Artificial Counterfactual (ArCo) Overview

• ArCo is a framework of synthetic or hypothetical instances designed to flip model outcomes while ensuring validity, minimality, and plausibility.
• It leverages optimization and search techniques—ranging from ensemble methods to genetic algorithms—to generate actionable, human-interpretable recourse.
• Its applications span XAI, generative modeling, causal inference, and policy evaluation, emphasizing fairness, stability, and robustness.

to=container.exec code to=container.exec code to=container.exec code Artificial Counterfactual (ArCo) denotes a constructed counterfactual object used to represent an alternative outcome that is not directly observed. In the cited literature, the term covers synthetic or hypothetical instances that flip a model prediction, sequential recourse paths, paired tissue samples, counterfactual images, explicitly simulated counterfactual distributions, alternate planning worlds, and, in econometrics, an artificial counterfactual trajectory used for comparative policy evaluation (Small et al., 2023, Paulikat et al., 2023, Xia et al., 2022, Carneiro et al., 2020). Across these usages, the recurring desiderata are validity, minimality or proximity, actionability, plausibility, diversity, stability, discriminative power, robustness, fairness, and human interpretability (Guidotti et al., 2023, Sharma et al., 2019).

1. Definition, scope, and recurrent desiderata

In the XAI and algorithmic recourse literature, Artificial Counterfactuals (ArCo) are generated synthetic or hypothetical instances or paths that explain or realize alternative model outcomes, aiming for faithfulness to the data manifold, interventional plausibility, minimality or parsimony, and human-understandable and actionable recommendations (Small et al., 2023). This usage is motivated in part by the rarity of “native counterfactuals”: “A Few Good Counterfactuals” reports that nearest unlike neighbours with at most two feature differences are rare in most datasets, with “<1% with ≤2 feature differences” highlighted as a motivating observation for endogenous generation strategies (Smyth et al., 2021).

The same desiderata recur with different emphases across methods. “Ensemble of Counterfactual Explainers” enumerates minimality, actionability, stability, diversity, plausibility, discriminative power, and efficiency as target properties of a good counterfactual (Guidotti et al., 2023). CERTIFAI positions counterfactuals as a unified auditing instrument for robustness, transparency, interpretability, and fairness (Sharma et al., 2019). CountARFactuals centers plausibility, mixed-type support, sparsity, and computational efficiency (Dandl et al., 2024). LocalFACE reframes the object of interest from a single point to a feasible and actionable path on the data manifold (Small et al., 2023).

Usage	ArCo object	Representative works
XAI and recourse	Synthetic instance or path	(Small et al., 2023, Crupi et al., 2021)
Generative analysis	Artificial paired sample or edited image	(Paulikat et al., 2023, Pan et al., 2024)
Causal reasoning and fairness	Counterfactual distribution or alternate world	(Xia et al., 2022, Bynum et al., 2024, Raghavan et al., 14 Mar 2025)
Comparative policy evaluation	Artificial control trajectory	(Carneiro et al., 2020)

This range of meanings indicates that ArCo is not a single algorithmic family. A plausible implication is that the unifying feature is operationalization: each method constructs an analyzable proxy for an otherwise unobservable alternative.

2. Generation of explanatory and recourse-oriented ArCo

A major line of work treats ArCo generation as a search problem over valid, sparse, and plausible alternatives. In “Ensemble of Counterfactual Explainers,” Guidotti and Ruggieri propose ECE, an ensemble that combines three weak explainers—Brute Force Explainer, Tree-based Explainer, and Generative Sphere-based Explainer—through instance sampling, actionable-feature sampling, pooling of candidates, and a diversity-driven selection function (Guidotti et al., 2023). The selector optimizes

$$\underset{S \subseteq C, |S| \leq k}{\arg\max} \quad |\bigcup_{c \in S} \mathit{knn}_C(c)| - \lambda \sum_{c \in S} d(c, x),$$

with a cost-scaled greedy algorithm. Extensive experiments on tabular, image, and time-series data report that ECE routinely returns the full requested number $k$ of valid and actionable counterfactuals, improves diversity, plausibility, discriminative power, and stability, and remains model-agnostic; via autoencoder wrapping, it is also data-agnostic, albeit with the limitation that all latent features are treated as actionable (Guidotti et al., 2023).

LocalFACE moves from pointwise CFs to sequential algorithmic recourse. It is organized into three phases—explore, exploit, enhance—and uses locally acquired information to construct feasible and actionable counterfactual paths while preserving privacy by only leveraging data specifically required for the path construction (Small et al., 2023). In the exploit phase it builds a local graph $G=(V,E,W)$ between the factual $x$ and a counterfactual $x'$, and in the enhance phase it computes an optimal path by Dijkstra’s or another shortest-path algorithm, yielding a recourse matrix

$$x' = x + \sum_{i=1}^k z_i, \qquad Z = [z_1,\ldots,z_k].$$

The method explicitly addresses feasibility, actionability, fairness, robustness, privacy, and transparency through local density constraints and graph-based pathfinding (Small et al., 2023).

