Singapore Consensus on Global AI Safety Research Priorities
Issued by
Expert committee convened by Singapore's IMDA at the 2025 Singapore Conference on AI
- September 20, 2026 · Correction — The 2026-09-17 daily pipeline run drafted a distinct 2026 academic paper (different jurisdiction, policy_type, issuing body, and a 14-month effective-date jump) and classify() incorrectly matched it as an update to this 2025 entry, overwriting it. Restored this entry to its original May 2025 content and moved the 2026 paper to its own entry, singapore-consensus-global-ai-safety-research-priorities-2026. 24 published news items link to this slug, all dated before the overwrite; this restores those links to accurate content. (Cody Maxwell)
- September 30, 2026 · Correction — Corrected the date and nature of the document: an expert-written research agenda from a scientific exchange, not a government declaration, since updated in 2026 (fact-check finding). (Cody Maxwell)
- September 30, 2026 · Correction — Checked the rewrite against the report itself; removed an unsourced ICLR link and matched the participant and authorship details to the source. (Cody Maxwell)
- October 1, 2026 · Correction — Rewrote the practical steps, newsletter hook, search description, and audience fields to match the entry's corrected content. (Cody Maxwell)
The Singapore Consensus is a research agenda for AI safety, written by an expert committee after a scientific exchange Singapore hosted in April 2025. It groups research priorities into development, assessment, and control. It is not a government declaration, has no signatories, and was updated by a 2026 edition.
Applies To
Overview
The Singapore Consensus on Global AI Safety Research Priorities came out of the 2025 Singapore Conference on AI. That scientific exchange was held in Singapore on 26 April 2025. More than 100 researchers, policymakers, and private sector participants from 11 countries took part. The report is dated 8 May 2025. An Expert Planning Committee drafted it, including Yoshua Bengio, Stuart Russell, Max Tegmark, and Dawn Song. Contributors took part in their individual capacities, and Singapore's IMDA convened the exchange rather than issuing the text. It sets out areas of broad agreement on AI safety research, organized on a defense-in-depth model: developing trustworthy systems, assessing their risks, and controlling them after deployment. It builds on the International AI Safety Report, which 33 governments backed, and has no signatories, endorsement, or reporting process. A 2026 Singapore Consensus, from a second exchange in May 2026, updates this edition.
Key Requirements
- •A research agenda, not a set of obligations; it has no signatories or reporting process.
- •Development: research on building AI systems that are trustworthy by design.
- •Assessment: research on measuring and evaluating AI risks.
- •Control: research on monitoring and intervening in AI systems after deployment.
- •Updated by the 2026 Singapore Consensus.
What Your Organization Must Do
- →Treat this document as a research agenda, not a source of compliance duties or reporting steps.
- →Read the 2026 Singapore Consensus first, since it updates this 2025 edition.
- →Map your own AI assurance work to the three themes: development, assessment, and control.
- →Brief engineering and risk teams on layering safeguards before and after deployment.
- →Check the International AI Safety Report, which 33 governments backed, for the underlying evidence base.
- →Avoid citing this text to regulators as evidence of binding commitments or endorsement.