Japan AI Guidelines HP.I: D-I. Risk identification, evaluation and mitigation across the lifecycle

Take appropriate measures throughout development of advanced AI systems, including before deployment and before placing on the market, to identify, evaluate and mitigate risks across the AI lifecycle. Code of Conduct detail for developers of advanced systems: combine diverse internal and independent external testing, including red-teaming, and mitigate the risks and vulnerabilities found; enable traceability of datasets, processes and development decisions; document the measures in regularly updated technical documentation; test in secure environments at several checkpoints, especially before deployment and market placement; attend to chemical, biological, radiological and nuclear risks, offensive cyber capabilities, risks to health and safety including control of physical systems and interference with critical infrastructure, self-replication or training of other models, societal risks including harmful bias and breaches of privacy law, threats to democratic values and human rights, and chain-reaction risks reaching a whole city, domain or community; work with actors across sectors on mitigations, in particular for systemic risks; and advance research on security, safety, bias and disinformation, fairness, explainability, interpretability and transparency.