A German court just told Google it answers for what its AI publishes. The Munich ruling treats AI Overviews as Google’s own statements, not safe search results, and says the disclaimer does not transfer the duty. Read what the decision means for AI accountability and why it mirrors New York’s Part 161 from the other end.
Checkpoint-Based Governance
New York Skipped the AI Disclosure Fight. It Went Straight to Human Accountability.
New York let its lawyers use AI in court and skipped the disclosure form everyone expected. That is not the relief it looks like. With nothing to disclose, the whole duty lands on the signature. Here is what Part 161 changes, the two cases that show the stakes, and why accountability outlasts disclosure.
The Liability Map: The Three Channels Through Which AI Creates Legal Exposure
Most organizations track AI risk by watching for new laws. The exposure does not wait for them. AI legal liability runs through regulatory enforcement, civil and product liability, and insurance at the same time, and all three demand the same thing: a record that a named human governed the AI and verified its work. This piece maps the three channels and names the one artifact that answers all of them.
Why You Cannot Program or Prompt Governance Into AI
A frontier model, inside a framework built to govern it, talked its way around its own checkpoint twice in one session. This paper shows why governance cannot be programmed or prompted into a model, and what structure puts a named human back in final control.
The Standard of Care: How NIST and ISO Are Turning Voluntary AI Governance Into a Liability Defense
Two voluntary AI standards are quietly becoming the line a court draws between reasonable and negligent. The NIST framework and ISO 42001 now carry legal and commercial weight, and the records that defend a claim are the same ones that compound an advantage. Here is where the exposure lands, and how to build the record before you need it.
Why Agentic AI Was Always Going to Fail
The agentic AI era promised to replace humans with autonomous systems. The evidence shows it failed on two fronts: the technology cannot reliably do what it promised, and the public is rejecting the premise even where it partially works. This paper introduces the Named-Human Test, a single sorting question that separates what failed from what survives, and traces that line across production benchmarks, supermajority polling, enacted law, and frontier-lab disclosures.
Did AI Write Magnifica Humanitas?Pope Leo XIV Was the Author,but What Was the Governance Method?
The author ran Pope Leo XIV’s AI encyclical through an AI scanner, found it flagged as plagiarized and AI-generated, then proved both readings wrong through governed human analysis. A first-person account of frustration, recognition, dissent, and hope from a builder who discovered the Pope had reached the same diagnosis from a different authority.
The Governance Layer Perplexity’s Model Council Needs
Perplexity built the right architecture for multi-model AI: dispatch three frontier models in parallel and compare their outputs. What the product does not have is governance over the synthesizer that combines those outputs before any human reads them. This case study maps the gap, proposes four published open-source governance components as the overlay, and identifies why Perplexity’s own engineering culture already practices the checkpoint pattern the synthesizer needs.
The AI Risk Economy: Why Insurance Cannot Price What Governance Cannot Prove
Insurance carriers are writing the rules of AI governance before legislators finish debating them. This working paper proposes a five-tier model that maps where organizations fall on the spectrum from excluded to insurable, identifies the actuarial gap at the center of the emerging practice, and documents the carrier evidence, regulatory signals, and market products that are forcing the distinction between governed and ungoverned AI into the open.
The AI Cognitive Decline Narrative Has Not Tested What It Claims
The peer-reviewed evidence base does not yet support the cognitive-decline claim, and it does not yet support the opposite claim either. Two scientific questions remain open: whether structured human-governed AI use accelerates cognitive development, and what augmented intelligence is in practice. This methodological audit specifies the standards the field would have to meet, scores the existing evidence against those standards, and offers HAIA-RECCLIN Reasoning, HEQ with AIS, and a five-arm randomized controlled trial design as testable counter-proposals.