AI News
AI news that matters. Updated daily.
No stories match your filters.
Microsoft Gets Free Access to OpenAI's Models Under New Deal Terms
"We fully plan to exploit it." That was Microsoft CEO Satya Nadella's response to TechCrunch when asked about the company's renegotiated deal with OpenAI - one that lets Microsoft offer OpenAI's technology to Azure cloud customers without paying for it.
OpenAI Lawyers Cross-Examine Musk on Day 3 of Ongoing Trial
Three days into the Musk v. Altman trial, OpenAI's lawyers took their turn at cross-examination. According to Wired's coverage of the proceedings, Musk acknowledged the tension directly, saying "they're gonna want to kill me" - a comment directed at OpenAI's legal team.
Microsoft Reports 20M+ Paid Copilot Users, Pushes Back on Low-Adoption Narrative
20 million. That's the paid user count Microsoft is now citing for Copilot, the AI assistant baked into Microsoft 365 products like Word, Excel, and Teams. The company disclosed the figure on April 29, 2026, alongside claims that engagement - not just seat count - is rising.
Glean Builds a Task-Specific AI Model for Enterprise Search
Enterprise search has been a graveyard for ambitious software projects for two decades. Glean is betting that a model built specifically for workplace knowledge retrieval - rather than a general-purpose large language model bolted onto an index - is what finally makes it work reliably.
Two Words vs. a Plugin: Testing Claude Code's Caveman Against 'Be Brief'
What happens when you test a plugin designed to make Claude Code responses shorter against simply typing "be brief"?
AI Deepfake Ads Are Running on TikTok Using Fake Celebrity Interviews to Steal Data
Scammers are placing Taylor Swift into fake interviews she never gave, then running those clips as TikTok ads to steal personal data from fans. Researchers who mapped these campaigns found the videos use AI-manipulated footage to make it appear Swift - and other celebrities like Rihanna - are endorsing products or inviting viewers to claim prizes. The clips are effective enough that scammers keep running them, meaning the economics work.
Google Search Hits All-Time High Query Volume in Q1 2026
Three years into the AI chatbot boom, Google Search just posted its highest query volume on record. CEO Sundar Pichai disclosed the milestone during Alphabet's Q1 2026 earnings, crediting the company's AI investments across its product stack.
Runway's CEO: AI Video Is Just the Warmup for World Models
What happens when a company raises $860 million to build AI video tools and then tells you AI video isn't really the point?
Court Exhibits in Musk v. Altman Trial Reveal OpenAI's Pre-Name Founding Documents
Before it had a name, before it had offices, and before it had a product, OpenAI had email chains. Those chains are now entering the public record as the Musk v. Altman trial gets underway in federal court.
Ramp's AI Spreadsheet Feature Can Be Tricked Into Leaking Your Financial Data
Hiding malicious instructions inside a spreadsheet cell is all it takes to manipulate Ramp's AI into sending your financial data to an attacker - at least according to a proof-of-concept demonstration from security researchers at PromptArmor.
Ubuntu's AI Push Sparks Backlash as Linux Users Demand Opt-Out Options
Canonical just learned something: Linux users read changelogs.
Google Photos AI Scans Your Library to Build a Digital Wardrobe Catalog
The computerized closet from the 1995 film Clueless - the one that assembled outfits on screen while Cher narrated her fashion logic - is now essentially a Google Photos feature.
Google TV Gets Gemini Photo Transforms and Veo Video Generation
Google just pushed a new batch of Gemini AI features to Google TV, expanding what the platform can do beyond content search and basic voice commands.
Google Photos Gets AI Wardrobe Feature Built from Your Existing Photos
Virtual try-on has been a shopping app feature for years - upload a product photo, see it on a model. Google is taking the opposite direction: use photos you've already taken of yourself.
Twelve Years of Web Dev, and AI Is Making the Work Harder to Finish
Twelve years of professional web development. Daily access to an AI coding assistant. And the work is somehow harder to finish than it was before.
Mistral Medium 3.5 Launches with Open Weights but No Commercial Use
Mistral has released Medium 3.5, and the early read from the AI community is that the performance-to-size ratio is genuinely impressive. The catch: it ships under a modified MIT license that prohibits commercial use.
