How a single label like 'Example' vs 'Instruction' can shift AI trust by 84%

A new arXiv paper (2606.04109) by Jianguo Zhu reveals that the simple label wrapping external context can dramatically distort how language models use that information. Across four models—GPT-5.5, DeepSeek V4 Pro, Llama-3-8B-Instruct, and Qwen2.5-7B-Instruct—the same misleading assertion was presented under different discourse-role labels (e.g., “Reference:”, “Evidence:”, “Instruction:”, “Note:”, “Example:”). The Misleading Adoption Rate shifted by 56 to 84 percentage points depending solely on the label. Labels that signal authority or binding (like “Instruction:” or “Reference:”) caused models to adopt the false information at much higher rates, while “Example:” consistently suppressed it. This effect held across multiple experimental controls: paired statistical tests, bootstrap intervals, final-instruction ablations, and log-probability probes on Qwen confirmed a label-conditioned candidate preference.

Boundary probes further clarified the phenomenon. Arithmetic tasks reduced overall adoption, but the label gap persisted; passage-shaped external context preserved smaller gaps; short-answer evaluation ruled out option-letter copying artifacts; and nested-label conflicts suggested that illustrative framing can actually delimit adoption scope. A 200-case manual audit verified that short-answer contrasts remained stable under conservative adjudication. The authors conclude that context-augmented language model systems—especially RAG pipelines—must report and control wrapper labels because presentation choices can fundamentally change measured reliance on supplied context. The paper is a preprint, available on arXiv.