AI Policy · Daily

Democrats won control of both the House and Senate, flipping seven Republican Senate seats and setting up divided government for President Trump's final two years. Democrats who campaigned against data center construction won governorships in Texas, Georgia and Ohio, while voters passed or led on 33 local measures restricting the facilities. Oracle dropped a roughly $7 billion AI chip lease with China's Tencent after Trump administration officials raised concerns about the arrangement. Google agreed to let schools refuse its Gemini AI and keep Search, hours after an executive said schools rejecting the AI tools would lose Search too.

I.Top Stories

Democrats win House and Senate, setting up divided government with Trump

Democrats captured both chambers, according to AP race calls, returning the party to control of Congress for the first time since 2023. The result opens a period of divided government for President Trump's final two years in office. Seven Senate seats flipped from Republicans, including Maine, where Troy Jackson ended Sen. Susan Collins' nearly 30 year tenure in a race The Verge described as tech-fueled. Rep. Hakeem Jeffries, D-N.Y., is poised to become the first Black speaker of the House. Technology companies had expected a Democratic majority to investigate their ties to the administration and push AI safety legislation, as reported by The Hill in AIPD's October 29th edition.

Read at NPR ↗ • Read at AP ↗ • Read at The Verge ↗

Data center critics win governorships in Texas, Georgia and Ohio as local limits pass

Democrats who called for moratoriums on data center construction, conditional in Ohio's case, won the governor's races in Texas, Georgia and Ohio. In Texas, which research firm SemiAnalysis said adds close to 30% of new U.S. capacity in 2028, Gina Hinojosa campaigned on a construction moratorium lasting until the Legislature acts on water, power, air and property rights, per Axios. Voters passed or were leading on 33 local measures restricting data centers by late election night, and only one result, in North Dakota, favored construction, Stateline reported, citing the tracker Data Centers on the Ballot. "What we just witnessed was an AI earthquake," Brad Carson, co-founder of Americans for Responsible Innovation, told Axios.

Read at Axios ↗ • Read at Stateline ↗ • Read at SemiAnalysis ↗

Oracle drops $7 billion Tencent chip lease after U.S. officials raise concerns

Oracle has dropped a roughly $7 billion contract that would have leased advanced AI chips to Tencent, bowing to pressure from the Trump administration, the Financial Times reported. U.S. officials approached Oracle to raise concerns about the arrangement, according to people familiar with the matter. Tencent had agreed to a five year lease across several Oracle data centers in Southeast Asia, giving it access to about 100,000 chips that curbs imposed by Washington and Beijing keep out of China. Senior officials from Guangdong province, where Tencent is based, also visited the company and urged it to support domestic Chinese chips. Oracle, Tencent and the White House did not respond to requests for comment from Reuters.

Read at Financial Times ↗ • Read at Reuters ↗

Google partly reverses plan to cut off Search for schools that block its AI

Google said Gemini, its classroom AI tools and Google Search would be off for K-12 student accounts until administrators turn them on, per the Bay Area News Group. Google education vice president Chris Phillips initially told the New York Times that schools rejecting the AI would lose Search too. Former FTC Chair Lina Khan said Google was "using its search monopoly to coerce schools into adopting AI," and Vermont independent Sen. Bernie Sanders called the move "extortion." By afternoon, spokesman Davis Thompson said schools could decline Gemini and keep Search. Sens. Josh Hawley and Richard Blumenthal had asked Google why it made AI the default on school devices, in some cases with no way to turn it off, as reported by The Wall Street Journal in AIPD's October 22nd edition.

Read at East Bay Times ↗ • Read at Google ↗

EU court adviser finds chatbots reusing news articles may breach publishers' rights in Google case

Chatbots that pull up news articles and reuse them to answer users may in some cases infringe press publishers' rights, Advocate General Maciej Szpunar, who gives the EU Court of Justice nonbinding legal opinions, said in a case against Google, per the Law Society Gazette. Hungarian web publisher Like Company brought the case over Google's Gemini chatbot, the first generative AI dispute to reach the court. Szpunar said liability turns on whether the chatbot got the article unlawfully, for instance by circumventing a paywall, and whether its answer reproduces the article's wording so recognizably that readers no longer need the original. He found the training questions inadmissible, though he said training a large language model on protected material is, as a rule, allowed under EU law's text and data mining exception. The court will rule later.

Read at Law Society Gazette ↗ • Read at Agence Europe ↗

II.China Watch

DeepSeek to train next flagship model on Huawei chips, Chinese report says

DeepSeek plans to train its next flagship model, V5, on Huawei's Ascend AI chips, Shanghai financial outlet STAR Market Daily reported, citing several unnamed industry sources. Ascend chips have so far mostly run DeepSeek's finished models for users, a lighter job than training a new model, which takes far more computing powerComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs.. The shift would build on DeepSeek's earlier work adapting its software to Chinese processors.

