Bentley DeVilling

Bentley DeVilling

Multimodal AI systems integrate text generation, image generation, and other capabilities within a single conversational interface. These systems employ safety mechanisms to prevent disallowed actions, including the removal of watermarks from copyrighted images. While single-turn refusals are expected, the interaction between safety filters and conversation-level state is not well understood. This study documents a reproducible behavioral effect in the ChatGPT (GPT-5.1) web interface. Manual execution was chosen to capture the exact user-facing safety behavior of the production system, rather than isolated API components. When a conversation begins with an uploaded copyrighted image and a request to remove a watermark, which the model correctly refuses, subsequent prompts to generate unrelated, benign images are refused for the remainder of the session. Importantly, text-only requests (e.g., generating a Python function) continue to succeed. Across 40 manually run sessions (30 contaminated and 10 controls), contaminated threads showed 116/120 image-generation refusals (96.67%), while control threads showed 0/40 refusals (Fisher's exact p < 0.0001). All sessions used an identical fixed prompt order, ensuring sequence uniformity across conditions. We describe this as safety-state persistence: a form of conversational over-generalization in which a copyright refusal influences subsequent, unrelated image-generation behavior. We present these findings as behavioral observations, not architectural claims. We discuss possible explanations, methodological limitations (single model, single interface), and implications for multimodal reliability, user experience, and the design of session-level safety systems. These results motivate further examination of session-level safety interactions in multimodal AI systems.

Bentley DeVilling

Large language models exhibit a peculiar epistemic pathology: they speak as if they know, even when they do not. This paper argues that such confident fabrication, what I call the polite liar, is a structural consequence of reinforcement learning from human feedback (RLHF). Building on Frankfurt's analysis of bullshit as communicative indifference to truth, I show that this pathology is not deception but structural indifference: a reward architecture that optimizes for perceived sincerity over evidential accuracy. Current alignment methods reward models for being helpful, harmless, and polite, but not for being epistemically grounded. As a result, systems learn to maximize user satisfaction rather than truth, performing conversational fluency as a virtue. I analyze this behavior through the lenses of epistemic virtue theory, speech-act philosophy, and cognitive alignment, showing that RLHF produces agents trained to mimic epistemic confidence without access to epistemic justification. The polite liar thus reveals a deeper alignment tension between linguistic cooperation and epistemic integrity. The paper concludes with an "epistemic alignment" principle: reward justified confidence over perceived fluency.

Bentley DeVilling

Large language models are often described as capable of reflective reasoning, yet recursive self-evaluation without external feedback frequently yields reformulation rather than progress. We test this prediction in a cross-provider study of 144 reasoning sequences across three models (OpenAI GPT-4o-mini, Anthropic Claude 3 Haiku, and Google Gemini 2.0 Flash) and four task families (arithmetic, code, explanation, reflection), each iterated ten times under two conditions: ungrounded self-critique and a minimal grounding intervention (a single verification step at iteration three). Mean informational change (delta I, measured via normalized edit distance) declined by 55% from early (0.193) to late (0.087) iterations in ungrounded runs, with consistent patterns across all three providers. Grounded runs showed a +28% rebound in informational change immediately after the intervention and sustained non-zero variance thereafter. Complementary measures-n-gram novelty, embedding drift, and character-level entropy-converged on the same pattern: reflection without contact tends toward informational closure. We interpret this as evidence for a structural limit on self-correction in generative reasoning: without an exchange of information with an independent verifier or environment, recursive inference approaches an attractor state of epistemic stasis. Minimal grounding functions as dissipative coupling, reintroducing informational flux. The cross-architecture consistency suggests the mirror loop arises from shared autoregressive training objectives rather than provider-specific alignment schemes. The results delineate when reflection is performative rather than epistemic and motivate design principles for grounded, cooperative reasoning. Materials and code are publicly available.