The Signal — September 27, 2026

The United States and China agreed to set up a channel for reporting AI incidents to each other, and neither government has said what would count as one. Separately, documents show that scans from Oxford's Bodleian Library went into OpenAI's training set under a deal announced as digitisation, and Blue Cross Blue Shield says hospitals' AI billing tools added nearly a billion dollars in claims without adding care.

Washington and Beijing agreed to a channel for AI incidents without saying what counts as one

At the close of Xi Jinping's three-day state visit to Washington, the two governments agreed on two AI items: a standing dialogue on the technology's risks and benefits, with the next round due by November, and a bilateral channel for AI-related incidents. The White House fact sheet, released late on September 25, calls the first the "U.S.-China Super Intelligence (SI) Dialogue" and says the leaders agreed to use "super intelligence" in place of "AI," according to Axios. China's Foreign Ministry and the state news agency Xinhua describe the same two items in ordinary AI terms. Both sit inside an eight-point consensus that also covers reciprocal tariff cuts on about $30 billion of goods and a planned memorandum on crisis communications between the two militaries.

Axios reports that Treasury Secretary Scott Bessent pushed for the incident channel in talks before the summit, and some have likened it to a Cold War-era "red telephone". Nothing published so far says what kind of incident would trigger a call, who answers on each side, or what either government would be expected to disclose. Trump drew his own limits on September 26, telling reporters the United States is "not going to be putting on brakes" and, on sharing with China, "I'm not looking to open it up," according to the Associated Press.

The published outcomes also do not say whether incidents inside private companies, such as the OpenAI sandbox escape we covered yesterday, would ever reach a government-to-government line. The two sides have agreed to a schedule for talking and a channel whose rules have not been published. The next round, due by November, will give us a chance to see whether either government puts definitions in writing.

Sources: Xinhua · Axios · PBS NewsHour (AP) · CNBC (Reuters)

Oxford's library deal also fed OpenAI's training set

Internal University of Oxford documents obtained by the Guardian through a freedom-of-information request say material OpenAI digitised from the Bodleian Library was used to "populate the OpenAI training set." When Oxford announced the partnership in March 2025, it described a digitisation project to make historical texts easier for students and researchers to reach, and did not mention model training. Oxford is the only UK member of NextGenAI, OpenAI's program with research libraries that also includes Boston Public Library, Caltech, MIT and the University of Michigan.

By June 2025 the Bodleian had shared 125,000 images scanned from 19th- and 20th-century PhD theses from European and American universities, and a collection of 10,000 16th-century broadside ballads was also scanned. Meeting minutes record staff, including members of the Bodleian's governance committee, raising concerns about reputational risk and about how a deal involving an energy-intensive technology fits the university's environmental commitments. The same minutes discuss possible mass digitisation of the library's 23 million items and an "Ask the Bod" chatbot. The documents do not say which works went into which models.

Oxford told the Guardian the material is modest in scale and out of copyright, that OpenAI's use is not exclusive, and that the Bodleian keeps the rights and will begin publishing the scans openly within the next few months. Oxford rejected the suggestion that the training use was hidden, saying staff had been open that the project would also contribute training data. An OpenAI spokesperson said the company was "proud" to help ensure "the AI models of today preserve the world's historical knowledge for the future." Unlike the purchased books some AI companies cut apart for scanning, the Bodleian's volumes stay intact.

Sources: The Guardian · The Next Web

Blue Cross says hospitals' AI billing tools added $942 million without adding treatment

On September 24, the Blue Cross Blue Shield Association published an analysis of claims from its member insurers, which together cover about one in three Americans. Using 2023 as the baseline, it estimates that hospitals added about $942 million in spending over two years by coding inpatient stays as more medically complex. Roughly 70 percent of that, about $650 million, came from secondary diagnoses that moved a claim into a higher-paying category. These are conditions recorded alongside the main reason for admission.

The association ties the rise to AI coding and documentation tools, which it says more than 60 percent of hospital systems now use to scan lab results and doctors' notes for conditions that can be billed. Its test case was major bowel surgery, where Reuters reports that recorded secondary diagnoses of partial intestinal blockage and too much acid in the body rose 55 percent and 33 percent, respectively, between early 2023 and late 2025. Anemia diagnoses also went up, but blood transfusions did not. "The disconnect between diagnoses and treatment suggests that AI is identifying more billable conditions, not sicker patients," said Luke Chalker, BCBSA's senior vice president of product and data science.

This is an insurer group's reading of its members' claims, and insurers are one side of a long billing fight with hospitals. The association's announcement does not explain how it matched the coding changes to specific hospitals' use of AI tools, so the link to AI rests on timing and the pattern of diagnoses rather than a direct comparison. The New York Times and TechCrunch covered the same study; their reports are not independent checks of it. Hospitals and insurers both use AI in these disputes now. In coverage of the study, Abridge founder Shiv Rao, whose company sells AI documentation tools to health systems, warned of "bots fighting bots, agents fighting agents," while Chalker called the current situation "a completely one-sided blood bath" with insurers on the losing side.

Sources: Healthcare Finance News · Reuters · TechCrunch

On the Editor's Desk

The Washington Post reported that U.S. and Russian diplomats weakened a draft text in this month's UN talks on autonomous weapons, including a requirement that humans review AI-generated targets. The talks ended on September 4, and we could not match that specific change against the public draft documents, so we are holding it until the text is clearer.

Walmart's chief executive published a letter promising the company will not use its AI shopping assistant or digital shelf labels to charge different customers different prices. For now the only check on that promise is Walmart's own testing. We also left out a letter from 26 state attorneys general asking Congress for federal AI rules that do not override state laws, because it is a request that has not yet produced any action.