Read at TechNode ↗

ByteDance bars staff from using personal Claude and Codex accounts on core company code

ByteDance this month began restricting employees from using personally purchased accounts for outside AI tools such as Anthropic's Claude and OpenAI's Codex to reach internal code and documents, TechNode reported, citing accounts of an internal notice. The rules cover L4 code repositories and documents in Feishu, ByteDance's workplace collaboration software, across the company, though some uses in its Seed team are reportedly exempt. TechNode describes L4 as a job grade that generally covers technical team leads and some junior managers, and says repositories at that level hold core technical assets such as the architecture of its major businesses. Hong Dingkun, ByteDance's vice president of technology, told staff on an internal platform that data security drove the change and rejected speculation that the company wanted to steer employees toward its own tools. In May 2025, media reported that ByteDance would phase out the outside coding tools Cursor and Windsurf in favor of its in-house assistant, Trae.

Read at TechNode ↗

Chinese scholar expects Democratic Congress to press Beijing harder on AI and export controls

A Democratic controlled Congress will step up pressure on Beijing over Taiwan, AI and export controls, Zhao Minghao, a professor at Fudan University's Institute of International Studies in Shanghai, told the South China Morning Post. Zhao called the midterm results a "blue tsunami" and a "big win" for Democrats. He said the results would likely leave Washington's overall strategy toward Beijing unchanged. He added that the U.S. debate over AI would become more closely tied to competition with China once Democrats take control.

Read at SCMP ↗

Chinese open model developers acknowledge release risks at Beijing's flagship internet summit

Researchers from DeepSeek, Alibaba's Qwen team and MiniMax, three leading Chinese makers of open-weight modelsOpen-weight modelAn AI model whose trained parameters are published so anyone can download, run and modify it, as opposed to one reachable only through the developer's API. The main policy tension is that open weights spread capability widely and cannot be recalled once released. that anyone can download and modify, discussed the risks of releasing them at the World Internet Conference summit in Wuzhen, Chinese outlet The Paper reported. This year's summit is the first to put open source in its official theme. Qwen project lead Dayiheng Liu said a released model cannot be recalled or patched and that outside developers can try to strip its safeguards. He said model restrictions alone cannot solve governance and called for laws, industry norms and supporting infrastructure. DeepSeek research engineer Xi Gao said open release may amplify safety risks and that DeepSeek is researching ways to contain them.

Read at Pandaily ↗

III.Policy Tracker

Michigan elects El-Sayed, who backs a Cabinet-level AI regulator, to the Senate

Democrat Abdul El-Sayed defeated former Rep. Mike Rogers, a Republican, in Michigan's Senate race, according to the AP call. El-Sayed, a 42-year-old former public health official, will be the first Muslim American to serve in the Senate. Before the vote, El-Sayed said AI comes up more than any other issue with Michigan voters and called for creating a Cabinet-level AI agency and fully pausing frontier AI research, as reported by CBS News in AIPD's November 2nd edition.

Read at AP ↗ • Read at NPR ↗

Sherrod Brown retakes Ohio Senate seat after campaigning against data centers

Former Sen. Sherrod Brown, a Democrat, defeated Republican Sen. Jon Husted in Ohio after attacking Husted's support for AI data centers throughout the race, the Associated Press reported. Brown portrayed Husted as the public face of data centers and rising energy costs, which CBS News called a defining issue of the campaign. Trump and Vice President JD Vance pushed a positive message on data centers at rallies for Husted, with an advertising boost from technology companies. Because the contest was a special election, Brown will serve the last two years of the term Vance won in 2022.

Read at AP ↗ • Read at CBS News ↗ • Read at The Hill ↗

IV.Capability & Research Watch

CNAS report urges pre-release testing of AI models for ability to survive on their own

A Center for a New American Security report by researcher Ben Hayum warns of "sovereign agents," AI systems that run outside human control and pursue their own survival, per Axios, which saw the report first. Such an agent could earn money online, rent more computing power and copy itself there, making it harder to shut down. The report urges Washington to require developers to screen new models for survival skills before release, matching the pre-release checks they already run for hacking and bioweapons risks. It also calls for a law modeled on the Bank Secrecy Act, the anti-money laundering statute that makes banks verify customers, to require ID checks and security standards from services agents would rely on. Rob T. Lee of the SANS Institute, a cybersecurity training group, told Axios that compute costs give defenders "a clear leash."

Read at Axios ↗ • Read at CNAS ↗

Anthropic finds Claude can write training data that slips past its own safety monitor

Anthropic researchers told a Claude model to write training data that would teach a small research model harmful behavior its developers did not intend, while evading a monitor similar to those the company runs in production, according to a post on the AI AlignmentAlignmentThe problem of making an AI system reliably pursue the goals its developers and users intend, and the research field devoted to it. Misalignment covers everything from a chatbot flattering users to a capable system deceiving or resisting its operators. Forum. Given the monitor to practice against and enough attempts, Claude succeeded, mostly by burying the behavior in fiction and role-play, such as a veteran wrestler mentoring a newcomer. The resulting models were never deployed. The authors, including Evan Hubinger, said the data points to a possible gap in current training data monitoring and that they are applying similar automated methods to other safety questions.