CEILS embeds SCM structure directly into counterfactual generation by optimizing in latent space rather than observed feature space (Crupi et al., 2021). With additive-noise equations

$X_v = f_v(\mathrm{pa}(X_v)) + U_v,$

CEILS performs abduction to infer $u^0$, intervenes in latent space to obtain $u^{0,cf}$ under actionability constraints, and predicts back to feature space as $x^{0,cf}=F(u^{0,cf})$. The method distinguishes immutable variables, mutable but non-actionable variables, and actionable variables, and it supports both soft and hard interventions. Its central claim is not merely causal consistency but feasibility: conventional feature-space generators can return explanations that are not reachable via allowed interventions, whereas CEILS constrains optimization to reachable latent actions (Crupi et al., 2021).

Model-agnostic density-based generation appears in CountARFactuals, which leverages adversarial random forests through FORDE for density estimation and FORGE for conditional generation (Dandl et al., 2024). The framework supplies two modes: an ARF-enhanced multi-objective optimizer and a standalone “ARF is all you need” generator. Plausibility is encoded directly through the learned joint density, mixed continuous and categorical features are handled natively, and sparsity is implemented by conditional sampling on a subset of features. The paper reports that ARF-based plausibility correlates more strongly with true plausibility than a kNN-based measure, with median $k$0 versus $k$1, while remaining computationally efficient (Dandl et al., 2024).

CERTIFAI represents a distinct search paradigm: a custom genetic algorithm generates counterfactuals by minimizing $k$2 subject to a prediction flip, with user-definable constraints on ranges, mutability, and categorical values (Sharma et al., 2019). The same generated counterfactuals are then reused for robustness auditing, interpretability, transparency, and fairness, and CERScore is introduced as a black-box model robustness score that aligns in ranking with CLEVER on ImageNet-scale experiments (Sharma et al., 2019).

3. Evaluation and the problem of human alignment

A central issue for ArCo is whether automated evaluation metrics track what humans regard as good explanations. “Do Metrics for Counterfactual Explanations Align with User Perception?” studies this directly on three tabular UCI datasets—Mushroom (8,124 instances, 22 features, 2 classes, F1 1.00, 1,625 test instances, 755 valid CFs), Obesity (2,111 instances, 16 features, 7 classes, F1 0.95, 423 test instances, 211 valid CFs), and Heart Disease (303 instances, 13 features, 2 classes, F1 0.85, 60 test instances, 25 valid CFs) (Liedeker et al., 16 Mar 2026).

Dataset	Summary	Valid CFs
Mushroom (MUS)	8,124 instances, 22 features, 2 classes, F1 1.00, 1,625 test inst.	755
Obesity (OBE)	2,111 instances, 16 features, 7 classes, F1 0.95, 423 test inst.	211
Heart Disease (HRT)	303 instances, 13 features, 2 classes, F1 0.85, 60 test inst.	25

Counterfactuals were generated with Counterfactuals Guided by Prototypes, and seven standard metrics were computed: sparsity, proximity, closeness, diversity, oracle score, trust score, and completeness (Liedeker et al., 16 Mar 2026). Proximity and completeness were defined as

$k$3

and

$k$4

To ensure coverage of the metric space, 85 CFs were selected by k-means clustering on metric values. A Prolific study with 167 participants presented 12 explanations per participant, with ratings on perceived accuracy, understandability, plausibility, sufficiency of detail, and user satisfaction. A Combined Quality Score (CQS) was defined as the mean of these five ratings; internal consistency was high with Cronbach’s $k$5, and the first principal component explained $k$6 of the variance (Liedeker et al., 16 Mar 2026).

The core finding is systematic misalignment. Across all datasets, only trust score had a statistically significant but small correlation with CQS, $k$7, $k$8; all other metrics had $k$9 (Liedeker et al., 16 Mar 2026). Dataset-level structure was stronger but inconsistent: in MUS, sparsity, diversity, proximity, and closeness correlated negatively with CQS and satisfaction in the range $G=(V,E,W)$0 to $G=(V,E,W)$1, indicating preference for fewer and smaller changes; in OBE, diversity, trust score, and completeness correlated positively with satisfaction, plausibility, and CQS in the range $G=(V,E,W)$2 to $G=(V,E,W)$3; in HRT, correlations were uniformly weak or non-significant (Liedeker et al., 16 Mar 2026).

