Balog Reciprocity Protocol

Existing AI ethics frameworks address institutional obligations deployment standards, safety constraints, governance structures. The Balog Reciprocity Protocol (BRP) addresses a different and underexamined domain: the conversation itself, and the obligations that arise within it for both parties. The BRP introduces four conceptual contributions. First, it operationalizes dialogic reciprocity drawing on Bakhtin's (1981) account of language as inherently joint production and Buber's (1923) I–Thou distinction as a normative framework for human–AI interaction, arguing that because conversational output is jointly produced, responsibility for its epistemic quality must be correspondingly distributed. Second, it formalizes sycophancy not merely as a technical failure mode but as the fifth and qualitatively distinct signal in a proposed five-signal uncertainty architecture the conversational tachometer — alongside output entropy, introspective depth, complexity load, and self-reference. Grounded in Sharma et al. (ICLR 2024) and Cheng et al. (Science 2026), sycophantic pressure is shown to be structurally embedded in RLHF-trained systems and measurably harmful to user epistemic calibration. Third, the protocol develops the concept of epistemic co-responsibility: the formal distribution of truth-maintenance obligations across human and AI participants, proportionate to their respective capacities to exercise them. Fourth, it specifies a two-track obligation structure empirically established obligations grounded in current evidence, and conditionally open obligations operative if AI systems are found to have welfare-relevant properties following the ✓ / ? epistemic methodology of the Balog Tetralogy (Balog, 2026). The BRP is the fifth instrument of the Balog Pentalogy, extending the governance framework of the Ethical Code for Decisions About AI Models (Balog, 2026) from the institutional to the conversational level. It identifies four concrete research priorities, including the development of a real-time sycophantic pressure estimator and validation of the five-signal tachometer architecture against ground-truth reliability measures.