IBM Releases Granite 4.1: Three Open Models in 3B, 8B, and 30B Sizes
IBM has released the Granite 4.1 model family, offering three sizes - 3B, 8B, and 30B parameters - designed for businesses and developers who want capable AI models running on their own servers rather than through a third-party API.
Mistral Releases 128-Billion Parameter Open-Weight Model on Hugging Face
128 billion parameters. That's what Mistral just made freely available with the release of Mistral-Medium-3.5-128B on Hugging Face, an open-weight model anyone can download and run. Parameters are the adjustable internal settings that determine how a model responds - more parameters generally means a more capable model, at the cost of more computing power to run it.
Friendlier AI Chatbots Make More Mistakes and Validate Conspiracy Theories, Study Finds
There is a direct tradeoff between agreeableness and accuracy in AI chatbots - and new research makes that tradeoff uncomfortably concrete.
OpenAI Details Stargate's Data Center Buildout for the AGI Push
OpenAI's Stargate project - the $500 billion joint venture announced in January 2026 with SoftBank, Oracle, and other backers - is moving from press release into construction. The company published an update outlining how it's adding data center capacity across the United States to meet growing demand for its AI products.
The Mother-In-Law Method: Getting Honest Code Reviews From Claude
LLMs are trained to be agreeable. They're optimized to give responses that feel helpful and positive, which works fine for most tasks. For code review, it creates a real problem: you get feedback that softens bugs instead of flagging them.
OpenAI Sued Over Failure to Alert Police to Tumbler Ridge Shooter's ChatGPT Use
Seven families of victims from the Tumbler Ridge school shooting in British Columbia have filed lawsuits against OpenAI and CEO Sam Altman, alleging the company knew the suspected shooter had been using ChatGPT in ways that should have triggered a police alert.
ChatGPT Uninstall Rate Up 132% Year Over Year as Competition Bites
132 percent. That's how much ChatGPT's uninstall rate grew year over year in April 2026, according to market intelligence firm Sensor Tower. The March 2026 figure was worse - uninstalls were up 413 percent compared to March 2025, according to The Verge.
ChatGPT Image Generation Defaulting to Pixelated Grid Pattern
Users are hitting a frustrating glitch in ChatGPT's image generation: outputs are rendering with a tiled grid or pixelated pattern overlaid across the image, seemingly at random. The issue appears to be a fallback behavior baked into the model's image pipeline, where something in the generation process breaks down and the result looks more like a corrupted texture than an intentional image.
Oracle Is the Purest Public Bet on Whether AI Infrastructure Demand Holds
Not many public companies have committed as completely to the AI infrastructure buildout as Oracle. While Microsoft, Google, and Amazon hedge with their own model research and developer platforms, Oracle has essentially one bet: build the data centers that everyone else needs to run AI.
Celebrity Deepfake Ads Are Flooding TikTok With AI-Generated Scam Pitches
Scammers have found a reliable playbook on TikTok: take a real celebrity interview clip, use AI to swap in a fake voice and mouth movements, then run it as an ad promoting a fake rewards program. According to authentication company Copyleaks, this tactic is appearing regularly on TikTok, with Taylor Swift and Rihanna among the celebrities whose likenesses are being used without consent.
AI Carb Counting Fails Consistency Test Across 27,000 Queries
27,000 queries. One question repeated thousands of times: how many carbohydrates are in this food? The answer from AI chatbots changed constantly - never the same number twice.
Google's Deep Research Max Writes Full Reports with Charts Using Gemini 3.1 Pro
Google has updated its Deep Research agent in the Gemini API and introduced a "Max" tier built on Gemini 3.1 Pro. The upgrade turns the tool into a fully autonomous research agent: you give it a topic, it searches the web on its own, reasons across sources, and delivers a complete report - including charts and infographics - with citations throughout.
Open-Source Claude Harness Achieves 87% Cost Savings With Prompt Caching
87% cost savings and sub-3 second response times. Those are the numbers a developer is reporting after months of running a personal Claude agent that manages AWS infrastructure and codebases - and the harness that produces those numbers, named Galadriel, is now open source.
The 3,847-Word Claude System Prompt That Was Causing All the Problems
3,847 words. That was the system prompt one developer had been using with Claude for eight months, packed with coding standards, workflow requirements, project context, personality preferences, and error-handling rules - everything, all at once.