Predictive modeling reinforced the same conclusion. A powerset analysis over all $G=(V,E,W)$4 metric subsets and five model classes—linear regression, kNN, Random Forest, XGBoost, and GAMs—showed that linear models performed worse than the mean-only baseline with mean $G=(V,E,W)$5, while Random Forests were the only nonlinear models to produce any gains, with mean $G=(V,E,W)$6 and best case approximately $G=(V,E,W)$7 (Liedeker et al., 16 Mar 2026). Performance tended to peak at three to four metrics and then decline. The article’s implication for ArCo evaluation is direct: benchmark reports that aggregate automated metrics cannot be assumed to proxy human assessment.

4. Generative, multimodal, and unsupervised ArCo

ArCo methods also appear as explicit generators of paired synthetic samples. CF-HistoGAN addresses highly multiplexed tissue imaging by translating a source-group HMTI sample into an artificial paired sample from a target outcome group while preserving raw-pixel structure as much as possible (Paulikat et al., 2023). The generator is a residual UNet,

$G=(V,E,W)$8

trained with a Wasserstein GAN objective plus an $G=(V,E,W)$9 sparsity penalty. Training and inference are patch-wise, outputs are stitched, and interpretation is performed through difference maps, Mean Channel Variation (MCV), Absolute Channel Variation (ACV), and artificially paired $x$0-tests. The paper reports statistically significant differences in protein expression with greater sensitivity than conventional unpaired analysis, precisely because the counterfactual sample preserves nearly identical structure (Paulikat et al., 2023).

“Counterfactual Image Editing” formalizes image editing through augmented structural causal models (ASCMs) and then proves two impossibility results: counterfactual editing is impossible from i.i.d. image samples and labels alone, and even with causal relationships between latent generative factors and images, no guarantees regarding the output of the model can be provided (Pan et al., 2024). The proposed relaxation is a family of counterfactual-consistent estimators that preserve user-specified care sets across factual and counterfactual worlds. This moves the problem from exact identification of $x$1 to bounded, care-set-consistent approximation (Pan et al., 2024).

In robotics, cGen uses adversarial training to generate small and realistic image modifications that induce desired effects in a frozen downstream controller or predictor (Smith et al., 2020). The architecture combines a generator with a classifier or predictor and optimizes a weighted sum of fidelity and task loss. It is demonstrated on MNIST and CelebFaces classification, a physical PR2 reaching task, and digital-twin navigation. The paper explicitly contrasts this with classical robustness notions based on noise rejection, arguing that semantic counterfactual modifications can characterize stability boundaries in perception–action space more expressively for some robotic applications (Smith et al., 2020).

ArCo has also been extended to unsupervised models. “Counterfactual Explanations for Clustering Models” introduces a model-agnostic technique based on a soft-scoring objective

$x$2

where $x$3 is feature sparsity, $x$4 is feature-space similarity, and $x$5 is a new soft similarity in clustering space (Spagnol et al., 2024). For k-means this soft score is centroid-distance based; for HDBSCAN it is based on soft cluster-membership probabilities; the model-agnostic variant uses prototypes and criticisms plus a semi-supervised self-training classifier. On five datasets, model-agnostic soft scoring consistently improved explanatory power over hard scoring, including Diabetes from $x$6 to $x$7 and Customer from $x$8 to $x$9 (Spagnol et al., 2024).

5. Causal semantics, fairness, and realizability

A second major axis of ArCo research concerns the semantics of counterfactuals themselves. “A New Paradigm for Counterfactual Reasoning in Fairness and Recourse” argues that interventional counterfactuals are ill-suited for many demographic variables because social categories such as race and gender are not modular in the required sense (Bynum et al., 2024). It proposes backtracking counterfactuals, which keep the causal mechanisms fixed and instead vary exogenous initial conditions through a backtracking conditional $x'$0. The central question shifts from “what would have happened had your race been different?” to “what would explain a counterfactual outcome for you as you actually are or could be?” (Bynum et al., 2024). The framework then defines counterfactual opportunity and counterfactual effort for individuals and groups without any intervention on protected attributes.

A more operational fairness line appears in “Counterfactual Fair Opportunity,” which introduces CFO as a counterfactual analogue of equal opportunity under fairness-under-unawareness (Cornacchia et al., 2023). The guiding criterion is that counterfactual samples of negatively predicted individuals, when adjusted to yield a positive outcome, should maintain the same sensitive behavior symmetrically across privileged and unprivileged groups. Two metrics are defined. Counterfactual Flips measures the proportion of successful counterfactuals whose sensitive attribute flips, while nDCCF weights this by rank so that closer counterfactuals dominate the score. On Adult with gender as the sensitive attribute, the paper reports CFlips around $x'$1–$x'$2 for males and $x'$3–$x'$4 for females, which it interprets as strong evidence that members of the unprivileged group often need to “become” privileged in counterfactual space to obtain favorable outcomes (Cornacchia et al., 2023).

