Imagine you see a stone strike a nut precisely—the shell cracks open. If a chimpanzee did it, you'd say "clever, tool use!" If the wind knocked the stone loose, you'd call it coincidence. The same event, different attributions—the key is judging "intent." The paper discussed here teaches AI exactly this: inferring whether an observed outcome was driven by a purpose.
Key points
- The paper extends the classic causal ladder (Pearl, 2009): association, intervention, counterfactual—adding a fourth level, teleology: asking "who wanted this outcome?"
- It introduces an Intentional Intervention operator: unlike the surgical do-operator (exogenous, imposed by experimenters), intentional interventions are endogenous, performed by a goal-driven agent observing the system.
- The intervention yields a twin model, the Structural Final Model (SFM): observed values are the results of intentional intervention; counterfactuals answer "what if the agent had NOT intervened?" Comparing the two reveals the agent's existence and intent.
- Cybersecurity: build a no-attack counterfactual baseline; if anomalous traffic shows goal-directed optimization (data theft, service disruption), classify it as an attack.
- Scientific research: compare real data against an "unbiased ideal experiment" counterfactual to detect confirmation bias.
- Animal behavior: is a crow placing nuts on roads tool use? Test whether success rates and spatial-temporal patterns exceed chance and show purposeful adjustment.
- SETI: compare signal compressibility/entropy with natural processes; efficient encoding supports an intelligent origin.
- Mechanism vs. teleology: teleological inference holds that some causal patterns require purpose-based explanations, and agent existence is a falsifiable empirical hypothesis.
- Objective free-will detection: free will could be quantified as purposeful interventions generated internally, measured by divergence between actual and deterministically-predicted trajectories.
- AI consciousness detection: if an AI shows goal-directed optimization not fully explained by its training objective, does that imply emergent intent?
- Exact counterfactuals require the full causal graph and noise distributions—approximations (e.g., variational inference) are needed.
- Distinguishing purpose from adaptation (e.g., evolved traits look purposeful): true purpose involves representing and planning future states.
- Multiple interacting agents: extend SFM with a game-theoretic framework.
- Short term: automated agent detectors for financial fraud, medical diagnosis, and autonomous driving intent prediction.
- Mid term: explainable AI that states "I did this to achieve X; otherwise Y would result."
- Long term: teleological reasoning as a key AGI component—understanding others' intentions, forming goals, and cooperating socially.
- Title: Teleological Inference in Structural Causal Models via Intentional Interventions
- Authors: Dario Compagno, Fabio Massimo Zennaro
- Institutions: University of Paris Nanterre; University of Bergen
- arXiv: 2603.18968
- Posted: March 20, 2026
Agent detection algorithm
1. Observe anomalies: statistically improbable events, entropy reduction, deviations from a baseline causal model. 2. Build counterfactuals: via SFM—"what would the system look like with no purposeful intervention?" 3. Compute intent probability: compare P(outcome | no intervention) vs. P(outcome | intervention). A significant gap supports the agent hypothesis. 4. Infer the specific intent: analyze which variables were intervened on, timing/order, and utility, to reverse-engineer the agent's goal function.
Applications
Philosophical implications
Challenges and limitations
Outlook
As SCM originally answered causal questions, the authors show it can also answer teleological ones—about goal-directed, state-aware agents intervening in causal systems. This may mark the boundary between "intelligent tools" and "intelligent partners."
---
Paper info: