 3 crashes, the same instrument set establishes event-level system contribution in one and negates it in the other, alongside three typed systemic findings resolved as established, negated, or indeterminate. For incomplete Level 2 records, an indirect path permits a rebuttable inference only when event anchors, a matched fleet propensity, no sufficient alternative account, and operator-controlled record gaps subject to retention duties coincide. Where burden-shifting instruments reallocate a proof object that neither party can establish, condition verification changes the object itself: from the unrecorded internal path to verifiable external conditions. Its indirect path gives the malfunction doctrine a version-indexed, exposure-normalized form for continuously updated fleets. Investigators can apply the protocol as a checklist linked to identified records and custodians. Insurers can incorporate its typed findings into claims analysis. Courts and regulators receive structured predicates for defect, negligence, causation, and apportionment, not final determinations.

Abstract Unlike conventional epidemiological studies that use observational data to estimate “associations” between risk factors and disease, the science of causal inference has identified situations where causal estimates can be made from observational data, using results such as the “backdoor criteria”. Here these results are combined with established epidemiological methods, to calculate simple population attribution fractions that estimate the causal influence of risk factors on disease incidence, and can be estimated using conventional proportional hazards methods. A counterfactual argument gives an attribution fraction for individuals. Causally meaningful attribution fractions cannot be constructed for all risk factors or confounders, but they can for the important established risk factors of smoking and body mass index (BMI). Using the new results, the causal attribution of smoking and BMI to the incidence of 226 diseases in the UK Biobank are estimated, and summarised in terms of disease chapters from the International Classification of Diseases (ICD-10). The diseases most strongly attributed to smoking and BMI are identified, finding 11 with attribution fractions greater than 0.5, and a small number with protective associations. The results provide new tools to quantify the causal influence of risk factors such as smoking and BMI on disease, and survey the causal influence of smoking and BMI on the landscape of disease incidence in the UK Biobank population.