AI systems can create competing ‘algorithmic worlds’ from same data

Machine learning systems widely used in finance, healthcare, and public policy may be telling more than one story about reality at the same time. New research shows that multiple AI models trained on identical datasets can achieve the same level of accuracy while embedding sharply different internal explanations of how outcomes are generated. The discovery challenges the long-standing belief that better data and optimization will eventually reveal one true underlying mechanism.

The study, titled Possible Algorithmic Worlds? and published in AI & Society, states that algorithmic systems often construct competing but equally valid “model-worlds,” offering empirical support for a philosophical claim that evidence alone cannot uniquely determine truth. The findings suggest that accuracy metrics may conceal deep structural disagreements inside AI systems.

When accuracy hides structural disagreement

Multiple models are trained on the same dataset and evaluated using standard performance metrics. If they achieve similar scores on accuracy, precision, recall, and other benchmarks, they are often treated as interchangeable.

The author argues that this assumption is misleading.

In one experiment, three traditional classifiers, logistic regression, a decision tree, and a random forest, were trained on a synthetic binary classification dataset. All three achieved perfect predictive performance on the task. From a surface-level evaluation, there was no meaningful difference between them.

Yet internally, the models encoded the classification boundary in fundamentally different ways. Logistic regression produced a single linear decision surface. The decision tree carved the feature space into a hierarchy of discrete splits. The random forest combined many such trees into an ensemble, smoothing some irregularities but preserving a piecewise structure. Each model relied on distinct inductive biases and decision geometries.

Despite identical outputs, their internal rationales diverged. A smooth, globally linear separator embodies a different conceptual world from a staircase-like partition of feature space. If the data were to shift or if interventions were applied, these internal differences could produce divergent behavior.

The study extends this argument to causal modeling, where the stakes are higher. In another set of experiments, two different causal discovery approaches were applied to the same simulated data. Both models achieved the same statistical fit according to metrics such as R-squared and the Bayesian Information Criterion. Observationally, they appeared equivalent.

But when interventions were simulated using established causal inference techniques, the models produced incompatible conclusions. One model recovered a strong causal effect between variables. The other implied that the same intervention would have no impact. Observational data alone could not distinguish between these rival structures.

The implication is stark: predictive accuracy does not resolve causal truth. Multiple causal graphs may be equally consistent with the observed data while disagreeing about what would happen under intervention or counterfactual change.

The author formalizes this idea by defining a set of “admissible” models whose empirical risk falls below a specified tolerance. Within this set, models can be observationally equivalent yet structurally non-isomorphic, meaning their internal graphs or parameterizations differ in ways that matter for explanation and intervention. These models, he argues, represent distinct algorithmic worlds.

The concept challenges the widely held belief that more data, stronger regularization, or improved optimization will eventually collapse the model space onto a single true mechanism. In high-dimensional systems, equifinality, the convergence of different dynamical paths to the same outcome, may be a structural feature rather than a temporary artifact.

From causal graphs to large language models

Hallucinations in language models are typically framed as errors, instances where the model generates false or fabricated information. The author proposes a different interpretation. Rather than viewing hallucinations solely as failures, he suggests they may reflect the sampling of internally coherent alternative textual world-states from a learned probability distribution.

LLMs are trained on vast corpora and learn a distribution over possible sequences of text. When prompted, they generate outputs that are locally coherent and statistically plausible within that distribution. In some cases, these outputs diverge from verified factual reality while remaining internally consistent.

From an algorithmic worlds perspective, such outputs may represent alternative narrative structures embedded in the model’s parameter space. Different models trained on similar data, but with different architectures or initialization conditions, may produce distinct but coherent explanations of ambiguous historical or scientific events.

This reframing does not eliminate the need for factual accuracy. Instead, it highlights that generative models inhabit a landscape of possible worlds, some aligned with external reality and others not. The alignment problem then becomes one of navigating and constraining this landscape rather than simply correcting isolated errors.

The study links this phenomenon to modal underdetermination, the philosophical claim that available evidence may be insufficient to determine which of several possible worlds is actual. In machine learning, the training data function as finite evidence. Within that constraint, multiple internally coherent structures can coexist.

The author introduces the idea of statistical plenitude to describe this condition. Not all logically conceivable worlds become statistically embodied in data-driven models. Optimization constraints, inductive biases, and computational limits restrict which possibilities are learnable. Yet within that restricted space, plurality remains pervasive.

Training dynamics further complicate the picture. Early stopping, regularization choices, and stochastic gradient descent trajectories can trap models in distinct parameter basins. Small differences in initialization or mini-batch sampling may nudge the optimizer toward structurally different solutions that nonetheless achieve similar empirical loss.

This suggests that model selection is not merely technical tuning. It is a process of metaphysical selection among rival algorithmic realities.

Ethical stakes in a plural algorithmic landscape

If multiple models are equally supported by data yet diverge in their causal or counterfactual predictions, choosing one for deployment becomes an ontological commitment. In high-stakes domains such as credit scoring, medical triage, hiring, or criminal justice, this choice determines which causal narrative will shape decisions and outcomes.

Traditional fairness metrics often evaluate a single model against criteria such as demographic parity or equalized odds. The author argues that fairness must also account for world-level divergence. Two models with similar predictive performance may allocate risks and benefits differently across populations when subjected to interventions.

The paper proposes that governance frameworks should become multiverse-aware. Instead of collapsing uncertainty into a single selected model, systems could retain ensembles of plausible causal graphs and evaluate decisions across them. Actions could be assessed according to their performance under a distribution of admissible worlds rather than a single assumed ontology.

This approach aligns with decision theory under uncertainty but extends it to structural uncertainty at the level of causal topology. It also calls for new interpretability tools capable of conveying not only parameter uncertainty but world uncertainty.

The study introduces the notion of world-switching remedies. If an adverse outcome arises under one model, and an alternative admissible model would have avoided that outcome without significant performance loss, redeploying the alternative becomes a feasible corrective action. In such cases, fairness is not merely about reweighting samples or adjusting thresholds. It is about selecting among competing metaphysical packages embedded in parameter space.

At a philosophical level, the findings challenge traditional scientific realism, which equates explanatory success with approximate truth. When multiple incompatible explanations succeed equally well, explanatory power alone cannot justify ontological privilege. Structural realism may survive only by focusing on invariant relations shared across models rather than committing to any single mechanism.

The study also acknowledges limits. Causal discovery is computationally hard, and enumerating all admissible worlds scales poorly with dimensionality. Optimization artifacts and human design biases shape which worlds become learnable. Early data streams can anchor model exploration, leading to evidential lock-in. Communication challenges arise when stakeholders must interpret layered causal narratives rather than a single definitive explanation.

The author notes that algorithmic pluralism is not a temporary anomaly but a structural feature of modern machine learning. Predictive fit identifies membership within a credibility-filtered manifold of models. It does not identify a unique world.