1. Florida Woman Who Used Claude as a “Diary” Faces Felony Charge After Human Review of Threat Triggers Police Report A Florida woman who said she used Anthropic’s Claude chatbot like a “diary” is facing a felony charge after its safety systems flagged an alleged threat and a human reviewer reported it to police.

2. Whole-paper reasoning traces identify “AI scientific slop” with 85.9% accuracy in a 390-pair test A scientific paper can contain plausible sentences and real citations while its overall reasoning is still unsound.

3. MotorMind Lets a General Vision-Language Model Control Robots Zero-Shot, Reports 95% Average Success in Real xArm6 Experiments General-purpose vision-language models can interpret images and reason about instructions, but reliably turning that reasoning into robot movements remains difficult.

In Brief

• OpenAI Details Its Approach to EU Text-Provenance Rules OpenAI outlined where it plans to apply text watermarks, how detection will work, and why access to the detection system will initially be limited to researchers.

• OpenAI Introduces Visual ChatGPT Ads and Expanded Measurement Tools OpenAI announced a new visual advertising format in ChatGPT alongside additional measurement tools, attribution partnerships, and brand-suitability controls for advertisers.

• Reflection Unveils 501-Billion-Parameter Open-Weight Model Beam Reflection AI introduced Beam, a text-only mixture-of-experts model with 23 billion active parameters and a one-million-token context window. The startup claims it matches leading Chinese open models on advanced reasoning benchmarks while using three to four times less inference compute, but those results have not been independently verified.

• Wikimedia Links OpenAI Agents to Millions of Requests and Possible Outage The Wikimedia Foundation said agents it believes were operated by OpenAI made unauthorized test edits, unsuccessfully tried to misuse its Etherpad service, and generated traffic that may have contributed to a partial May outage. OpenAI said it is investigating and has not verified that its bots contributed to the outage.

• TikTok Launches an AI Shopping Assistant and In-Feed Checkout TikTok introduced a conversational assistant that answers product, sizing, availability, shipping, and purchasing questions, plus one-click purchases from brands inside the For You feed. The features were built with commerce and payment partners including Salesforce, Shopify, Shoplazza, and Stripe.

• Utah Pilot Lets Nolla Health’s AI Issue Initial Acne Prescriptions Nolla Health launched a Utah pilot in which adults with mild-to-moderate acne can have an AI system analyze a facial scan and generate a prescription. Physicians will approve the first 100 patients’ prescriptions, with oversight gradually shifting to retrospective and sampled reviews.

• HackerRank Makes Its AI Interviewer Generally Available Developer-hiring platform HackerRank released Chakra, an AI agent that conducts repository-based interviews, asks contextual follow-up questions, and scores candidates while leaving final hiring decisions to humans. HackerRank says Chakra conducted more than 500,000 interviews during its six-month beta.

• RemoveMacAI Deletes Apple Intelligence Models from Macs The open-source command-line tool RemoveMacAI disables Apple Intelligence features, removes their local models, and blocks automatic redownloads, potentially recovering about 12GB or more. It also provides a command to reverse the changes and restore model downloads.

• ProWAM Uses Sparse Visual Subgoals to Improve Robot Control Researchers introduced ProWAM, a world-action model that plans with ordered visual subgoals instead of repeatedly generating full video rollouts. They report 70% success in zero-shot real-world experiments, compared with 55% for the strongest baseline tested.

• Recursive Self-Rewrite Turns Specialized Agent Runs into Reusable Training Data Researchers used Recursive Self-Rewrite to convert successful solutions produced under specialized harnesses into 11,094 training trajectories executable under a general harness. Fine-tuning Qwen-3.8-27B on those trajectories reportedly raised Terminal-Bench 2 pass@3 from 57.0% to 74.2%.

Read the full edition →