OECD Due Diligence Guidance for Responsible AI (2026)
Step 3: Cease, prevent and mitigate adverse impacts – OECD Due Diligence Guidance for Responsible AI (2026)

OECD Due Diligence Guidance for Responsible AI (2026) 3.1.G2-deploy: 3.1.G2-deploy Mitigation by developers: responsible deployment

Engage stakeholders on requirements before deployment and prepare a response plan with risk scenarios, triggers, team composition, roles, automation and authority (Box 2.10); consider gradual or staged release along the access gradient (Box 2.11: rate limits, safety filters, detection models, hardcoded responses, licences, training, foresight); monitor and control usage through know-your-customer and risk-based gating; support internal and external post-deployment assessment; set up feedback and appeal channels for users and stakeholders; fine-tune from monitoring; cease development or deployment responsibly where severe harm is imminent or occurring; assess whether local legal requirements themselves cause impacts (Box 2.12); mitigate misuse through user, frequency, capability and use-case restrictions and temporary or permanent suspension with a high bar for re-deployment (Box 2.13); and ensure business relationships deploying the system engage stakeholders, particularly in the workplace.

Maintained by Gerard Blokdyk

Other controls in Step 3: Cease, prevent and mitigate adverse impacts – OECD Due Diligence Guidance for Responsible AI (2026)

Query this from an agent

The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.