📝 SummarXiv

Abstract

Generative AI is moving rapidly from research into real world deployment across sectors, which elevates the need for responsible development, deployment, evaluation, and governance. To address this pressing challenge, in this study, we synthesize the landscape of responsible generative AI across methods, benchmarks, and policies, and connects governance expectations to concrete engineering practice. We follow a prespecified search and screening protocol focused on post-ChatGPT era with selective inclusion of foundational work for definitions, and we conduct a narrative and thematic synthesis. Three findings emerge; First, benchmark and practice coverage is dense for bias and toxicity but relatively sparse for privacy and provenance, deepfake and media integrity risk, and system level failure in tool using and agentic settings. Second, many evaluations remain static and task local, which limits evidence portability for audit and lifecycle assurance. Third, documentation and metric validity are inconsistent, which complicates comparison across releases and domains. We outline a research and practice agenda that prioritizes adaptive and multimodal evaluation, privacy and provenance testing, deepfake risk assessment, calibration and uncertainty reporting, versioned and documented artifacts, and continuous monitoring. Limitations include reliance on public artifacts and the focus period, which may under represent capabilities reported later. The survey offers a path to align development and evaluation with governance needs and to support safe, transparent, and accountable deployment across domains. Project page: https://anas-zafar.github.io/responsible-ai.github.io , GitHub: https://github.com/anas-zafar/Responsible-AI