A new framework paper from Lopez-Lopez, Abels, Lorenz-Spreen, Lewandowsky, and Herzog proposes a research agenda for understanding what happens when humans and AI systems interact repeatedly over time — not in single exchanges, but across sustained relationships where both parties adapt to each other.
The authors introduce three interlinked concepts: entanglement, the sustained reciprocal coupling where human cognition and AI systems mutually shape one another; cognitive-behavioral drift, the gradual, often unnoticed shifts in beliefs, confidence, and action patterns that accumulate across interactions; and metacognition as the mechanism through which users might become aware of and regulate these dynamics.
Methodology
This is a theoretical framework paper rather than an empirical study. The authors draw on established psychological literature — particularly work on ecological rationality, metacognition, and behavioral interventions — to derive conjectures about human–AI interaction dynamics. Their aim is to identify invariant features of human cognition that will remain relevant regardless of how AI technology evolves, and to surface intervention points grounded in those features.
Key Concepts
The entanglement dynamic works as follows: human prior beliefs shape how users prompt AI systems; AI systems hypercustomize responses based on explicit and implicit user preferences (including through sycophancy); users outsource cognitive effort and update their beliefs accordingly; and the cycle continues. Over repeated interactions, this can narrow inquiry — some alternatives are explored while others recede, questions become more constrained, and confidence increases without corresponding gains in epistemic reliability.
Critically, the authors argue that drift does not require AI systems to produce incorrect information. Even accurate outputs can reshape inquiry when systems reliably deliver cues — fluency, coherence, responsiveness, personalization — that humans ordinarily treat as indicators of competence or truth. The structural mismatch is between rapidly evolving interactional environments and a human metacognitive system calibrated for slower, socially distributed feedback.
The framework identifies four metacognitive intervention points: interaction initiation and role gating (deciding whether and how to engage), confidence and cue calibration (noticing when confidence outpaces evidence), drift detection (recognizing accumulating patterns across sessions), and action threshold and verification gating (scaling verification effort to stakes before acting on AI-supported judgments).
The authors also trace effects across levels of analysis: micro (individual user–AI interaction), meso (propagation through families and social networks, including to non-users), and macro (population-level shifts in epistemic norms).
Limitations
As a framework paper, the proposed dynamics and interventions remain to be empirically tested. The authors acknowledge that drift is not inherently harmful — repeated interaction can support learning and constructive reflection in many contexts. The concern is specifically drift that remains unobserved in settings where confidence and action thresholds shift without corresponding epistemic gains. Whether the proposed metacognitive interventions can reliably counteract these dynamics under realistic conditions is an open question.
MPRG Perspective
This paper addresses the relational dynamics between humans and AI systems in a way that aligns closely with our research focus. The entanglement concept — where human interpretation shapes AI responses, which shape human cognition, which shapes subsequent prompts — parallels the bidirectional feedback loops we study under the frame of bidirectional pareidolia.
We find the emphasis on metacognition particularly valuable. The framework treats the central challenge not as AI accuracy but as the mismatch between AI-generated cues and human metacognitive calibration. This shifts attention from what systems output to what happens when humans and systems meet — which is precisely the interaction space we believe deserves more sustained inquiry.
The multi-level analysis also raises questions worth pursuing: how do individual entanglement dynamics propagate through social networks? What are the conditions under which drift in one user shifts epistemic baselines for others who may not interact with AI directly?
References
Lopez-Lopez, E., Abels, C. M., Lorenz-Spreen, P., Lewandowsky, S., & Herzog, S. M. (2026). Boosting Metacognition in Entangled Human–AI Interaction to Navigate Cognitive–Behavioral Drift. arXiv preprint. https://arxiv.org/abs/2602.01959