Generative models have made it possible to synthesize convincing content across language, vision, speech, and even biological sequences. This progress has opened powerful avenues for privacy-preserving data sharing, scientific discovery, and creative work, but it has also introduced new risks: misuse for deception, copyright infringement, leakage of sensitive training data, and the potential design of harmful biological agents. This dissertation presents a unified body of work on generating synthetic content responsibly and detecting and constraining it when misused.

The dissertation contributes five complementary frameworks. FakeDB synthesizes relational databases that preserve semantic constraints and statistical properties, enabling secure information exchange in place of sensitive real data. GEM steers multimodal generation toward user-specified objectives through prompt tuning and discriminative feedback, and is paired with an efficient finetuning strategy that makes high-quality synthesis more scalable. O3 addresses unintended information retention in large language models by supporting continual unlearning through orthogonal adapters and out-of-distribution detection. CADD, a context-based audio deepfake detector, leverages transcripts and contextual metadata such as news and social media signals to identify fabricated speech, particularly in public statements where the cost of deception is high. Finally, SafePro extends these ideas beyond conventional modalities to protein language models, introducing a multi-objective alignment framework that jointly optimizes structural foldability while steering generation away from toxicity and virulence.

Together, these contributions form a coherent agenda for promoting authenticity, privacy, safety, and trust in the generative AI era, and they suggest concrete paths forward for deploying generative systems in domains where the stakes of getting it wrong are highest.