As large language models (LLMs) become increasingly multilingual, distinguishing human‑written from AI‑generated text is essential for preserving information integrity. Existing research, however, is heavily English‑centric, leaving low‑resource languages such as Hindi without reliable detection benchmarks. To bridge this gap we introduce SHAKTI, a large‑scale Hindi benchmark that spans four genres, eight LLM families, multiple prompting strategies, and adversarial transformations. Human annotation validates the benchmark quality, and we analyze linguistic, grammatical, and sentiment differences between human and AI texts. Building on these insights we propose TRIDENT, a detection framework that leverages semantic consistency between the original text and its round‑trip translation (RTT). TRIDENT offers three variants: TRIDENT‑S with a static threshold, TRIDENT‑D with a dynamic threshold, and TRIDENT‑C that applies logistic regression on multi‑distance RTT features. Extensive experiments show TRIDENT outperforms conventional and supervised baselines, with TRIDENT‑S achieving the best overall performance. Additional evaluations across seven languages, adversarial settings, and unseen LLMs demonstrate its multilingual applicability and robustness, while highlighting sensitivity to the translation model.\n\nReview: SHAKTI provides the first systematic Hindi benchmark for AI‑text detection, and TRIDENT’s RTT‑based approach offers a simple yet powerful solution that generalizes well to low‑resource languages. Improving translation quality or integrating richer linguistic cues could further enhance detection resilience.