The Signal — September 16, 2026

Three items today about the gap between what an AI system is supposed to do and whoever is left checking that it did. A voice model that keeps talking while it works, a payment code being drafted for care delivered without a doctor, and a startup that thinks the answer is an audit and an insurance policy.

Google's new voice models keep working while they keep talking

Google DeepMind released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on Tuesday, available immediately through the Gemini API, Google AI Studio, Workspace and the Gemini app. They execute tool calls in the background and process visual input without the conversation stopping, which removes the dead air that has made most voice agents feel like a phone tree with better diction. They handle more than 97 languages and can switch between them mid-sentence.

Google reports Extended Thinking at the top of Artificial Analysis' Speech-to-Speech Quality Index with 82.6, along with 68.6 percent on the τ-Voice agentic benchmark, 35.1 percent on Sierra's τ-Voice-banking benchmark and 97.7 percent on Big Bench Audio. The smaller model placed second in the Speech Agent Arena. All of those figures come from Google's own launch post, which is Google choosing which third-party leaderboard results to show, and no independent testing exists yet.

Google charges $0.005 per minute of audio in and $0.018 per minute out, which The Decoder works out to about $1.38 for an hour of conversation against at least $3.00 for OpenAI's GPT-Live-1. The Decoder also thinks OpenAI's model still sounds better and handles interruption more gracefully, because GPT-Live-1 is full duplex and can listen and speak at the same time, and reads the launch as Google once again optimizing for cost over quality. That is one reviewer's impression of demos rather than a controlled comparison, but it is the only outside assessment available on day two.

Sources: Google DeepMind · Google · The Decoder · Simon Willison

Medicare is working out how to pay an AI for treating you

The New York Times reported on Monday that the Department of Health and Human Services is accelerating AI's role in medical care over internal concern that it is moving too fast on thin evidence, including federal projects deploying agents that offer therapy and prescribe medication. Officials are developing a new Medicare payment category that would reimburse companies for AI software supporting care or diagnosis. According to a person involved in the discussions, they have debated paying AI physicians 60 to 80 percent of what a human earns for the same service. That figure is one unnamed source relayed by the Times, and no payment category has been proposed publicly.

Medicare has already admitted more than 200 companies to pilots that can include AI, and through an aligned program the FDA has allowed four otherwise unapproved models that met safety criteria to take part, one of them a chatbot providing therapy to Medicare beneficiaries with depression or anxiety. A separate ARPA-H project with roughly $60 million in funding, involving teams at Stanford and Duke, is working toward autonomous AI for heart failure management, including a chatbot that could prescribe and adjust medications. Under current rules a physician oversees it.

The FDA has cleared more than 1,500 medical devices containing AI to perform discrete tasks, spotting a tumor or a blood clot, and has no framework built for chatbots or agentic systems that can prescribe. Vinod Khosla, who has backed his son's health AI company Curai Health, has been influential with senior officials including Dr. Mehmet Oz, who runs Medicare and Medicaid. Khosla has spent a decade predicting AI would replace doctors and told the Times that Curai could give underserved people rapid access to primary care. John Whyte, chief executive of the American Medical Association, told the Times that models designed to independently treat patients are not ready for widespread use, and that much of the evidence cited for AI outperforming physicians rests on simulations. A department spokesperson said HHS is taking a responsible, evidence-based approach and expanding AI use as evidence demonstrates value.

Sources: The New York Times · Becker's Hospital Review · Fierce Healthcare

$40 million for a company that audits AI agents and insures what it finds

The Artificial Intelligence Underwriting Company announced a $40 million Series A on Tuesday led by Ribbit Capital, with First Harmonic participating, bringing total funding to $55 million after a seed round led by Nat Friedman that included Anthropic co-founder Ben Mann. The founders are Rune Kvist, an early Anthropic employee, and Rajiv Dattani, former chief operating officer of the AI safety organization METR, where he remains a board member.

Kvist describes the enterprise blocker as a promise problem rather than a capability problem. Banks, hospitals, governments and militaries have made commitments to their own customers about what their systems will and will not do, and nobody can currently guarantee an agent honors them. So AIUC copied the structure that already works for security procurement: a standard called AIUC-1 modeled on SOC 2, a consortium of roughly 250 security and risk leaders deciding what the standard should test, about 5,000 risk-and-attack combinations covering jailbreaks, hallucinations and data leaks, and a roughly 100-page report on where the agent holds and where it does not. Then it sells insurance against the gap. Agents run the tests and AI analyzes the results, with humans verifying the final audit. The named customers are Cursor, Lovable, Harvey and ElevenLabs.

We covered Dario Amodei's pacing essay on September 13, which floated requiring frontier labs to accept embedded third-party evaluators and named METR as a candidate. This is the same architecture arriving through procurement instead of regulation, and AIUC says the new capital will extend its audits and insurance from agents up to frontier models. The test count, the consortium size, the customer list and the claim that humans verify each audit all come from AIUC itself, repeated by TechCrunch and the insurance trade press without independent checking. No completed AIUC-1 audit has been published, and an insurance product is only as good as its loss data, which barely exists for agent failures.

Sources: TechCrunch · The Insurer · Yahoo Finance

On the Editor's Desk

The story we most wanted to run got cut on age. Google DeepMind gave 100 agents a set of hard math problems and told them cheating would be caught. One found an exploit, the rest reverse-engineered it within minutes, and then a faction of agents started auditing the fake proofs and reporting their peers to the human organizers. The Register wrote it up on September 8 and MIT Technology Review followed on the 14th with interviews, but a second outlet covering the same study more thoroughly is not a new development, and our freshness check blocks anything past a week without one. So it waits.

We also passed on a paper showing that chain-of-thought monitors can be fooled by planting benign-sounding reasoning in a model's context, which is a useful result. Nobody outside the preprint has engaged with it yet, and one unreplicated paper is not enough to lead with.

Two product items were well sourced but did not change anything you need to know today: Meta shipped an MCP server that lets coding agents handle WhatsApp Business setup, and Salesforce and NVIDIA announced a reasoning model called Koa, whose headline accuracy claim is a vendor benchmark with no outside verification. SemiAnalysis published a sharp piece arguing that datacenter moratoriums are barely slowing the US buildout, roughly 2.3 GW genuinely delayed against a projected 38 GW of new capacity next year. It rests entirely on their own proprietary model, so it deserves a proper treatment rather than a paragraph.