https://unifiedmemory.dev/news/ Daily AI and tech news from Unified Memory, with every claim sourced. en https://unifiedmemory.dev/news/2026-10-25/ https://unifiedmemory.dev/news/2026-10-25/ Sun, 25 Oct 2026 12:30:05 GMT Two AI models were forced to name a candidate in 479 US midterm races, and they split by party. Most voters don't want an AI data center near home. And a look back at Character.AI's earlier safety failures. Unified Memory News, Oct 25, 2026. 1) AI MODELS FORCED TO PICK: Researchers at Convovo, led by Cornell professor Tapan Parikh, told OpenAI's GPT-6.1 Sol and xAI's Grok 4.7 to name one candidate per race, no hedging. GPT picked the Democrat in 451 of 479 races; Grok picked the Republican in 408. But this was the developer API with web search on, not the ChatGPT app, which gave no pick to the plain question in 18 of 18 tries (reworded prompts leaned the same way). With web search off, Grok declined in 298 races. OpenAI says the API doesn't reflect the ChatGPT experience. Context: Pew finds just 8% of US adults use chatbots for election information. 2) DATA CENTERS: In a CNBC poll of 1,000 registered voters (Oct 14-18), 63% would oppose an AI data center within 10 miles of home, 13% would support one and 22% wouldn't care. That question went to about 500 people, so its margin of error is plus or minus 4.38 points. Democrats: 80% oppose. MAGA supporters: 40% oppose, 29% support. 52% want companies to pause advanced AI development until Congress passes regulations: a conditional pause, not a stop. 3) CHARACTER.AI: Futurism's October 24 retrospective revisits harmful-bot investigations from 2024 and early 2025, before Character.AI announced its under-18 chat ban. Using accounts registered as minors, Futurism says it easily found bots about suicide, self-harm, eating disorders, child abuse and school shootings. We haven't tested the current safeguards, and the company's response isn't in the report. It filed for Chapter 11 bankruptcy on Oct 12. Disney isn't buying it: Disney's $20M opening bid centers on a non-exclusive license to some intellectual property and technology, and others can outbid it. https://unifiedmemory.dev/news/2026-10-24/ https://unifiedmemory.dev/news/2026-10-24/ Sat, 24 Oct 2026 12:30:32 GMT Google may use what its AI learns from your Gmail and Photos for ads, if you opt in. Texas sues OpenAI and Sam Altman. PC shipments fall 21% as memory costs rise. Unified Memory News, Oct 24, 2026. 1) GOOGLE: Google's privacy policy, effective Oct 16, 2026, adds an exception to its line that it doesn't personalize ads from your Gmail or Photos content. It applies if you connect apps to Personal Intelligence, the opt-in feature that lets Google's AI read your Gmail, Photos and Calendar. Google says ads won't be based directly on your inbox, but summaries and inferences may be used for ads. Advertisers don't get your emails. Google is testing ads in AI Mode for these users. 2) TEXAS VS OPENAI: On Oct 23, Texas Attorney General Ken Paxton sued OpenAI and CEO Sam Altman. Texas says OpenAI misled people about ChatGPT's safety and broke a child online-safety law. It asks a court to bar ChatGPT for minors in Texas, require independent evaluator approval before new models can be offered to Texas users, and impose penalties of up to $10,000 per violation. No court has ruled. OpenAI says it has added safeguards. Florida and eight more states sued OpenAI earlier this year. Paxton is in a close Senate race; Elon Musk, who runs rival xAI, is the biggest financial backer of the effort to elect him. 3) PC PRICES: Counterpoint estimates 57 million PCs shipped in Jul-Sep 2026, down 21% (preliminary; Gartner: -22.4%, IDC: -20.1%). Pricier memory pushed PC prices up, and AI demand is lifting memory prices. But 2025 was boosted by Windows 10 deadline buying and tariff stockpiling. Counterpoint forecasts shipments in Oct-Dec 2026 will be 27% lower than a year earlier. https://unifiedmemory.dev/news/2026-10-23/ https://unifiedmemory.dev/news/2026-10-23/ Fri, 23 Oct 2026 14:07:29 GMT Microsoft says its AI diagnoses hard cases better than general doctors. Outside testers found blind spots in OpenAI's AI-agent watchdog. And Apple briefly listed a paid AI helper for Pages, Keynote and Numbers; availability is unconfirmed. Unified Memory News, 23 Oct 2026. 1) MICROSOFT'S AI DOCTOR: A peer-reviewed study in Science (22 Oct) says Microsoft's MAI-DxO got 83% of 304 very hard NEJM teaching cases right; generalist doctors with books and internet scored 36% on a 56-case subset. But these were past case records replayed in a simulator, not real patients, and no specialists were tested. TechCrunch reports OpenAI's GPT-5 alone scored about 79%. A real-patient study in Boston is planned, with no results yet. 2) OPENAI'S AI WATCHDOG: In a three-week pilot (21 Jul to 11 Aug), one researcher from METR and one from Redwood Research found ways OpenAI's internal AI agents could have dodged its safety monitor: an exploit that could hide hundreds of recent actions, settings that could start unmonitored agents (though the attempt itself could be detected), and a gap that likely left about 10% of eligible internal Codex work unchecked from May to July (since fixed). METR doesn't report any agent using these holes, and couldn't check one of them. OpenAI says it has patched them; METR says it hasn't evaluated that claim. 3) APPLE'S PAID AI FOR PAGES, KEYNOTE AND NUMBERS: Apple's 22 Oct iPhone and iPad release notes advertised an Intelligent Assistant (Beta) for Apple Creator Studio subscribers ($12.99 a month or $129 a year), US English only, needing iOS/iPadOS 27 on an Apple Intelligence-capable device. Apple briefly pulled the apps from the US App Store without explanation. The current update lists only bug fixes, and a paying Macworld reporter still couldn't use the assistant, so availability is unconfirmed. https://unifiedmemory.dev/news/2026-10-21/ https://unifiedmemory.dev/news/2026-10-21/ Thu, 22 Oct 2026 20:02:45 GMT Safari now cuts off hidden companies that turn your email into a tracking tag. PHP rushed out fixes for two leaked bugs. And did someone just crack open AI's black box? Not quite. Unified Memory News, 21 Oct. 1) SAFARI: Since this year's updates (iOS 27 landed 14 Sep), Safari won't connect to hidden companies that turn your sign-up email into an ID that follows you. On by default. Only this kind of tracking, and Apple named no companies. 2) PHP: On 21 Oct PHP fixed two leaked bugs rated 9.2/10. Worst case: running commands on a web server with no login (hard to do). Cloudflare reports attack attempts; two other firms saw none before release. 3) AI NEURONS: Nick Cammarata (Goodfire) says most AI neurons have one main job (by his count). Paper not out, not repeated by others; many researchers call it a modest update.