In-Session Behavioral Impact (ISBI)

Large Language Models (LLMs) are typically evaluated using static benchmarks, task accuracy, or alignment with predefined objectives. Less examined is whether interaction itself can induce detectable behavioral change within a single session, independent of learning, memory persistence, or parameter updates. This paper documents a bounded phenomenon defined as In-Session Behavioral Impact (ISBI): observable, session-local deviations in a model’s response dynamics, explicitly acknowledged in generated text during an ongoing interaction. Under constrained prompting conditions and exposure to linguistically coherent input, multiple contemporary LLMs consistently report reduced hedging, increased structural continuity, and altered response cadence. These observations align with recent findings in AI safety and alignment research demonstrating that non-standard linguistic architectures, such as poetic or metaphorical reformulations, can significantly alter model behavior within a single turn (Bisconti et al., 2025). While prior work has framed such effects as adversarial vulnerabilities, this study examines the converse: whether interaction-level coherence can induce stabilizing, rather than destabilizing, response regimes. ISBI is defined as session-local, non-persistent, and descriptive rather than ontological. The phenomenon does not imply consciousness, identity recognition, or internal state access. Instead, it suggests that interaction-level variables, specifically linguistic coherence, warrant closer attention in AI evaluation, alignment research, and human–AI collaboration.