Who is Responsible? The Data, Models, Users or Regulations? A Comprehensive Survey on Responsible Generative AI for a Sustainable Future

Benchmark gaps leave critical risks unmonitored: While AI safety benchmarks thoroughly evaluate bias and toxicity, they provide sparse coverage of privacy, data provenance, deepfakes, and system-level failures—meaning organizations may deploy generative AI systems without adequately testing these significant risks.

Current evaluations are static and not portable across systems: The survey found that most AI evaluations are designed for specific tasks and don't transfer well to new contexts or updated models, making it difficult for auditors and regulators to consistently assess safety across different generative AI applications and versions.

Poor documentation undermines accountability: Inconsistent documentation practices across AI releases prevent meaningful comparison between different versions of models, complicating efforts by regulators, organizations, and researchers to track whether safety improvements are actually being made over time.

Responsibility requires action across four domains: The survey identifies that responsible AI depends on improvements in data practices, model development, user governance, and regulatory frameworks—suggesting that no single stakeholder (data providers, AI developers, users, or policymakers) can solve this alone.

Agentic AI systems need urgent safety attention: The research highlights that autonomous AI systems operating in pipelines present unique system-level failure risks that current benchmarks don't adequately address, representing an emerging governance challenge as these systems become more prevalent.