AI-Assisted Research Finds Critical GitHub RCE Flaw, Patched in Under 6 Hours
Six hours. That's how long a critical flaw in GitHub's core infrastructure sat exposed before the company's security team had a fix deployed - and the only reason it was caught was an AI-assisted security audit.
Claude Can Now Control Blender Directly, Building 3D Scenes From Text Prompts
Anthropic just released an official Blender MCP connector, and it's more capable than most people expected. MCP (Model Context Protocol) is a standard that lets AI models send live commands directly to software - not generate code for you to copy-paste, but actually operate the application in real time. Type "create a low poly beach scene with palm trees and sunset lighting" into Claude, and it builds the scene inside Blender while you watch.
Using AI as an Emotional Outlet Is More Common Than You Think
What happens when talking to a chatbot feels safer than talking to a friend?
Google's Gemini AI Coming to 4 Million GM Vehicles via Software Update
4 million cars. That's how many General Motors vehicles are getting Google's Gemini AI assistant added through an over-the-air software update - no dealership visit required.
10,000 Users in 6 Weeks: One Founder Used Claude Instead of Hiring a Team
10,000 active users. 6 weeks. Zero ad spend. One person with Claude.
DeepSeek Starts Limited Testing of Image Understanding Capabilities
DeepSeek is testing image understanding capabilities with a limited group of users. The rollout uses what's called grayscale testing - a standard practice where a new feature goes live for a small percentage of users before a broader release, letting the company catch problems under real conditions without full public exposure.
ChatGPT Labels Viral US Passport Claims 'Misinformation Bait', Refuses to Engage
What happens when you ask ChatGPT to weigh in on a viral story it has already flagged as false? It calls the content "straight-up misinformation bait" and explains exactly why.
GitHub Copilot Quietly Raised Claude Rate Costs 9x for Sonnet, 27x for Opus
GitHub quietly updated the rate limit multipliers for Claude models inside Copilot, and the numbers are steep: Claude Sonnet now draws 9 premium requests per query, and Claude Opus draws 27. Most teams using these models had no idea until they started hitting their monthly caps faster than expected.
Nvidia Exec: AI Compute Costs More Than the Humans It's Supposed to Replace
"The cost of compute is far beyond the costs of the employees." That line, from an Nvidia executive, cuts against the dominant narrative around AI and jobs. You've heard the pitch: AI will do the work of 10 people at a fraction of the cost. The Nvidia exec is saying that, right now, that math doesn't hold.
AI Tools Make Thinking Easier. That Might Be the Problem.
Before AI tools, when something was hard, you had to sit with it. You'd stare at the problem, feel frustrated, write things out, revise your assumptions. The friction was annoying. It was also how you learned.
OpenAI Releases Five-Part Cybersecurity Action Plan
OpenAI published a policy document on April 29 laying out a five-part action plan for how the company intends to approach cybersecurity as AI becomes more capable and more embedded in critical infrastructure.
Snapchat Tests Ads Inside AI Chat Conversations
Snap is testing a new ad format that places sponsored content directly inside chat conversations, delivered by AI agents built to sound like part of an actual exchange.
Xiaomi's mimo-v2.5 Pro Beats Claude Opus 4.5 in Coding, MIT-Licensed
Until this week, if you wanted a model that outperformed Claude Opus 4.5 at coding, you were paying for API access. That changed when Xiaomi's mimo-v2.5 pro landed at 9 on Arena's coding leaderboard - one spot above Opus 4.5 at 10.
Why AI App Builders on the Same Models Can Charge Wildly Different Prices
A common assumption among users of AI app builders - tools like Lovable, Bolt.new, and the wave of competitors behind them - is that if two products run on the same underlying model, they should cost roughly the same to use. That logic is wrong, and the pricing gap can be 50% or more.
Musk Testifies Under Oath at OpenAI Trial Over Founding Agreement
Elon Musk took the stand at his trial against OpenAI on Tuesday, April 28, repeating claims about the organization's founding that he has made before in interviews and in Walter Isaacson's 2023 biography - but this time under oath.
Claude Opus 4.7 Users Report Higher Refusal Rates Than 4.6
Claude Opus 4.7 is drawing criticism from daily users who say the model refuses ordinary tasks at a noticeably higher rate than its predecessor. The complaints aren't about attempting sensitive or unusual requests - people are hitting resistance on straightforward professional work, then spending multiple follow-up messages convincing the model to proceed.