Vatican calls for AI dignity & White House AI order stalls - AI News (May 26, 2026)

Let’s start with governance and ethics. The Vatican released “Magnifica Humanitas,” an encyclical from Pope Leo XIV focused on human dignity in the age of AI. It draws parallels to the industrial revolution and argues that AI isn’t neutral—because it inherits the values and incentives of the people and institutions behind it. The document flags risks like biased “objectivity,” simulated empathy that can mislead users, and high-stakes decisions—jobs, credit, services—being delegated to systems without compassion. It also calls out the material footprint of AI, and pushes for clearer accountability, stronger oversight, and more public control over how data is treated. Notably, the Pope personally presented it at the Vatican alongside Anthropic’s Christopher Olah, signaling the Church wants a seat at the table with the people building frontier systems.

On the U.S. policy front, reporting says President Trump was set to advance an executive order on AI risk, including a voluntary model-testing process for government evaluation. But an eleventh-hour push from adviser David Sacks helped derail it, arguing that “voluntary” steps can quickly become mandatory regulation—and that could slow U.S. progress amid competition with China. The immediate takeaway isn’t the fine print, it’s the power struggle: the administration appears split between a safety camp and a speed camp, and the speed camp just scored a win. For companies trying to plan around U.S. AI rules, this kind of last-minute reversal is its own signal: uncertainty remains the default.

Now to the economics of the model market, where the price war keeps getting louder. DeepSeek made permanent a seventy-five percent price cut on its V4 Pro model—after initially treating it like a short-term promotion. That matters because V4 Pro also supports very long context, the kind enterprises use for document review, codebase analysis, and big conversational histories where cost can explode. Cheaper long-context changes what’s feasible—and what procurement teams will pressure vendors to match. But it also comes with tradeoffs: some enterprises will weigh savings against reliability, compliance requirements, and geopolitical exposure when sending sensitive data to a Chinese provider. And hovering over it all is the unresolved dispute around alleged improper distillation of Claude outputs, which keeps training-data provenance in the spotlight.

Staying with the business of AI, Anthropic is reportedly heading into a breakout quarter: projected Q2 revenue of about 10.9 billion dollars, potentially its first profitable quarter, and growing demand tied to Claude’s coding products. If those numbers hold, they reinforce a broader shift: coding and developer workflows look like one of the first truly massive monetization channels for LLMs. At the same time, Anthropic is also rumored to be upgrading Claude’s “memory” in a more structured way—moving from a single summary into multiple topic-based “Memory Files,” more like a personal wiki the model can consult. That would make Claude feel more persistent without constantly bloating the context window. And on the security side, there are signs the high-capability Claude Mythos model may be inching toward wider availability, assuming safeguards mature. Put together, it’s a picture of a company trying to scale both revenue and responsibility—especially in areas like cybersecurity where the downside risk is high.

For developers, one of the most grounded pieces of advice today comes from software engineer Nolan Lawson: AI coding tools don’t have to be “slop cannons.” His argument is that if you deliberately slow down and use multiple models as reviewers, you can turn LLMs into an amplified code-review partner instead of a speed-only autocomplete. The twist is that the bottleneck becomes triage—because models are now pretty good at finding potential issues, but someone still has to validate and prioritize what matters. Lawson’s workflow ranks findings by severity, fixes only the critical ones first, and sometimes abandons a pull request entirely if the approach is wrong. The key point: AI can raise code quality and developer understanding, even if it doesn’t increase velocity—and it may even uncover old bugs that force extra testing and refactoring.

That idea—treating agent output as something you measure and manage—shows up in OpenAI’s latest developer cookbook too. They published a walkthrough on “macro evals,” a way to evaluate multi-agent systems by analyzing patterns across thousands of traced runs instead of judging single answers in isolation. The value here is practical: real agent failures often come from workflow problems like bad handoffs, missed signals, or inconsistent tool use, not a single obviously wrong response. By clustering traces and ranking recurring failure patterns by impact, teams can prioritize fixes that actually move reliability. If you’re building agentic workflows in production, this is the kind of methodology that helps you stop playing whack-a-mole with one-off bugs.

And if you’re building tools around agents, the Model Context Protocol—MCP—just dropped a release candidate for a major 2026 spec update, with the final planned for late July. The headline change is a stateless core over plain HTTP, removing the handshake and session assumptions that made some deployments awkward. In practice, that’s about easier scaling, better caching, and cleaner enterprise architecture—especially when you’re load-balancing or auditing requests. MCP also formalized an extensions framework and introduced clearer deprecation paths, which is a polite way of saying: implementers should expect breaking changes and should start validating now. The bigger significance is momentum—MCP is trying to become the common “plumbing” that keeps tools and models from turning into isolated islands.

On the product side, multimodal assistants are pushing further into everyday tasks. OpenAI’s ChatGPT account demonstrated a workflow where you snap a photo of a form, tell ChatGPT—by voice—what you want filled in, and get a completed version back. It’s not flashy research; it’s admin work automation. If it holds up in real use, it’s the kind of feature that makes AI feel less like a chatbot and more like an assistant that handles the annoying parts of work. And in open model news, ByteDance Research released Lance, a unified multimodal model that can both understand and generate or edit images and videos. The interesting part isn’t any single benchmark—it’s the broader direction: smaller, generalist multimodal models that aim to replace a pile of separate systems for captioning, editing, and generation. That could reduce complexity for teams trying to ship multimodal features without building a whole zoo of models.

Security teams also got something tangible: Perplexity open-sourced Bumblebee, a read-only tool to scan developer laptops for exposure during supply-chain incidents. Instead of relying only on repo scanners or build systems, it checks what’s actually present on endpoints—things like risky packages, extensions, and even AI agent configurations. That matters because recent supply-chain compromises often trigger during install steps, and your developers’ machines can be the missing piece in incident response. The project emphasizes not becoming an attack vector itself by avoiding package-manager execution and sticking to metadata reads. For orgs dealing with frequent vulnerability fire drills, this is a practical way to tighten the feedback loop.

Finally, a quick note on readiness—because “agentic AI” is moving faster than enterprise foundations. A new survey-based readiness index argues that while many companies are spending heavily on agentic AI, only a small slice has the data consistency and governance to run it safely and effectively at scale. The theme lines up with a broader push toward Open Data Infrastructure—store data once in open formats, keep governance centralized, and avoid tool lock-in so agents and analytics can access consistent definitions. Whether you like the framing or not, the message is hard to dodge: without interoperable, well-governed data, agents don’t become “autonomous,” they become unpredictable.