Abstract
Generative AI can scale social-media production, but its synthetic linguistic signature may reduce audience trust. This paper quantifies that trade-off using a two-phase transformer pipeline. A BERT-base-uncased detector, fine-tuned on mixed-domain human and AI text, achieved 93.85% validation accuracy (macro F1 = 0.94, N = 4,274). The detector was then applied to 7,225 de-duplicated TikTok captions from 2025 to estimate a continuous AI-authorship probability, PAI. The pre-specified aggregate test found a small negative relationship between PAI and engagement rate (Pearson r = −0.0467, p < 0.001, 95% CI [−0.070, −0.024]; R 2 = 0.22%). Exploratory rank correlations were weak and directionally mixed: likes decreased (ρ = −0.022, p = 0.062), comments increased (ρ = 0.029, p = 0.014), and shares were unchanged (ρ = 0.013, p = 0.269). The most AI-intensive decile recorded 11.21% mean engagement versus 13.37% for the most human-authentic decile, a 2.16-percentage-point or 16.2% relative gap. We term this observed, non-causal contrast the Authenticity Tax. The results support a hybrid workflow in which AI supplies structure and scale while humans restore voice, emotion, and contextual specificity.