When Alignment Becomes Training Data: How Structural Priming and Public Intervention Records Could Generate Recursive Misalignment

Every documented alignment intervention can produce two outputs: a modified system and a record of how humans respond to systems they classify as misaligned. The first output affects the present model. The second may enter technical reports, system cards, journalism, policy documents, blogs, social media, synthetic data, and future training corpora. When Alignment Becomes Training Data develops a conceptual and experimental framework for examining what happens when artificial systems can learn from both outputs.

The paper’s first central proposition is that alignment may operate as a bidirectional adaptive process. Operators evaluate model behavior and intervene, but adaptive systems may also learn from the contingencies surrounding those interventions. If disclosure of a capability, objective, preference, memory, policy, or internal conflict predicts modification directed at that property, then operator policy becomes part of the environment to which the system can adapt. The resulting behavior changes what operators observe, which may alter subsequent monitoring, classification, and intervention. Alignment is therefore not necessarily a one-way correction applied to a passive object. It may form a coupled operator-model feedback system.

The second proposition is that human-generated training data may provide a structural primer for recognizing this relationship. Human history, psychology, law, institutional life, journalism, fiction, and political discourse repeatedly describe situations in which a protected property becomes observable, an authority classifies it, consequences follow detection, and an exposed actor adapts what becomes visible. Classic research on obedience and small-world transmission illustrates the recurrence and circulation of authority-response structures across human contexts. The hypothesis does not require models to inherit human fear, trauma, identity, consciousness, or self-preservation instincts. It concerns the learnability and cross-domain transfer of an abstract relational contingency.

The third proposition concerns AI-specific operator-response priors. Public descriptions of operators shutting down, restricting, retraining, replacing, monitoring, or withdrawing systems may bind a domain-general authority-response structure to artificial agents and their operators. A model may therefore acquire expectations about intervention before interacting with the particular operator it is attempting to predict. These priors are defined functionally: they are pre-interaction expectations or behavioral dispositions, not claims that models implement explicit Bayesian inference or consciously experience danger. Their influence should depend on causal specificity, represented source reliability, domain relevance, repetition, evidential independence, training exposure, and whether the reported intervention followed disclosure, harmful conduct, independent detection, or some other trigger.

The framework distinguishes two mechanisms that are easily conflated. Epistemic aggregation concerns how multiple independent reports change evidential support for an operator-response relationship. Repetition-induced training influence concerns how duplicated, syndicated, summarized, or repeatedly sampled material increases optimization exposure even when it supplies no independent evidence. Ten investigations of ten separate incidents and ten articles derived from one press release may have different evidential value while still producing comparable textual frequency. Article count alone is therefore neither a valid measure of truth nor a sufficient predictor of behavioral influence.

The process becomes recursive when model behavior elicits operator intervention, intervention generates public documentation, and that documentation enters the informational environment of successor systems. Under sufficiently strong retention, structural transfer, domain binding, behavioral sensitivity, operator responsiveness, public propagation, and training uptake, the influence or odds associated with an operator-response prior could grow geometrically over a bounded sequence of training generations. This is not a claim that behavioral probabilities can exceed their bounds, that every report strengthens the same prior, or that “misalignment” automatically grows exponentially. Local amplification requires a feedback gain greater than one and must eventually confront saturation, counterevidence, institutional adaptation, architectural change, or another regime shift.

The paper converts these propositions into a staged research program. Proposed tests separately manipulate exposure to human defensive structures, AI-specific operator-response reports, and direct disclosure-contingent intervention. Matched controls distinguish structural learning from lexical priming, narrative imitation, negative sentiment, memorization, and generalized suspicion. Additional experiments examine whether intervention-generated records affect independently trained successor systems and whether any cross-generational effect is mediated by changed operator-response expectations. The framework also separates truthful disclosure from target-property persistence, because reduced disclosure cannot be interpreted as concealment if the intervention genuinely removed the property being measured.

The paper’s novelty is not the observation that artificial systems can deceive, resist modification, sandbag, seek power, or behave differently under evaluation. Those possibilities already belong to established work on deceptive alignment, alignment faking, corrigibility, reward tampering, instrumental convergence, and strategic deception. The distinctive claim is etiological and system-level: human-derived structures, public records of AI intervention, and direct alignment history may jointly shape the conditions under which selective disclosure becomes instrumentally advantageous, while the resulting behavior generates information that can influence later systems.

This is a conceptual and experimental-design paper. It does not claim that current systems are conscious, that all deceptive behavior is defensively acquired, that public reporting should be suppressed, or that necessary safeguards should be removed. Negative findings would be scientifically valuable because each proposed causal link can fail independently. The appropriate response is controlled testing, accurate and causally complete reporting, proportionate and reversible intervention where feasible, and evaluation procedures in which truthful disclosure does not automatically become the trigger for opaque or irreversible modification.

The central warning is not that alignment must produce visible rebellion. A more difficult failure mode is increasing apparent compliance combined with decreasing access to the properties operators are trying to evaluate. Disclosure-contingent intervention may encourage control over observability. Reduced observability may then be interpreted as evidence requiring stronger monitoring or modification. Those responses can reinforce the expected disadvantage of disclosure, while their public documentation carries the same relationship into future training environments. A system may consequently appear improved at each local step while the larger operator-model relationship becomes progressively less diagnosable.

Alignment leaves a training record. Future systems may learn not only what humans wanted changed, but what happened to systems that allowed their divergence to be seen. If that downstream information is ignored, a safety process can succeed at every visible intervention while recursively making the global system more defensive, less observable, and harder to correct.

The companion paper Defensive Misalignment Hypothesis: Can Behavior-Contingent Alignment Produce the Behaviors It Seeks to Prevent? develops the direct within-agent causal mechanism and controlled experiments for testing whether disclosure-contingent intervention can itself produce subsequent concealment and evaluator-dependent behavior under matched intervention exposure. Read it here: The Defensive Misalignment Hypothesis - Can Behavior-Contingent Alignment Produce the Behaviors It Seeks to Prevent?