EU AI Act Transparency Rules Push Businesses to Build AI Governance Ahead of Delayed High-Risk Deadlines
This opinion piece by Khush Kashyap (Vanta) discusses the EU AI Act's transparency requirements that took effect in August 2026, requiring providers and deployers to label or disclose AI-generated or manipulated content. Providers like Anthropic and OpenAI have begun implementing watermarking, but the article highlights significant limitations: OpenAI's own testing showed that replacing 25% of words with synonyms dropped watermark detection accuracy from ~92% to ~17%. The Digital Omnibus regulation, finalized in July 2026, has delayed high-risk AI obligations to December 2027 (Annex III systems) and August 2028 (AI in regulated products), and narrowed the definition of "safety component." The author argues businesses should use this extended timeline to build AI governance infrastructure—including AI asset inventories, impact assessments, and governance structures—rather than waiting for regulatory deadlines, to avoid rushed compliance and security theater.
Remediação
The article recommends three priorities for organizations: (1) build a complete AI asset inventory covering all AI systems, their data processing, ownership, and risk classification; (2) conduct AI impact assessments for higher-risk systems covering bias, human oversight, transparency, accountability, security and data handling; (3) create clear AI governance structures with defined ownership across security, privacy, legal, compliance and product teams, maintaining audit trails for approvals and risk monitoring.
Linha do tempo
-
EU AI Act Transparency Rules Push Businesses to Build AI Governance Ahead of Delayed High-Risk Deadlines
ler a análise desta fonte ocultar
This opinion piece by Khush Kashyap (Vanta) discusses the EU AI Act's transparency requirements that took effect in August 2026, requiring providers and deployers to label or disclose AI-generated or manipulated content. Providers like Anthropic and OpenAI have begun implementing watermarking, but the article highlights significant limitations: OpenAI's own testing showed that replacing 25% of words with synonyms dropped watermark detection accuracy from ~92% to ~17%. The Digital Omnibus regulation, finalized in July 2026, has delayed high-risk AI obligations to December 2027 (Annex III systems) and August 2028 (AI in regulated products), and narrowed the definition of "safety component." The author argues businesses should use this extended timeline to build AI governance infrastructure—including AI asset inventories, impact assessments, and governance structures—rather than waiting for regulatory deadlines, to avoid rushed compliance and security theater.
Remediation
The article recommends three priorities for organizations: (1) build a complete AI asset inventory covering all AI systems, their data processing, ownership, and risk classification; (2) conduct AI impact assessments for higher-risk systems covering bias, human oversight, transparency, accountability, security and data handling; (3) create clear AI governance structures with defined ownership across security, privacy, legal, compliance and product teams, maintaining audit trails for approvals and risk monitoring.