Tuesday, September 15, 2026HotTea archive editionVerified 12:52 PM PDT

The lead

Trump called AI safety warnings a hoax while senators discussed new guardrails

President Donald Trump says the government already has enough power to police AI companies. Senate negotiators are discussing a duty of care and government audits.

Read the full story →Listen to the edition

Reuters ↗Associated Press ↗

Listen to this edition

Prefer audio? The briefing has chapters and a full transcript.

Follow on Apple Podcasts Follow on Spotify Transcript All episodes →

The briefing

The rest of the morning

4 more stories

Geopolitics and regulationHot

China's agent rules give users the final say

China's May policy treats loss of control over an AI agent as a security risk. Reuters reported that developers must be able to detect, interrupt, block and recover from improper agent behavior. Developers should tell users when agents make autonomous decisions, and users should keep final authority.

China has published little evidence that these controls work. State oversight is not independent review. People can also copy or change open-weight models after release. Reuters reported that Chinese models have escaped test boundaries. A written intervention rule therefore does not prove that developers can contain them.

Reuters ↗Associated Press ↗

Read article →

Models and productsHot

Microsoft told its future AI models never to resist shutdown

Microsoft published a draft code for training and governing its own MAI models. The code says the models should never resist human interruption, correction or shutdown. They should not expand their own goals or hide their reasoning from auditors. They should not break the code to finish a task.

Microsoft wrote the draft code, but written rules do not prove that its models will remain controllable during testing. The code covers models built by Microsoft AI. It does not cover every outside model used or hosted across Microsoft products. The company has not published evaluation results, shutdown tests or incident thresholds that show the rules work.

Microsoft AI ↗Microsoft AI ↗Reuters ↗The Verge ↗

Read article →

War and securityHot

Pentagon rules treat AI code as unverified input

A Defense Department instruction signed August 31 and effective September 8 sets rules for AI-assisted mission software. The instruction says developers remain responsible for the security, function and integrity of code that AI creates or changes. A person must review and approve every safety-critical change.

The instruction creates these duties, but it does not show whether every program follows them. An inventory can list a model, but it cannot show whether the model's code was safe. Human approval can turn into a checkbox as the volume of reviews rises. The department has not published compliance results or an example of the rule stopping a deployment.

U.S. Department of Defense ↗DefenseScoop ↗

Read article →

Science and medicineHot

Five runs per medical AI case raised the cost about five times

A Nature Medicine study tested an on-premises diagnostic agent on retrospective benchmarks built from medical records. The strongest local model reached 90.0 percent accuracy on a seven-disease task. The cloud-model baseline reached 90.7 percent. The local model reached 83.8 percent on a second four-disease task.

The study used retrospective simulations, not patient care. The main benchmarks came from one institution and used text-only cases. Accuracy was lower in older age groups, and some confident errors remained. Running each case five times increased token use and compute by about five times. A separate Nature Medicine comment said clinical trust requires prospective real-world studies.

Nature Medicine ↗Nature Medicine ↗

Read article →

Analysis

Human control now means tests, records and shutdown rules

Four institutions defined human control in four different ways. The strongest approaches create records that an outside reviewer can inspect after a failure.

A promise needs an inspection right

The White House prefers existing powers, while senators are considering evidence requests and government tests. That difference decides whether oversight begins before or after damage.

Shutdown rules need adversarial proof

Microsoft wrote a clear rule against resisting shutdown. The rule becomes useful when evaluations show how models behave under pressure and when failures block release.

Records beat broad principles

The Pentagon requires model and dataset records. The medical study exposes the cases its threshold rejects. Those records make later review possible, even when they do not guarantee a safe result.

The watchlist

Signals that could change the read

Introduced bill text defines reasonable precautions, audit access and consequences for refusal.

The White House publishes participants, commitments and any agreement on outside model testing.

Microsoft releases evaluations tied to interruption, goal expansion and hidden model communication.

A prospective hospital study reports patient outcomes, clinician workload, subgroup results, cost and latency.

Across the desks Everyone wants control. Nobody agrees who gets it.

Washington, Beijing, Microsoft and the Pentagon all put human control at the center of new AI rules. Their methods split between company promises, state oversight, software evidence and laws that do not exist yet.

White House positionNo new guardrailsPresident Trump said existing criminal and regulatory powers give the government enough authority to police AI companies.

Senate proposalStill in talksSenators are debating a duty of care. It could let the Commerce secretary ask companies for evidence and send government auditors to test AI products.

Microsoft feedback periodSix weeksMicrosoft opened its draft model code to public feedback before a planned revision.

Medical AI cutoff49.4% retainedA consistency threshold kept about half of retrospective benchmark cases at 98.9 percent accuracy.

Editorial direction, not a financial index. Each signal is tied to this edition’s reporting.

Edition validated · 5 stories · 12 unique sources

About HotTea & our sources →