Advances in LLM and speech technology are diversifying human-AI interactions. We investigated phonetic entrainment in spontaneous human-AI dialogues, examining its acoustic characteristics and relation to user attitudes. Analyzing 18 h of conversational audio and post-interaction questionnaires (AIAS-4, TES) from 40 mandarin speakers engaging with a chinese LLM (Doubao), we found widespread entrainment encompassing both synchrony and convergence. Overall synchrony was frequent (avg. 52% of inter-pausal units), with intensity and speech rate showing synchrony, while pitch often exhibited counter-synchrony. Convergence was feature specific: the mean f0 was dominated by convergence (43.6%), while most other characteristics remained largely stable or fluctuated between convergence and divergence. Compared to human-human dialogues, human-AI entrainment appeared weaker and less differentiated across features, supporting the interpretation of entrainment as also serving a social distance function in the human-AI interaction (HAII). Specifically, task type modulated entrainment modes: Open-ended chat favored immediate synchrony in intensity and speech rate, while collaborative planning elicited stronger long-term convergence in the f0 and wav2vec2 embeddings, highlighting that phonetic adaptation in HAII is dynamically tailored to communicative goals. Finally, an Entrainment Index (EI) quantifying both synchrony and convergence approximated a normal distribution with a slightly heavier left tail, revealing a wide spectrum of individual differences. However, these differences were not significantly correlated with conscious attitudes as measured by the AIAS-4 or TES scales, suggesting that entrainment may be unconscious behavior or driven by deeper psychological factors rather than explicit perceptions.