AI News: May 18, 2026

1. Musk loses $134B fraud suit against OpenAI after a two-hour jury verdict

Oakland courthouse. A California jury rejected Elon Musk’s claims against Sam Altman and OpenAI in roughly two hours of deliberation, throwing out Musk’s $134 billion damages request and finding unanimously that the underlying suits had been filed too late. Closing arguments had narrowed the case to whether Altman is trustworthy enough to be running OpenAI’s transition to a capped-profit, with Musk’s team reserving the right to appeal. The verdict removes the largest legal cloud hanging over OpenAI’s planned IPO path and the for-profit conversion. Source

2. Pope Leo XIV to publish the first papal encyclical on AI, with Anthropic’s Olah as a guest speaker

Vatican. Pope Leo XIV will present his inaugural encyclical on artificial intelligence on May 25, with Anthropic interpretability lead Christopher Olah scheduled as a guest speaker at the presentation. The encyclical is the first papal-level Catholic doctrinal document focused specifically on AI, and is expected to carry weight across global Catholic policy positioning and over the long arc, into ethics curricula at Catholic universities and hospitals. Olah’s invitation signals that the document is engaging with the interpretability and alignment research strand of the field rather than treating AI purely as a labor or media question. Source

3. MAGA-aligned groups press the White House for mandatory pre-release testing of frontier models

Washington. A coalition of conservative organizations led by Humans First submitted an open letter to President Trump asking for an executive order that would require safety testing of frontier AI models prior to release. The framing is notable because it pushes from the political right for hard pre-deployment regulation of US-built frontier systems, complicating the administration’s earlier moves to centralize AI policy at the federal level and roll back state-level mandates. The letter lands as the FSB and IMF have both begun warning that the most capable models can already function as offensive cyber tools at scale. Source

4. Top-line AI startup revenue clears $80B, with Anthropic and OpenAI capturing 89%

Market analysis. A new tally cited in The Decoder puts the aggregate revenue of the leading AI startups at roughly $80 billion, with Anthropic and OpenAI between them capturing about 89% of it. The concentration is a sharp reframe of the “vibrant AI startup ecosystem” narrative: outside the two model labs, the rest of the named startup cohort splits a single-digit-billion long tail. The piece is useful context for VC and enterprise buyers trying to assess durability of mid-tier model and tooling vendors as the model layer consolidates. Source

5. World Action Models let robots simulate consequences from raw video, no labels required

Robotics research. A new survey writeup walks through World Action Models, a research direction in which robots learn to imagine the consequences of an action from raw everyday video footage rather than from labeled demonstrations. The thread links the approach to recent generative video models repurposed as physics simulators, with the implication that robot policies can be pre-trained against learned forward dynamics before any real-world interaction, then fine-tuned on small amounts of supervised data. The framing is timely against this week’s Cosmos Predict 2.5 LoRA tutorial from NVIDIA and the broader push toward simulation-first robotics. Source

6. Amazon turns Alexa+ into an on-demand AI podcast generator

Amazon. Amazon launched Alexa Podcasts, an Alexa+ feature that lets users ask the assistant to produce a podcast episode on any topic. The system researches the subject using partnerships with the Associated Press, Reuters, The Washington Post, and over 200 US local newspapers, generates an episode preview, and lets users adjust length, tone, and focus before AI-generated host voices narrate the final cut, with episodes saved in the Alexa app’s Music and More sections. The feature is rolling out to US customers and is positioned as the start of Alexa+ becoming a personalized AI content creator rather than a single-turn voice assistant. Source

7. South Korea’s LetinAR positions thumbnail-sized lenses as the optical backbone for AI glasses

Hardware. LetinAR, a South Korean startup, is manufacturing thumbnail-sized waveguide lenses that the company is pitching as the optical backbone for the next wave of AI glasses. The piece frames the optics layer as the underappreciated bottleneck for the consumer AI glasses category that Meta, Apple, Google, and Snap are all chasing, where compute and battery are tractable but field of view, brightness, and form factor still come down to which optical stack a device maker can license. LetinAR’s wager is that consolidation in optics will mirror what happened with mobile camera modules. Source

8. Stanford student op-ed describes a campus ChatGPT culture of “just a little bit of fraud”

Education. Stanford student Theo Baker published an op-ed reflecting on how ChatGPT has reshaped his graduating class, arguing that the model deepened a pre-existing pattern of low-grade academic dishonesty at the university rather than creating it from nothing. The piece is one of the cleaner first-person accounts of the everyday classroom dynamic now driving institutional pushback against AI tutoring, and is worth reading alongside this week’s AI policy coverage because it is the texture that shapes parent-voter sentiment behind the regulatory letters. Source

9. Import AI 457 covers AI Stuxnet scenarios, a “cursed” Muon optimizer regression, and positive alignment

Newsletter. Jack Clark’s Import AI 457 dropped May 18 with three threads. The lead is an analysis of a sophisticated computer virus targeting precision calculation software, framed as an early example of the AI-augmented offensive cyber playbook several labs have warned about. The middle section dissects a regression mode in the Muon optimizer that causes specific neural network configurations to deteriorate under prolonged training, which is timely given Cursor’s separate Muon-related work on Composer 2.5 the same day. The third piece is a research note on “positive alignment,” shifting the framing from harm prevention toward AI systems explicitly designed to support human flourishing. Source