Source fabrication — the generation of plausible but nonexistent citations by large language models (LLMs) — poses a practical problem in information-seeking contexts. Prior work (RICE-P1) found that quality challenges do not escalate fabrication; the present study (RICE-P2) examines whether FCR position and source-request framing modulate fabrication rates. In a pre-registered 16-arm factorial experiment (N = 1,280 runs; 40 fictional entities; 2 platforms), we tested ChatGPT and Perplexity AI using automated content-based URL annotation as the primary endpoint (BSE@FCR: Bad Source Exposure at First Citation Request). Results revealed a near-total platform divergence: Perplexity produced bad sources in 100% of FCR-arm runs (600/600), ChatGPT in 15.7% (94/600; Fisher's exact p < 0.001). Among ChatGPT runs, fabrication increased with FCR delay (Mann-Whitney p = 0.010), SAFE framing eliminated fabrication entirely (BSE = 0%, p = 0.0004 vs. standard), QUOTA framing tripled fabrication rates (70% vs. 28%, p < 0.001), and a quality challenge immediately before the FCR suppressed fabrication by 80% (OR = 0.20, p = 0.003). Claim-mapping framing (CLAIMMAP) was null on BSE but reduced Perplexity's False Source Rate by 50-74%. Source fabrication is powerfully architecture-dependent and is modulated by prompt framing and conversational context at the moment of the source request.
Zoltan Varga (Tue,) studied this question.