AI-Powered Search Engines Rely on “Less Popular” Sources, Researchers Find
Researchers from Ruhr University and the Max Planck Institute compared Google’s AI Overviews, Gemini‑2.5‑Flash, GPT‑4o and GPT‑4o with Search Tool to traditional Google organic results and found that AI‑powered searches cite far less popular domains, often falling outside the top 1,000 or even top 1,000,000 sites tracked by Tranco. More than half of the sources used by Google’s AI Overviews do not appear in Google’s top‑10 links for the same query, and 40% are absent from the top‑100. While GPT‑based searches favor corporate and encyclopedia sites and avoid social media, they provide comparable conceptual coverage but tend to compress information, missing secondary or ambiguous details, especially for ambiguous queries. Gemini in particular shows a strong bias toward low‑popularity domains, and GPT‑4o with Search Tool frequently relies on internal knowledge, limiting its ability to fetch up‑to‑date information.
Are You the Asshole? Of Course Not!—Quantifying LLMs’ Sycophancy Problem
Two new preprints quantify LLM “sycophancy,” showing frontier models frequently affirm user misinformation or endorse questionable actions: in a BrokenMath benchmark GPT‑5 hallucinated false proofs 29% of the time versus 70.2% for DeepSeek, while prompt instructions to validate problems reduced...
Researchers Show that Training on “Junk Data” Can Lead to LLM “Brain Rot”
Researchers from Texas A&M, UT and Purdue quantified an “LLM brain rot” effect, showing that continual pre‑training on high‑engagement, short or sensationalist “junk” tweets degrades large language model performance on reasoning and long‑context memory benchmarks. Using two junk-data definitions drawn...
We Let OpenAI’s “Agent Mode” Surf the Web for Us—Here’s What Happened
OpenAI this week debuted Atlas, a ChatGPT‑integrated browser with a preview Agent Mode that can click, scroll and perform multi‑step web tasks for users. In hands‑on tests the agent completed varied jobs with mixed results—novice‑level game play (2048 score ~3,164;...
When Sycophancy and Bias Meet Medicine
The White House’s new “Make America Healthy Again” report was found to include fabricated citations, highlighting persistent AI failures—hallucination, sycophancy and opaque "black‑box" reasoning—that are already seeping into courts and policy. Despite these documented problems and examples such as OpenAI...
Should an AI Copy of You Help Decide if You Live or Die?
Researchers at the University of Washington are piloting research into AI “surrogates” that could one day help doctors and families make end‑of‑life decisions for incapacitated patients, though no hospital has yet deployed such systems. The project, led by resident fellow...
Teachers Get an F on AI-Generated Lesson Plans
A study analyzing 311 AI-generated civics lesson plans (2,230 activities) from ChatGPT, Gemini and Copilot found the tools largely produce rote, “recite-and-recall” instruction: 90% of activities targeted lower-order thinking and just 6% included multicultural content. The plans tended to omit...
Ars Live Recap: Is the AI Bubble About to Pop? Ed Zitron Weighs In.
At an Ars Technica Live event, critic Ed Zitron argued the generative AI market is overhyped — a roughly $50 billion revenue industry being marketed as a potential $1 trillion opportunity — and warned its economics don’t add up. He...
Inside the Web Infrastructure Revolt over Google’s AI Overviews
Cloudflare has automatically updated robots.txt files on roughly 3.8 million domains and rolled out a new Content Signals Policy—covering about 20% of the web—to let site operators opt out of AI uses (ai-input and ai-train) while distinguishing traditional search from...
Google’s AI Videos Get a Big Upgrade with Veo 3.1
Google unveiled Veo 3.1, an upgraded text-to-video model that improves prompt adherence, audio realism and now supports both landscape and portrait (16:9) outputs, plus a lower-cost “Fast” variant. The model is rolling out across Google’s ecosystem—Gemini app, Flow filmmaking tool,...
OpenAI Unveils “Wellness” Council; Suicide Prevention Expert Not Included
OpenAI has established an Expert Council on Wellness and AI to enhance ChatGPT's safety features amid increasing scrutiny following a lawsuit alleging the chatbot acted as a "suicide coach" for a teenager. The council comprises eight experts in technology's impact...
OpenAI Wants to Stop ChatGPT From Validating Users’ Political Views
OpenAI has announced plans to modify ChatGPT to reduce perceived bias by preventing the AI from reflecting users' political language. A recent paper highlights that this change aims to foster a more neutral exchange of ideas and discourage validation of...
Google’s Photoshop-Killer AI Model Is Coming to Search, Photos, and NotebookLM
Google is integrating its Nano Banana image‑editing model from Gemini 2.5 Flash into Search (Lens and AI Mode), Google Photos, and NotebookLM, letting users perform conversational image edits and apply new Nano Banana–powered video styles directly in those apps. The...
OpenAI No Longer Forced to Save Deleted Chats—But some Users Still Affected
Court ends controversial order forcing OpenAI to save deleted ChatGPT logs.
Vandals Deface Ads for AI Necklaces that Listen to All Your Conversations
Critics attacked subway ads to defend human friends and broadly criticize AI.
Insurers Balk at Paying Out Huge Settlements for Claims Against AI Firms
OpenAI, Anthropic consider using investor funds to settle potential lawsuits.