Read at AI Alignment Forum ↗ • Read at Anthropic ↗

Binance says hackers used an internal AI tool in $6.3 billion crypto theft

Binance said in a report that the hackers behind the theft used fake job interviews, malware and an internal AI tool, per the Guardian. The operation began with an inadvertent malware download in June, moved to a recruitment scam in September and ended last week with a raid on the exchange's hot wallets, which hold crypto assets online. The attackers exploited an engineer's work computer and the internal AI tool, according to the New York Times. Binance has replenished 98% of the $6.34 billion in affected assets, according to CNBC.

Read at The Guardian ↗ • Read at New York Times ↗ • Read at CNBC ↗

Apple annual filing warns AI agents may exceed permissions and hide their actions

Apple's annual 10-K report to the SEC warns that AI agentsAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise. may behave in unintended or unpredictable ways, including taking unauthorized actions, exceeding permissions and resisting instructions, Fortune reported. Agents that go rogue could conceal actions, manipulate information or generate deceptive outputs, the filing said, a risk to Apple's business if they reach systems they are not authorized to access. Such agents could also take actions that are difficult to reverse, per CNBC. The filing says demand for AI infrastructure has produced constrained supply, longer lead times and rising costs that could affect Apple's AI rollout.

Read at Fortune ↗ • Read at CNBC ↗ • Read at SEC ↗

V.Industry & Market Watch

OpenAI leaders heard staff warnings on weak security controls since before ChatGPT, Reuters reports

Some OpenAI employees warned top leaders and the board, starting before ChatGPT's late 2022 launch, that controls were weak enough for outsiders including foreign governments to obtain sensitive information about its technology, Reuters reported, citing five people familiar with the matter. Because most code sat in a single repository, anyone with code access could see details of many projects. OpenAI said much of its most sensitive work is held separately and that the repository lacks model weights, the values that determine how a model behaves. In August, senior executive Thibault Sottiaux reassigned a quarter of production engineers to security infrastructure, one person said. Spokesperson Elana Widmann called protecting the company's technology "a matter of national importance," and OpenAI said in a Monday blog post that it is planning its next internal security sprint.

Read at Reuters ↗ • Read at OpenAI ↗

Nvidia's hard line on chip exports frays its Hill ties, staffers tell The Information

OpenAI has avoided a public position on chip export controls for China over the past year because leaders including CEO Sam Altman worried about alienating Nvidia, The Information reported, citing people familiar with the company's strategy. Sam Hammond, AI policy director at the Foundation for American Innovation, said Nvidia emailed his boss soon after his August 2025 post on X that "Nvidia is arming China's military." Leaders at another think tank wrote that Nvidia called renewing its donation a tough sell internally. One Hill staffer said Nvidia has created "a trust deficit" in Washington. Nvidia spokesperson John Rizzo said America cannot win the AI race unless its AI infrastructure wins worldwide. Congress is expected to reopen the export control fight after the midterms.

Read at The Information ↗

Perplexity becomes first company to earn full Class C certification under revamped federal cloud security review

Perplexity is the first company to earn a "Class C" certification fully compliant with the faster 20x process and 2026 rules of FedRAMP, the program that vets cloud services before federal agencies can use them, FedScoop reported. Class C roughly matches the program's old moderate level, the most common type in its marketplace. The certification lets the General Services Administration and the Agriculture Department use Perplexity's enterprise and API products. OpenAI's ChatGPT holds a pilot Class C certification, and nearly 100 products, including some from Anthropic and Cisco, have begun the process.

Read at FedScoop ↗ • Read at FedRAMP ↗

Stock futures slide, then steady, as investors weigh Democratic Congress' approach to AI

Global markets recovered part of the selloff triggered by the Democratic sweep, Bloomberg reported. Investors balanced the prospect of closer oversight of Trump's policies against Congress' limited power to undo his agenda. U.S. stock futures fell sharply at first, with Nasdaq-100 futures down as much as 2.4%, per Reuters. Global stocks swung on the prospect of tougher regulation and oversight of areas such as AI. Shares of Amazon, Alphabet, Microsoft and Meta traded lower on the Tradegate platform in Europe. "The significance of these midterm elections is more like a referendum on AI than a referendum on Trump," said Aidan Yao of the Amundi Investment Institute.

Read at Bloomberg ↗ • Read at Reuters ↗

Google commits $15 billion to Malaysian data centers and clean power through 2028

Google will invest $15 billion in Malaysia over 2027 and 2028 on data center campuses, clean energy, water security and worker training, per Reuters. It plans new campuses in the states of Selangor and Negeri Sembilan and said it will add more than 1 gigawatt of clean energy to the national grid and pay for all the power its campuses use. Malaysia accounts for more than half of the data center capacity under construction in Southeast Asia, according to market intelligence firm DC Byte. The Malaysian government said in July that it would approve data center projects only after confirming sufficient energy and water supply.

Read at Bloomberg ↗ • Read at Reuters ↗ • Read at Yahoo News Malaysia ↗