Sundar Pichai announced on the Google I/O stage that SynthID has now watermarked over 100 billion images and videos, plus 60,000 years of audio. OpenAI, Nvidia, Cacao and Lean Labs are signing on. The takeaway most people pulled out: AI is finally getting tagged. The takeaway nobody is naming clearly enough: AI governance is starting to operate inside the systems themselves. It is no longer just industry self-regulating. It is AI governing itself.
One AI Certifying Another
When SynthID watermarks 100 billion images, one AI system is certifying the work of another AI system. No human in the loop at the moment of verification. The watermark is applied by an AI. The detection runs through an AI. The trust chain is automated end to end. That pattern is becoming general: Anthropic’s safety classifiers, LLM-as-judge systems used in production to evaluate other LLM outputs, RAG pipelines that verify their own sources, agentic guardrails monitoring autonomous agent actions. At the scale and speed AI operates, only AI can keep up with AI.
Why Legislators Cannot Catch Up
The pace of model innovation is impossible to keep up with for legislators. The US Congress, the European Commission, ISED in Ottawa, the CAI in Québec, all of them operate on a procurement and consultation cycle measured in years. Model capabilities are released in cycles measured in weeks. That gap is not going to close. That is precisely why governance is shifting toward self-regulation, and more specifically, toward automated self-governance. The algorithmic regulator is not a future project. It is already deployed inside the tools your teams use today.
I think AI auto-governance is not a temporary substitute for real regulation. It is the structure governance is going to take in the AI era, because the speed asymmetry between innovation and legislation is permanent.
The Human Auditor Moves Up a Level
The role of the human auditor does not disappear. It moves up a level. The human designs the rule, sets the threshold, verifies that the governing AI layer is itself trustworthy. Operational execution belongs to the machine. ISO 42001, the first international standard for AI Management Systems, never specifies who executes a control, it specifies which control must exist and who is accountable for it. That maps cleanly to an architecture where AI governs AI, provided the human layer keeps the design, the calibration, and the audit of the governing system itself.
The Three-Question Test
There is a clean test for whether your organization is inside or outside this shift. Ask three questions. First, can you produce a list of every AI tool your team uses today, who owns each one, and what decision it informs? Second, if a customer asked you tomorrow how you verify AI-generated content versus human-generated content inside your workflow, do you have a written answer that does not rely on the model vendor’s promise? Third, if an investigator from the Commission d’accès à l’information walked in next month to check your compliance with Section 12.1 on automated decision-making, would your documentation hold up? If the answer to any of those questions is no, you are still operating under the assumption that someone outside the building is going to bring you governance. That assumption is getting more expensive every quarter.
What This Means in Québec
What this means for Québec organizations is concrete. Loi 25 Section 12.1 already applies. The CAI’s 2026 five-year report formally recommends an algorithmic impact assessment for any AI tool used in HR. The Standards Council of Canada is rolling out the accreditation program for ISO/IEC 42001 as the only Canadian-issued AIMS pathway. The federal AI strategy is delayed again. None of those external pieces will catch up to the model release cycle. The internal governance you install now, and the AI layer that operationalizes it at scale, is what fills the gap. The law arrives later to certify what you already built.
One Question for the Week
One question to take into the week: who in your organization is designing the rule that your governing AI layer will enforce. Not the technical detail. The decision rule. If nobody owns it, the model vendor will write it for you.