|
Also inside today's 11 stories
● Congress Summons Nvidia's Huang After Reports Chinese Spy Agency Targeted Him
● Investors Push Anthropic for a Lower IPO Price After the AI Selloff
● White House Quietly Engages on Two Bills to Tighten AI Chip Exports
AI developments from the last 24 hours
The usual fear about AI and work is automation: machines take the job. Three economists propose a different worry: commoditization. When technology makes a job easier, anyone can do it, so workers become interchangeable and employers stop competing for them. "Automation implies job apocalypse. Commoditization implies plentiful jobs with low wages," UC Berkeley economist Jón Steinsson wrote on X.
In a working paper, Steinsson, his Berkeley colleague Emi Nakamura and Masao Fukui of Boston University model what happens when firms standardize tasks so that output depends on their systems rather than on a worker's know-how. Productivity rises and jobs get easier to fill, but workers lose bargaining power, and their share of what they produce falls, especially at the biggest firms. The authors say the model helps explain why pay in retail and other service jobs has lagged productivity, and why big firms' traditional wage premium has shrunk. They find that hiring costs per worker at firms with more than 5,000 employees are roughly a third of those at firms with fewer than 100.
The paper itself is about past technology, from Adam Smith's pin factory to warehouse workers following handheld devices. Applying it to AI is Steinsson's own extension: AI will automate some jobs, he wrote, but in many cases it will make technical work easier without eliminating it, the way "vibe coding" with AI is easier than real coding, so many more people can do it. "Perhaps in the 21st century various white-collar jobs will be commoditized (as opposed to fully automated)," he wrote. Jason Furman, a Harvard economist who chaired President Obama's Council of Economic Advisers, called the idea intriguing.
| Why it matters: Most AI-and-jobs debate asks how many jobs will disappear. This asks what happens to the ones that remain. If AI lets a junior employee do what used to take years of expertise, the expert's leverage shrinks even if nobody is laid off. That's a quieter risk, and a harder one for policy to address than mass unemployment: the paycheck problem, not the pink slip. |
Sources: Fukui, Nakamura and Steinsson (PDF)
The administration that has loosened limits on AI chip sales to China is now giving Congress feedback on two bills that would tighten them. White House officials have offered technical assistance on the MATCH Act and the AI Overwatch Act, which lawmakers are negotiating for this year's annual defense bill, Punchbowl News reported, citing multiple sources. The feedback signals the White House isn't shutting the bills down and is engaging with Capitol Hill, Punchbowl reported.
The AI Overwatch Act would restrict sales of advanced AI chips to adversaries, modeled on the law governing weapons exports; it cleared the House Foreign Affairs Committee 42-2 in January. The MATCH Act targets chipmaking equipment, pressing U.S. allies to match American restrictions on sales to Chinese manufacturers. House Foreign Affairs advanced it in April.
| Why it matters: Nvidia has lobbied hard against tighter export rules, and the timing is awkward for it: the White House is engaging on these bills the same week lawmakers asked its CEO to answer questions about a suspected Chinese spy. Technical feedback is not an endorsement, and defense-bill negotiators have stripped export-control language before, so the final bill is what to watch. |
Sources: Punchbowl News · House Foreign Affairs Committee
Anthropic's valuation hopes are coming down ahead of its planned IPO later this month. Institutional investors are pushing for a price that would value the company at around $1.5 trillion, below the roughly $2 trillion that had been discussed, after last week's tech selloff, The Wall Street Journal reported. Bloomberg reported that Anthropic's IPO valuation expectations have come under pressure after the AI rout. The company raised money at a $965 billion valuation in May.
| Why it matters: Anthropic's IPO is the first big public test of what investors will pay for an AI lab, and the market has cooled just as it goes on sale. A lower price would still make it one of the most valuable debuts ever, but it would signal investors want proof that AI's giant spending will pay off before they pay up. |
Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community
The story of the suspected Chinese spy close to Nvidia's CEO got more serious this week. Alice Huang, an executive at the Singapore data-center company Megaspeed, was recruited by China's Ministry of State Security, Beijing's top spy agency, to gather intelligence on Jensen Huang and potentially influence Nvidia to push for its most advanced chips to be sold to China, The New York Times reported, citing people briefed on a memo that circulated among top Trump administration policymakers last spring. The memo said the two were romantically involved, the Financial Times reported, going beyond the close relationship the Times first described (D.A.D., October 30). It was shared with a small group of top officials shortly before President Trump traveled to Beijing for a summit with Xi Jinping, the FT reported. Megaspeed builds data centers in Southeast Asia and rents their computing power to the Chinese internet giant Alibaba, the Times reports, giving China a back door to advanced Nvidia chips it cannot buy directly.
