AI news · Thursday, October 29, 2026

Google releases Gemini 4 Carbon, but you can't use it yet

Google announced Gemini 4 Carbon, calling it its strongest model yet, with big gains in finding and fixing security flaws (88.2% on a vulnerability-finding test called SEC-bench Pro). It costs $2 per million input tokens and $10 per million output tokens, same as predecessor Argon. For now it's only going to cybersecurity defense partners, not the public.

The catch: Two frontier models in two weeks shows labs racing on release speed, even while the public still can't touch either one.

via Google DeepMind

Anthropic's IPO filing admits its AI biology hype needs time

Anthropic CEO Dario Amodei has said AI could cure most diseases within 5-10 years and deliver a 'deluge' of new drugs. But the company's new IPO filing is more cautious, admitting it lacks experience in physical lab work and that any AI-found drug candidates still face the same years-long testing all new medicines require. The filing also walks back earlier claims about a gene-editing enzyme Claude reportedly discovered.

The catch: SEC filings carry legal liability for overstatement, which is likely why Anthropic's investor pitch sounds so different from its CEO's public claims.

via Wired

xAI offers to buy homes near its data center if owners drop lawsuit

A subsidiary of Elon Musk's xAI is offering to purchase up to 418 homes near its Southaven, Mississippi data center and power plant, but only if owners sign away their right to sue over noise and vibration from the facility. Three residents already filed a lawsuit over the issue; 18 homeowners have reportedly agreed to sell.

The catch: Buying out complaints rather than fixing the noise sets a template other data center operators may follow as opposition spreads nationwide.

via The Verge

Texas voters, left and right, unite against AI data centers

A new coalition called We the People Texas, representing an estimated 1 million voters across the political spectrum, is fighting data center construction in Texas. A recent poll found 57% of Texans oppose new data centers, including 43% of Republicans and 75% of Democrats, citing water use, tax breaks, and new gas plants.

Zoom out: When data centers become a top-tier election issue in a reliably red state, it signals the AI buildout has outpaced public patience nationwide.

via The Verge

Russian-linked fake videos target two Senate candidates

AI-made or AI-edited videos falsely attributed statements to real people to spread disinformation about Democratic Senate candidates Sherrod Brown and James Talarico days before the midterms. Researchers tied the accounts sharing the clips to the same Russian network, Storm-1516, that targeted Kamala Harris and Tim Walz in 2024.

What's next: X removed the original posts, but copies with hundreds of shares stayed up, showing takedowns can't keep pace with AI-made political fakes.

via Fast Company / Associated Press

UK watchdog may probe asylum decisions aided by ChatGPT

A UK government inspector is considering investigating the Home Office's use of AI tools, built on OpenAI's GPT-4, to summarize asylum interviews and research country conditions. A digital rights group found the error rate in these summaries is higher than the 9% the government reported in its own pilot study.

The catch: A UK judge already found one asylum refusal was based on an AI 'hallucinated' document, showing the stakes when errors affect life-or-death decisions.

via Wired

Google's AI search dodges who won the 2020 election, in the US only

Asking Google's AI Mode who won the 2020 presidential election gets a dodge: it tells US users to go check the Federal Election Commission's website instead of just answering. The same question gets a direct, correct answer when asked from Taiwan, Russia, France or Canada. Google says it's 'looking into' the glitch.

The twist: Google blamed overly cautious safety filters, but the fact the caution only applies inside the US raises harder questions about what triggered it.

via Ars Technica

NVIDIA's free transcription AI understands 40 languages

NVIDIA released Nemotron 3.5 Transcribe, a free, open-weight AI model that transcribes speech in 40 languages, translates 31 of them, and labels up to eight different speakers in recordings up to an hour long. It tops leaderboards for both accuracy and speaker labeling among open models, and can run locally on compact hardware like a Jetson device.

For you: A free, locally-runnable model this capable lowers the bar for small businesses, like a clinic or law office, to build their own transcription tools without sending audio to a cloud service.

via Hugging Face / NVIDIA

Microsoft made $10 billion from OpenAI in one quarter

Microsoft reported $10 billion in revenue from its commercial deals with OpenAI in the quarter ending September 30, about 11% of Microsoft's total $92.3 billion quarterly revenue. OpenAI shares roughly 20% of its revenue with Microsoft under their partnership, which runs through 2030. Microsoft's overall revenue grew 19% year over year, its 10th straight record quarter.

Follow the money: Even as OpenAI diversifies to other cloud providers, this shows how financially entangled the two companies remain heading into OpenAI's own expected IPO.

via TechCrunch

A judge says AI should decide some court cases

Geoffrey Vos, one of England's most senior judges, is retiring and using his final days in office to argue AI should adjudicate simple civil cases like personal injury claims, freeing up courts for complex disputes. He's seen courts in Shenzhen, China already using AI heavily to decide civil cases and came away impressed. Vos warns the risk is real: lawyers have already been caught citing fake, AI-invented case law.

Zoom out: A sitting senior judge publicly endorsing AI-decided verdicts, not just AI-assisted paperwork, marks a real shift in how seriously courts take the technology.

via Wired

What is a 'reasoning model'? Explained simply

A reasoning model is an AI chatbot that works through problems step by step before answering, instead of blurting out the first guess. Think of scribbling math on scratch paper versus answering instantly. For example, asked to plan a road trip with five stops and a budget limit, it works through the constraints one by one instead of just listing cities that sound right.

For you: Knowing this helps you pick the right tool: a quick question doesn't need a reasoning model, but a complex spreadsheet formula or trip budget might.

Original explainer

AI emergency alerts are coming, but humans write them badly

Emergency managers are increasingly turning to AI chatbots to help write disaster warnings, after years of confusing human-written alerts (one 2016 case had neighboring California counties calling the same wildfire a 'voluntary' and 'mandatory' evacuation on the same day). But a chatbot already produced an inaccurate storm-risk graphic for an Alabama county during Hurricane Isaias. Most agencies officially ban generative AI, but experts believe it's already widely used anyway.

Zoom out: The tool meant to fix confusing emergency communication can itself generate confidently wrong information at the exact moment accuracy matters most.

via Wired / Grist