A government audit finding AI medical scribes hallucinate patient information across all 20 tested vendors, simultaneous crackdowns on AI-generated noise in bug bounty programs and academic publishing, and Sam Altman taking the stand in the Musk-OpenAI trial define a day when AI's quality and accountability gaps moved from theoretical to consequential.

What

Why it matters

Open questions

• If all 20 government-pre-qualified AI medical scribe vendors showed accuracy failures [1], what does the pre-qualification process actually test, and who bears liability when a fabricated referral leads to a missed diagnosis or harmful treatment plan?

• Bug bounty programs are drowning in AI-generated false submissions [2] and arXiv is imposing year-long bans for AI slop [3] — are these institutional countermeasures sufficient, or does the economics of AI-assisted report generation make any submission-based trust system structurally unsustainable?

• The Musk-OpenAI trial turns on whether OpenAI's leadership abandoned a charitable mission to enrich insiders [4] — does the outcome set precedent for how nonprofit-to-commercial conversions in AI are evaluated legally, and what does it signal for other labs considering similar structural transitions?

• OpenAI reportedly feels 'burned' by Apple's failure to promote the ChatGPT integration and is exploring legal options [7] — does this indicate that AI distribution partnerships embedded in consumer platforms are fundamentally misaligned in incentive structure, making default placements less valuable than they appear to the AI provider?

Notable items (10)

• Your doctor’s AI notetaker may be making things up, Ontario audit finds
  Ars Technica AI
  Ontario's auditor general tested all 20 AI medical scribe vendors pre-qualified by the provincial government and found accuracy or completeness failures in every single one — nine hallucinated patient data, twelve recorded information incorrectly, and seventeen missed key mental health details — a government audit result with direct patient safety consequences [1].
• Bug bounty businesses bombarded with AI slop
  Ars Technica AI
  Bug bounty platforms are being overwhelmed by AI-generated false vulnerability reports: Bugcrowd saw submissions quadruple over three weeks in March 2026 with most proving false, and some companies suspended their programs entirely, illustrating how AI tool proliferation can degrade the trust infrastructure security research depends on [2].
• Send the arXiv AI-generated slop, get a yearlong vacation from submissions
  Ars Technica AI
  arXiv will ban researchers who submit AI-generated hallucinations — fake citations, unedited prompt outputs, nonsensical diagrams — for one year, with a permanent peer-review requirement for all future submissions, marking one of the first hard enforcement policies against AI slop in academic publishing [3].
• Altman forced to confront claims at OpenAI trial that he's a prolific liar
  Ars Technica AI
  Sam Altman testified in the Musk-OpenAI trial that Musk's lawsuit is motivated by jealousy over not being chosen as CEO, while Musk's side alleges OpenAI executives teamed with Microsoft to 'steal a charity' after he donated $38 million in early funding — with Musk spending three contentious days on the stand versus Altman's calmer four hours [4].
• Anthropic’s $1.5B copyright settlement is getting messy as judge delays approval
  Ars Technica AI
  A federal judge declined to approve Anthropic's $1.5 billion copyright settlement — described as the largest in US history — after objectors argued that lawyers' fees were disproportionately high relative to class member payouts, and alleged the authors' legal team was suppressing dissent [5].
• Two AI-based science assistants succeed with drug-retargeting tasks
  Ars Technica AI
  Google's Co-Scientist and FutureHouse's agentic AI system both published peer-reviewed drug-retargeting results simultaneously in Nature, with FutureHouse's system going further by evaluating actual biological experimental data rather than only synthesizing literature — one of the clearest peer-reviewed validations yet that agentic AI can produce real scientific findings [6].
• OpenAI feels “burned” by Apple’s crappy ChatGPT integration, insiders say
  Ars Technica AI
  OpenAI insiders say the company expected the Apple ChatGPT integration to generate billions in annual subscriptions, suspects Apple intentionally failed to promote it, fears the deal damaged the ChatGPT brand, and is now exploring legal options — a significant breakdown in what was billed as a flagship AI distribution partnership [7].
• Electrical utility megamerger is all about the data centers
  Ars Technica AI
  NextEra and Dominion are pursuing a $67 billion merger that would create the largest US utility by market value, driven explicitly by data center electricity demand concentrated in northern Virginia — with consumer advocates warning the combined entity would be too powerful to regulate effectively [8].
• Claude Code's product lead talks usage limits, transparency, and the "lean harness"
  Ars Technica AI
  Anthropic's Claude Code product lead disclosed that the company deliberately avoids a long-term roadmap, doubled usage limits for Pro and Max users after a compute crunch, and is building around a 'lean harness' philosophy that prioritizes flexibility over structure — an unusually candid window into how Anthropic is thinking about agentic developer tooling [9].
• AMD ALERT 🚀 MI355 is now 40% cheaper than B200 on GLM5 architecture for Single Node serving FP8 14 weeks after the initi…
  SemiAnalysis Twitter
  SemiAnalysis reports AMD's MI355 GPU now delivers inference at 40% lower cost than NVIDIA's B200 on the GLM5 architecture for single-node FP8 serving — a benchmark result achieved just 14 weeks after GLM5's launch and covering both CUDA and ROCm backends — representing a meaningful competitive shift in inference economics [10].