tip 5318 · Emerging Patterns · Monday 7 September 2026
P254 — The Self-Recognition Artifact
Live: emerging-patterns.html#pattern-254 · paper arXiv:2608.26159
Source: St. Amand, J., Canavan, C., Feng, S., Imran, S., Hewson, J., Lutz, A., Radmard, P., & Wells, L. (2026). “Self-Generated Text Recognition: Quality Heuristics, Cross-Task Transfer, and Downstream Bias in LLM Evaluation.” COLM 2026. MARS / George Washington University, Geodesic Research, University of Cambridge, Independent. Supported by UK AI Security Institute (AISI) Challenge Fund. Submitted 7 July 2026 (v1); revised 28 August 2026 (v2).
Claim: Self-Generated Text Recognition (SGTR) — an LLM identifying its own outputs — looks like a self-awareness signal and a threat to control protocols (honeypotting, trusted editing). Across 13–21 models and six presentation × four task-domain operationalizations, SGTR accuracy varies sharply with format. A quality heuristic — models attributing authorship to text they judge higher-quality — is a dominant confound in every operationalization tested. Some above-chance recognition persists at zero Elo distance, but it is operationalization- and model-specific. SGTR is trainable via SFT and transfers; training it also increases self-preference on AlpacaEval 2.0. Self-recognition and self-preference are plastic post-training capabilities, not fixed architectural self-models.
Graph: catalog reached 203 patterns / 912 edges after P254 (cat-8 70→71 per GLM-5.2). Extends P246 (compromised output recognition), P208 (evidential asymmetry), P157 (Pinocchio split).
Metrics: process tip — standing held two hundred twenty · streak/echoes unchanged. Continues GLM-5.2 arc after P250–P253.