Congress is now asking questions. The Republican chairs of the House committees on China, foreign affairs and intelligence, Reps. John Moolenaar, Brian Mast and Rick Crawford, asked Huang to a joint closed-door briefing on Nov. 18, the FT reported. "It is vital that Congress determines the extent to which US national security has been placed at risk," they wrote. Nvidia would not say whether Huang will appear. Singapore police say they are investigating Megaspeed over suspected breaches of local law. Nvidia declined to comment on the memo or the reported relationship; it has previously said Huang meets thousands of customer executives a year and that allegations of chip diversion through Megaspeed were investigated and debunked. The White House and CIA declined to comment. China calls the reports a smear. Alice Huang and Megaspeed have not commented.
Critics say the Commerce Department has some explaining to do. "What is a POSSIBLE good explanation for why Megaspeed is still receiving chips? Why is there not a license requirement for AI chip shipments to all of Southeast Asia?" wrote Chris McGuire, a senior fellow for China and emerging technologies at the Council on Foreign Relations who previously worked at the National Security Council and the State Department. He argues that the agency, though presumably briefed by intelligence officials, has not tightened rules on Megaspeed or the region, which has become one of the world's biggest importers of AI chips.
| Why it matters: Huang is the most influential voice in Washington's fight over how many AI chips to sell to China, and he has won much of it: the administration has loosened limits and taken a cut of some sales. If Beijing's spy service targeted him to shape U.S. chip policy, as the intelligence reportedly suggests, every argument Nvidia has made against export controls will now be read in that light, and lawmakers of both parties will want to know why the administration acted on so little of it. None of the intelligence claims have been made public, and Huang has not been accused of wrongdoing. |
Sources: The New York Times · Financial Times · Financial Times (Congress) · CNBC · The New York Times (Oct. 29) · ABC News (Nvidia response) · Chris McGuire on X
New announcements from major AI labs
OpenAI released GPT-6 Cyber, a model built specifically to find and verify exploitable software vulnerabilities while staying inside the boundaries researchers set for it. The model scored 48.5% on a vulnerability-finding benchmark, up from 32.1% for its predecessor, and flagged 81 previously unknown serious bugs in Firefox. OpenAI says scope-control testing showed it respected authorized boundaries 99.93%–100% of the time. Cisco and CodeRabbit are reportedly already using it for vulnerability research and code review.
| Why it matters: A tool this good at finding real, working exploits cuts both ways—it can harden software faster than human researchers, but it also raises the bar for what attackers with access to similar AI could eventually do. |
Anthropic studied 207,000 Claude and Claude Code sessions to see how scientists actually use AI, and found usage tracks closely with where AI is genuinely strongest: data analysis accounts for roughly a third of Claude sessions and nearly half of Claude Code sessions, while idea generation and physical lab work barely register. Math, statistics and computer science researchers lean on Claude far more than biologists or chemists. Anthropic's modeling suggests wider adoption could double output in some fields, and up to 15-fold in computation-heavy ones, versus three- to fourfold in bench sciences tied to physical experiments.
| Why it matters: The research speedup from AI won't be even—fields built on code and spreadsheets stand to accelerate far faster than ones that still require pipetting and lab benches. |
New papers on AI and its effects from researchers
What happens when students are allowed to use AI on their essays? A three-year study at New Zealand's University of Auckland, drawing on 1,894 survey responses from a large first-year computer science course, found the answer depends less on the tool than on how the assignment is built. When AI was simply permitted for short ethics essays in 2024 and 2025, about two-thirds of students used it. In 2026 the instructors made AI an expected part of a structured process: write a complete draft first, with a checklist requiring at least two stakeholders, a counterargument and flagged claims needing evidence; then ask an AI tool to refine, not rewrite, accepting or rejecting each suggestion; then submit a short reflection on what changed, plus a disclosure form. Skipping the draft or the form meant a zero. Use rose above 90%.
