Every causal claim hides two quiet promises. First, that each thing we do has one version, a “treatment” that doesn’t keep changing its mind. Second, that one person’s outcome doesn’t depend on another’s treatment. Together these form SUTVA, the Stable Unit Treatment Value Assumption. They give causality its backbone. Without them, the experiment collapses into noise. The AI era breaks these promises with mathematical precision.
The end of causal isolation
In classic experiments, we could imagine flipping one switch at a time. Give a medicine to Patient A, and Patient B stays untouched. In an algorithmic system, that boundary doesn’t exist. The ad shown to A affects what B sees, because both are inside a shared feedback loop, a recommender, a ranking model, a social graph. Your “treatment” leaks.
When a learning algorithm updates itself based on everyone’s clicks, each action becomes part of the next treatment definition. We no longer have a single intervention; we have a living ecosystem of evolving exposures. Causality assumes separability. AI destroys separability.
The mirage of stable treatments
In classical inference, “the treatment” is a fixed thing, but in AI, every treatment is a moving target. An algorithm serving you content is different at 9 a.m. than it was at 8 a.m., because it has just learned from millions of others. A pricing model retrains nightly; a translation model improves hourly; a recommendation shifts per scroll. So there are not one but infinite versions of treatment. The idea that an outcome corresponds neatly to a single “assigned” version of the world becomes absurd. The treatment itself is learning. Causality assumes consistency. AI evolves consistency out of existence.
When invariance melts
The second foundational idea in causal inference, invariance, says that the relationship between cause and effect doesn’t change across contexts. If X causes Y, it should still do so when the weather, website, or month changes. That’s how we generalise. But AI systems rewrite their own rules. A self-optimising platform doesn’t preserve causal structure; it chases local optima. Its policies, user incentives, and environmental conditions shift together, creating moving equilibria. The world starts to “learn back.” The causal map redraws itself faster than we can trace it.
The reflexive era of causality
What used to be “interference” has become the default mode of the digital world. When every unit interacts, adapts, and co-learns, the unit itself loses meaning. We are dealing with systems where cause and effect are entangled across agents, algorithms, and time.
To survive this reflexive regime, we need new scaffolds:
- Dynamic structural causal models (DSCMs): Treat causal relations as functions that evolve with policy.
- Causal reinforcement learning: Model learning itself as part of the causal process.
- Reflexive SUTVA (RSUTE): Admit interaction, but formalise how far the ripple travels.
These aren’t academic tweaks. They are the only way to make sense of systems that rewrite their own data-generating processes.
Causality after equilibrium
AI doesn’t just automate prediction, it alters the world it predicts. Every optimisation loop closes the gap between model and environment until both are co-dependent. In this regime, causality is no longer about estimating static effects. It’s about tracking how interventions reshape the rules themselves. In the age of self-learning systems, the question “What causes what?” mutates into “What remains stable enough to call a cause?” That’s not a statistical problem. It’s a structural one.