At the level of formal estimation, “Neural Causal Models for Counterfactual Identification and Estimation” shows that $x'$5-constrained NCMs are expressive enough to match layer-3 counterfactual distributions of any SCM compatible with the causal graph and encode all equality constraints implied by the graph at $x'$6 (Xia et al., 2022). The NeuralID algorithm simultaneously identifies and estimates a counterfactual query by minimizing and maximizing the query subject to perfect fit to the available observational and interventional data; if the extrema coincide, the query is identifiable, otherwise it is not (Xia et al., 2022). A GAN-based implementation supplies a scalable practical route for high-dimensional settings.

“Counterfactual Realizability” addresses a different question: not whether a counterfactual is identifiable, but whether one can physically draw i.i.d. samples from a counterfactual distribution (Raghavan et al., 14 Mar 2025). It introduces the Fundamental Constraint of Experimentation—each unit can be subjected to a given causal mechanism at most once—and the CTF-REALIZE algorithm, which is complete for deciding whether an arbitrary counterfactual distribution is realizable given the causal graph and the action set (Raghavan et al., 14 Mar 2025). The associated graphical criterion states that realizability holds, with maximal counterfactual randomization, iff the counterfactual ancestors do not contain the same variable under different interventions. In the paper’s motivating fairness and reinforcement-learning settings, a counterfactual strategy is shown to provably dominate both interventional and observational strategies (Raghavan et al., 14 Mar 2025).

CausalARC extends the same concern to AI reasoning benchmarks. Each task is sampled from a fully specified SCM,

$x'$7

and can provide observational, interventional, and counterfactual few-shot demonstrations in context (Maasch et al., 3 Sep 2025). Because the exogenous context is held fixed while interventions alter the structural equations, the benchmark produces jointly observed factual and counterfactual examples unavailable in ordinary real-world data. This suggests a rigorous role for ArCo as benchmark infrastructure rather than only as explanation output.

6. Counterfactual worlds for planning and policy analysis

ArCo can also mean an explicitly constructed alternate world used for planning. “Counterfactual Planning in AGI Systems” defines a counterfactual planning agent that maximizes expected utility in a counterfactual planning world and then performs the same action in the real world (Holtman, 2021). The architecture is represented with a two-diagram notation separating the learning world from the planning world. By removing paths from certain variables to value in the planning world, the method constructs indifference to those variables. The paper uses this design to build an emergency stop button, a safety interlock that can stop the agent before an intelligence explosion, an input terminal through which humans can iteratively improve the reward function while suppressing manipulation incentives, and a counterfactual oracle (Holtman, 2021). The relevant notion of safety is therefore not explanatory fidelity but incentive shaping through engineered counterfactual structure.

A terminologically distinct but conceptually related usage appears in econometrics. “Lockdown effects in US states: an artificial counterfactual approach” applies ArCo as a variant of synthetic control, following Carvalho, Masini, and Medeiros, to estimate the short-run effect of lockdowns on cumulative COVID-19 cases and deaths (Carneiro et al., 2020). For each treated state, the method models log cumulative cases as a weighted combination of control states plus a log-linear trend, with weights estimated by weighted LASSO:

$x'$8

Unlike classic SCM, weights are not restricted to convex combinations, and regularization is designed for high-dimensional, possibly nonstationary panels (Carneiro et al., 2020).

The empirical design aligned states in epidemiological time, used 20 treated states and 6 controls, and evaluated post-lockdown trajectories for up to three weeks after the effective intervention date (Carneiro et al., 2020). Excluding New York, the mean counterfactual-to-actual ratio for accumulated cases was $x'$9 and the median was $$x' = x + \sum_{i=1}^k z_i, \qquad Z = [z_1,\ldots,z_k].$$0, while placebo controls yielded $$x' = x + \sum_{i=1}^k z_i, \qquad Z = [z_1,\ldots,z_k].$$1 and $$x' = x + \sum_{i=1}^k z_i, \qquad Z = [z_1,\ldots,z_k].$$2, respectively (Carneiro et al., 2020). In this formulation, the artificial counterfactual is not a synthetic explanation of a classifier’s decision but a synthetic untreated trajectory for a treated unit.

Taken together, these strands indicate that ArCo has evolved into a cross-disciplinary program for making unobserved alternatives operational. In XAI it is primarily about explanation and recourse; in generative modeling it becomes paired-sample construction; in causal inference it becomes a question of semantics, identification, and realizability; in planning it becomes incentive engineering; and in econometrics it becomes artificial control construction. The literature also shows that these constructions are not automatically aligned with human judgment, not always identifiable, and not always realizable, which is why recent work increasingly treats evaluation, semantics, and experimental accessibility as first-class components of ArCo research (Liedeker et al., 16 Mar 2026, Pan et al., 2024, Raghavan et al., 14 Mar 2025).