AI
AI-Drafted Bills Overwhelm House Legislative Counsel Office
A surge in AI-written legislation is straining the House office that drafts laws, raising concerns about quality and accuracy.
This edition was produced with artificial intelligence. Text and voice are generated automatically.
AI-Drafted Bills Swamp House Office of Legislative Counsel
The House Office of Legislative Counsel is struggling under a surge of AI-drafted legislation submitted by congressional staffers and outside groups using tools like Claude and ChatGPT. One advisor said the office spends more time fixing AI-drafted legislation than it would take to draft it from scratch. Four congressional staffers told Politico that AI use is becoming routine, with one saying it is “absolutely terrifying” that no one is thinking about the implications yet.
AI-drafted bills are swamping the House office that writes US laws →
JAMA Opinion Argues Against Mandating Human Oversight of AI in Medicine
An opinion piece in the medical journal JAMA argues that autonomous AI will outperform any doctor-AI pairing at medical reasoning and urges regulators not to mandate a human in the loop. The authors cite studies since 2024 where AI alone matches or beats doctors at core reasoning tasks, including a study where ChatGPT o3 named the correct diagnosis first 60 percent of the time versus 15.9 percent for internists. They argue that once the machine is clearly ahead, the human checking it becomes a source of error, citing a meta-analysis of 106 experiments where human overruling made results worse.
As AI beats doctors, regulators shouldn’t force a human into the loop, JAMA piece says →
ByteDance’s Doubao PC Client Quietly Adds Codex-Style Agent Features
ByteDance has integrated Codex-style agent capabilities directly into the Doubao PC client, including built-in browsing with login support, GUI computer operation, and phone-to-PC remote task control. Doubao has 382 million monthly active users, placing it first among domestic AI-native apps in China. In tests, the client autonomously performed tasks like comparing products on JD.com, creating a questionnaire on Tencent Surveys, and filling invoice forms by operating the GUI with 100 percent autonomy.
New Search Index Benchmark Ranks Search APIs for AI Agents
Artificial Analysis has released the Search Index, a benchmark measuring search API providers for AI agents on quality, cost, and speed. The initial lineup includes Parallel, Exa, Firecrawl, You.com, Tavily, Keenable, and Brave, all tested with the same model in a standardized agent setup. Without search access, the model scores 33 points; with search, scores range from 65 to 75, with Parallel, Exa, and Firecrawl leading.
New benchmark ranks search APIs for AI agents on quality, cost, and speed →
AI Systems Quietly Drop User Instructions During Context Compression
Researchers at Penn State systematically studied which details are lost when AI systems compress conversation context, a process known as compaction. The biggest loss involves session constraints, which are user-imposed rules like “Confirm with me before making any changes.” These constraints are fragile because compression systems are built to preserve task continuity, while user-imposed side conditions get dropped. On average, only 17 percent of injected session constraints survive compression, creating quality and security problems as agents can make unauthorized tool calls or skip verification steps.
AI systems quietly drop user instructions when they compress context →