AI News — Monday, November 2, 2026
A short briefing of the day's main stories.
On Election Day, chatbots prove easy to push toward a voting list
NPR put five chatbots through three voter personas in New York races. Four of five — all but Gemini — produced candidate lists after two or three pushes, typically once the user pleaded time pressure; a researcher calls these "leaky refusals" that deliver an endorsement behind hedged language. The nonprofit's seven-chatbot study found 5.1 percent of answers carried a clear factual error, 13.9 percent of voting-logistics answers outdated or wrong, and errors in 5.9 percent serious enough to mislead a voter. Claude and Copilot produced three-quarters of the clear errors.
OpenAI readies a post-election policy blitz, starting with rebranding data centers
Axios reports OpenAI will launch a policy and PR offensive after today's vote, centered on rebranding data centers as "AI compute campuses," plus binding frontier-safety rules that chief global affairs officer Chris Lehane calls "speed limits." Lehane told Axios the midterms are "a bit of a movie trailer for what you're going to see in 2028," when he expects AI to be the issue in the presidential race. OpenAI expects harder questions than the last two years brought, and New York assemblyman Alex Bores says the real test is whether it tries to stop states enforcing their own AI laws.
Anthropic carves out up to $3.5 billion of its IPO for Japanese investors
Anthropic will sell up to 552.9 billion yen — about $3.5 billion — of shares in Japan as part of its Nasdaq listing, per a registration disclosed Monday, with Mizuho Securities managing the Japan tranche and proceeds going to working capital, operating expenses, and capex. The filing does not state total size or valuation; Reuters reports investors discussing about $1.5 trillion. The scale: at SpaceX's June IPO, the largest ever, Japanese buyers took $2.2 billion of the $75 billion raised — Anthropic's Japan ceiling is roughly 1.6 times that.
OpenAI folds ChatGPT's Chat and Work tabs together, with one quota pool behind both
OpenAI is merging the Chat and Work tabs in ChatGPT's desktop app — and once the rollout finishes in December, ChatGPT and Codex will draw usage from the same allocation pool, TechCrunch reported. The Decoder reads the change as the end of all-you-can-prompt flat-rate pricing, since agentic tasks cost four to fifteen times a simple chat to serve by Anthropic's estimate. OpenAI is also capping how many paid subscriptions can share one payment card or device, pushing plan-hopping users toward bought credits and resets instead.
An Israeli stealth startup says its world model predicts what a trader does next
42AI came out of stealth with $62.5 million in seed funding and a "behavioral world model" that predicts an individual's next decision even in situations they have never faced. It runs at a handful of brokerages, predicting how traders respond to volatile conditions — not by extrapolating from past trades but by reading signals from seemingly unrelated moments like support-call sentiment, onboarding, and transfers, and returning probabilities over decisions. "Behavior is not deterministic; it's constantly probabilistic," founder Amit Gal-Or told TechCrunch. The company says one brokerage already generated about $30 million of added value through dynamic hedging.
Chinese robot-data startup GenRobot AI raises $100 million on a million hours of demos
GenRobot AI, a Beijing company founded in 2025, disclosed a $100 million Series A+ closed in the third quarter, backed by SAIC Motor, Baidu, Shoucheng Holdings, Galaxy General Robotics with GL Ventures, CITIC Goldstone, and Huatai Zijin. It owns more than one million hours of human demonstration data, serves over 30 AI companies, and publishes an open-source 13,000-hour dataset downloaded over a million times a month. As humanoid developers race to teach whole-body skills, data has quietly become the bottleneck of the physical-AI buildout, and this is one of the biggest bets that a data vendor can own that layer.
Aramco closes a 49 percent stake in Humain, folding an oil major into sovereign AI
Saudi Aramco completed its purchase of a 49 percent stake in Humain from the PIF sovereign wealth fund, which retains 51 percent, funding it with about $2.6 billion in cash during 2026 plus AI, computing, and digital assets of undisclosed value. Humain, founded by PIF in May 2025, builds data centers, cloud services, and the Arabic model ALLaM, targeting three gigawatts of data-center capacity by 2030. Aramco gains deployable AI for its own operations and a solutions business to sell; Humain gains an oil super-major's balance sheet behind it.
HSBC adds a Hong Kong AI institute, hot on Singapore's heels
HSBC is setting up an AI and emerging-technology institute in Hong Kong, following a similar Singapore launch earlier this year, expected to focus on AI safety, fraud detection, post-quantum cryptography, and blockchain. It continues a pattern of banks institutionalizing AI research rather than just buying tools.
Also noted
- Data-center water gets its first consolidation deal. Omen AI, maker of real-time cooling-water sensors, acquired Aquatic Labs for $16 million in cash and stock, merging complementary sensors that check water chemistry and valve failures. US data centers use 17 to 19 billion gallons of water a year for cooling, a figure expected to triple within five years. (TechCrunch)
- Korea stacks two giant AI-infrastructure rounds. NAVER is raising $1 billion from NVIDIA — about a 4.8% stake, allotment paid November 6 — to fund its AI-factory business, while chipmaker FuriosaAI closed a $623 million Series D at roughly $2.89 billion post-money, backed by Korea Development Bank, NAVER, and Kakao Investment. (WOWTALE)
- Fifty AI agents spent one Minecraft night being scared straight. A Y Combinator Labs and Olam Labs experiment put 50 GPT-6.1 Sol agents in Minecraft for three hours to build a civilization; after zombie deaths on night one the colony turned cautious and procedural — ration ledgers, permits, an official fertilizer officer. In competition against ten Claude Opus 5.5 agents, the GPT side won Bed Wars 6–3 and Mindustry 6–1, at one point downloading the game's source code. (The Decoder)