Trustworthy AI for the ocean: bridging the science-policy divide

Thousands of ocean science papers are published each year, yet much of this knowledge remains inaccessible to the non-expert decision-makers advancing ocean sustainability. General-purpose large language models (LLMs) offer a pathway to synthesis but often fall short on reliability, transparency, and scientific grounding. We present IPOSGPT, a domain-specific LLM grounded on a curated corpus of peer-reviewed literature, intergovernmental reports, and environmental policy documents. Across 25 policy-relevant queries, IPOSGPT outperformed leading generalist models (gpt-4o, gemini-2.0-flash, sonnet-4, sonar-pro) on citation credibility and traceability, producing zero fabricated references compared with over 95 percent invalid citations on average for baselines. On answer quality, IPOSGPT was competitive, frequently ranking second while maintaining strict faithfulness to evidence. In a Seychelles case study, the system accelerated assessment of circular economy options to reduce marine pollution. These results show that domain-specific LLMs can deliver trusted synthesis for high-stakes sustainability policy, addressing hallucination, source integrity, and accountability.