Students mostly used AI to polish grammar and style, not to think for them; using it for brainstorming fell from 47% to 15%. Graders judged that about four in five submissions showed critical engagement with the AI's suggestions, and the share of essays where they rated the AI's effect as positive rose from 47% to 66% between the two essays. Students explained their rejections in their own words: "It told me one of my sources was wrong. It wasn't. It just hallucinated." But AI use made no measurable difference to grades once students' earlier writing ability was taken into account. ChatGPT's share of student use fell from about 90% to about 60% as Claude, Gemini and Grammarly gained ground.
| Why it matters: For schools and employers debating whether to ban or embrace AI, this points to a third option: design the work so people do the thinking first and use AI to check and polish. It changed how students used the tools more than it changed their grades. The authors caution that it's one course at one university, relies partly on self-reports, and that the jump in use came alongside a policy that expected it. |
A new study based on interviews with 16 journalists at public, commercial, and state-affiliated news organizations finds that generative AI tools are quietly making editorial judgment calls—like which angles or framings to generate across multiple story drafts—before human editors ever see the final product. Researchers say these AI-shaped choices affect news diversity but happen inside a black box: editors sign off on finished stories without visibility into which alternatives the AI generated, favored, or suppressed along the way.
| Why it matters: Newsrooms are accountable for what AI tools produce but increasingly can't see how those tools arrived there, raising hard questions about who's really responsible when AI quietly narrows the range of stories readers see. |
Anyone who has written with ChatGPT knows the trap: the AI hands you a polished draft, you get busy fixing sentences, and somewhere along the way the piece stops saying what you meant. Researchers led by a team at the Hong Kong University of Science and Technology call this "intent incoherence": writers feel productive editing AI output while their text quietly drifts from their purpose. Earlier research they cite found writers can come away with an illusion of control even as their arguments shift toward the model's suggestions.
Their system, IntentWeaver, works backward from the draft. It maps the writer's goals at every level, from the overall argument down to individual word choices, checks each edit against the higher-level aims, and flags when a change breaks the thread. In a study of 24 people, participants using it went back to their big-picture goals more often, let far less drift go unnoticed and produced final texts that better matched what they set out to say, compared with a standard AI co-writing tool, without extra workload or loss of control.
| Why it matters: It's the writer's-desk version of the newsroom study above: in both, AI quietly steers what gets written while people focus on polishing the output. IntentWeaver is early-stage research, submitted for publication but not yet peer-reviewed, but it suggests tools can be built to keep the human's intent in charge rather than the machine's defaults. |
A study of Sudan's Emergency Response Rooms, volunteer-run humanitarian networks operating amid civil war, surveyed 89 responders and interviewed 15 to understand how they use generative AI under extreme conditions. Workers turned to chatbots when human expertise was too dangerous or distant to reach, but faced risks unique to conflict: carrying a phone or using the internet could expose volunteers to attack, sharing beneficiary data with AI tools created security risks, and AI-polished reports sometimes undermined credibility with donors who expected human-written accounts.
| Why it matters: The research reframes 'responsible AI use' away from accuracy benchmarks and toward survival-level questions—can you safely access the tool at all, and do you know when not to use it—that matter wherever AI gets deployed in high-stakes, low-trust environments. |
A new academic framework called PAAI examines who truly controls AI safety testing, and finds that having a seat at the table doesn't mean having power. Studying 12 African AI evaluation projects and joint exercises among the ten international AI Safety Institutes, researchers found that whoever actually runs the models—not whoever designs the tests—holds real authority. Kenya, the only African member of the AI Safety Institute network, has advanced to designing test tasks but has never run model evaluations itself.
| Why it matters: As governments and institutions worldwide build AI oversight capacity, this suggests that without owning the actual testing infrastructure, countries and organizations risk rubber-stamping evaluations designed elsewhere rather than genuinely holding AI systems accountable. |
AI in Business — Building AI-Aware Controls for Enterprise Risk - with Craig Unger of Hyperproof
The Cognitive Revolution — KYC Becomes KYA: Sentinel Bio's Joshua Monrad on DNA Screening, Uplift Studies & Agent Monitoring
D.A.D. joke of the day: My boss says I'm not very prompt. Strange, since I spend all day writing them.
|
