ITU report sets global norms for AI safety testing and watermarking
The ITU has released its Annual AI Governance Report 2025, which provides a non-binding global framework for AI pre-deployment safety testing, red-teaming, and watermarking. The report is expected to serve as a key reference for regulators moving these practices from voluntary recommendations toward mandatory enterprise compliance obligations.
Key Takeaways
- Identifies three primary enterprise pressure points: model release gates, safety evaluation evidence, and content provenance maturity.
- Categorizes content watermarking as a mandatory post-deployment transparency obligation rather than an optional technical feature.
- Proposes formal licensing and registration for AI models to ensure documented approval gates and audit trails.
- Signals regulatory convergence across the EU, UK, and U.S. states regarding adequate pre-deployment safety controls.
Why It Matters
The ITU's framework serves as a multilateral baseline that will likely inform upcoming rulemaking by the EU AI Office and U.S. federal agencies. For streaming and media enterprises, this shifts watermarking from a technical experiment to a required audit trail for synthetic content. Failure to formalize these release gates now creates significant exposure as voluntary norms harden into enforceable licensing and disclosure regimes by late 2026. Watch for whether specific ITU testing standards are cited in the EU's final implementation guidance for high-risk AI systems.
Additional Context
The ITU report’s emphasis on watermarking coincides with the fast-approaching August 2, 2026, deadline for Article 50 transparency obligations under the EU AI Act. Per ComplianceHub (June 2026), these rules require all generative AI outputs to be machine-readable and detectable, with deepfakes and public-interest text requiring explicit labeling. While a transitional period allows legacy systems until December 2026 to comply, new deployments must meet the August mandate or face significant penalties. Google and other major providers have already committed to these standards, signing the EU’s Code of Practice on Transparency in July 2024 to accelerate the adoption of tools like C2PA and SynthID. Simultaneously, the OECD updated its AI Principles in May 2024 to address the risks posed by general-purpose and generative AI, specifically focusing on information integrity and accountability throughout the AI lifecycle. According to OECD reports from July 2025, over 47 governments have now committed to these principles, which utilize the same definitions of AI systems found in the EU AI Act and the new ITU framework. This alignment indicates a global move toward interoperable governance, where technical standards for safety testing developed by the ITU will provide the evidentiary basis for audits under national laws. Industry readiness remains a critical concern. A study cited by Resemble AI in June 2026 found that only 38% of AI image generators had implemented adequate watermarking practices. As the 'compliance splinternet' merges into a unified global baseline, enterprises are being pushed to move beyond ad hoc safety reviews. The convergence between the ITU standards and the Singapore Consensus on Global AI Safety Research Priorities suggests that the window for voluntary self-regulation is closing, as multilateral bodies move toward mandatory registration and licensing for frontier models.
Read full article at aigovernance.com
Get this in your inbox → Subscribe
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source