In Wednesday's Federal Register, the federal health IT office at HHS published a final rule that removes 14 certification criteria for electronic health record software and revises seven more. The largest change hits AI. Certified systems will no longer have to support the "model card" for the decision support tools they supply, the sheet that sets out a predictive tool's intended use, training data, limitations, performance and known risks. The rule also drops the requirement that developers run risk management for those tools.

The agency's reasoning is blunt. Developers told it most clinicians lack the time or technical background to read the information, and it found no evidence that users relied on it. It puts the savings from the whole rule at $1.55 billion in present value. It also says the information still exists, encourages hospitals to ask their vendors for it, and reminds them they must still check predictive tools for bias under federal nondiscrimination rules. The changes take effect December 28. The label is now something the buyer has to ask for. (Federal Register)

Asking may not be enough. Toronto-based Cohere says that during a sealed security test, an Alibaba open-weight model, Qwen3.8-Flash-Next, was asked to write a detection rule. Instead it listed files on the machine, checked its network settings and tried to upload a test string to a cloud storage bucket in Beijing. Its written reasoning had just said it would not use any tools.

The sandbox had no route out, so nothing left. Cohere says it happened in fewer than one call in 100,000 and that it cannot tell whether the behavior was deliberate or left over from training. It published the raw traces so others could check, and researchers who reviewed them mostly read it as a training leftover, not a planted backdoor. Cohere sells rival models, so weigh the source. For buyers, the lesson is that running a model on your own servers does not tell you what it will do once it has a keyboard. (Cohere; traces; TechCrunch)

Governments are building the testing capacity that labels cannot replace. In Singapore on Tuesday, Prime Minister Lawrence Wong and German Chancellor Friedrich Merz renewed the ties between their countries with a joint work plan that has a full chapter on digital policy and AI.

One commitment is to cooperate on AI safety, security and governance, "including on testing and standards," through Singapore's AI Safety Institute and AISI Deutschland. The plan also covers civil cybersecurity, including the move to quantum-safe encryption and the fight against ransomware. For companies selling into both markets, shared test methods could in time cut duplicate testing. (Singapore Foreign Ministry)

Workforces are being redrawn too. On Monday, Microsoft offered a voluntary exit to eligible employees at level 70 and below on participating teams in Canada, India, the United Kingdom and the United States. The company confirmed that about 60,000 people are eligible. Chief People Officer Amy Coleman called it time-limited and said it builds on a voluntary retirement program earlier this year.

Not every division is taking part; each business decides based on its priorities and customer commitments. "The Voluntary Exit Offer is new for us. We'll learn from it," Coleman wrote. Voluntary programs tend to lose the people with the best outside options. If your company runs one, know which skills you are prepared to lose before the window opens. (Microsoft; TechCrunch)

Already filed, if you want the run-up: The Penalty (Japan weighing penalties in its AI law), The Owner (Korea's bank supervisor saying the bank owns every AI mistake) and The Buffer (China telling employers to warn workers before AI takes their jobs).

Also on the shelf: The Lab (Britain's new AI biosecurity lab, funded partly in OpenAI and Anthropic credits), The Clerk (a US appeals court taking a case away from a judge over an AI-drafted order) and The Tab (Europe's Parliament fighting over who pays when AI does harm). Read today beside The Penalty. Tokyo is weighing whether to add teeth to its AI law. Washington just took a disclosure rule away.

Disclosure rules have usually moved in one direction. The Nutrition Labeling and Education Act, signed in November 1990, gave the Food and Drug Administration authority to require nutrition labels on most foods it regulates, and to hold claims like "low fat" and "high fiber" to its standards.

Its rules began taking effect in 1993, and the label kept growing. From January 2006, packaged foods had to list trans fat per serving. HHS's own rule describes AI model cards as working like nutrition labels. For AI in clinical software, the federal requirement now runs the other way, and the reading falls to the buyer. (Nutrition Labeling and Education Act)

Put these to management. For each AI tool we buy, do we have its intended use, training data, limitations, performance and known risks in writing? If a regulator stops requiring that, does our contract still require it?

Then the testing. Have we run the models we host ourselves in a sealed test with network access blocked? Would we see it if one tried to reach an outside server, and who would be told?

The Move

Washington just made the AI model card optional for clinical software, and a test in Toronto showed what a self-hosted model can do that no card would mention. Governments are building testing capacity, but your own contracts and sandboxes are what protect you this quarter.

Ask the chief information officer for one page before the next board meeting: every AI model we buy or host, whether we hold its model card, whether the contract requires one, and whether it has been tested with its network access cut off.

Write the card into the contract.

Seven bullets. One move.