Daily News · 6 min read

AI News: May 26, 2026

ai-policy ai-ethics future-of-work research math agents coding-agent content-partnership journalism
Listen

1. Pope Leo XIV’s first encyclical “Magnifica humanitas” puts the Catholic Church squarely against optimizing humans, AI-driven warfare, and concentrated AI power

Vatican. Pope Leo XIV released “Magnifica humanitas: On safeguarding the human person in the time of artificial intelligence” on May 25, a 42,300-word encyclical that extends the Church’s social teaching tradition from Rerum Novarum through Centesimus Annus into the AI era and is dated May 15 to mark the 135th anniversary of Pope Leo XIII’s 1891 labor encyclical. The text warns against treating human beings “as things to be perfected or optimized,” critiques AI’s role in warfare, and frames a small technological elite as the modern equivalent of the industrial concentration Rerum Novarum was written against, calling explicitly for stronger AI regulation. Anthropic co-founder Chris Olah was invited to speak at the Vatican presentation and used the platform to argue for outside oversight of frontier labs; the encyclical itself rebuts Olah’s claims about model introspection, stating that AI systems “merely imitate certain functions of human intelligence.” TechCrunch’s reading of the document is that the AI framing is mostly a vehicle for older arguments about concentrated power and democratic erosion rather than a treatise on the technology itself. Source

2. ClickUp cuts 22% of staff, builds 3,000 internal AI agents, and introduces $1M salary bands for employees who manage them

ClickUp. Productivity company ClickUp laid off 22% of its workforce and CEO Zeb Evans framed the cut as a restructuring around what he calls a “100x organization” model in which roughly 3,000 internal AI agents outnumber human staff three-to-one, with employees expected to direct agents and review their output rather than do the underlying work themselves. The company is pairing the layoffs with new compensation bands that top out at $1 million per year in cash, available to anyone who delivers “100x impact” by building or managing AI systems, an unusual structure for a SaaS firm last valued at $4 billion in 2021. Industry trackers count more than 100,000 tech layoffs across roughly 250 events in 2026 already, and a Gartner survey cited by TechCrunch found that 80% of companies adopting autonomous tech have cut jobs but that the reductions are not reliably translating into financial returns — making ClickUp’s bet a closely watched test case for the agent-replaces-headcount thesis. Source

3. Twenty-author survey on AI in academic research identifies five principles and finds 80% of fully autonomous research systems produce fabricated results

Alpha Signal. A 20-author survey covering the full academic research lifecycle published via Alpha Signal on May 25 lays out five principles every AI research stack now has to solve, anchored by the empirical finding that roughly 80% of fully autonomous research systems produce fabricated results. The authors argue the bottleneck has shifted from raw model capability to verification and governance, with artifact generation consistently outpacing the ability to prove outputs are correct, faithful, or meaningful. They recommend layered architectures that explicitly separate exploration, tool-based execution, and verification, and conclude that the most reliable deployment mode is human-governed collaboration in which AI reduces friction in retrieval, drafting, coding, visualization, and review while researchers retain accountability for judgment, experimental design, and interpretation. The paper positions orchestration, provenance, and feedback design as comparably important to model scale for any team building an AI research stack. Source

4. Google DeepMind’s AlphaProof Nexus autonomously cracks nine open Erdős problems and 44 OEIS conjectures at a few hundred dollars of inference per problem

Google DeepMind. Coverage of DeepMind’s AlphaProof Nexus surfaced widely on May 25, detailing a framework that pairs Gemini 3.1 Pro with the Lean formal proof assistant to autonomously solve nine of 353 open Erdős problems — two of which had stood unresolved for 56 years — and prove 44 of 492 open conjectures from the Online Encyclopedia of Integer Sequences. DeepMind reports the inference cost per solved problem at a few hundred dollars, an order of magnitude lower than most prior automated theorem-proving work, with the Lean compiler providing machine-checkable verification rather than relying on natural-language plausibility. The overall solution rate remains low at roughly 2.5%, so the system is best read as a research tool that helps mathematicians narrow the search space rather than a replacement for proof discovery. The Lean proofs and accompanying natural-language write-ups have been released on GitHub under the google-deepmind/alphaproof-nexus-results repository. Source

5. xAI opens its Grok Build coding agent to all SuperGrok and X Premium Plus subscribers with plan mode, parallel subagents, and worktree support

xAI. xAI moved its Grok Build coding agent out of the SuperGrok Heavy-only tier and into early beta for every SuperGrok and X Premium Plus subscriber, expanding access to the same CLI it shipped in restricted form on May 20 alongside the grok-build-0.1 model. The CLI installs with a single command and ships a plan mode that drafts a step-by-step plan a user can approve, comment on individual steps of, or rewrite before any code is touched, after which every change lands as a reviewable diff. For larger jobs Grok Build can delegate to specialized subagents that run in parallel and supports deep worktree integration so each subagent operates in its own working copy. The grok-build-0.1 model behind the CLI has a 256K-token context window with $1.00 per million input tokens and $2.00 per million output tokens, putting xAI directly against Claude Code, Codex, and Antigravity for the agentic coding workflow. Source

Also covered in today’s xAI roundup.

OpenAI. OpenAI announced a content partnership with Grupo Folha and Grupo UOL, the first commercial AI content agreement between Brazilian publishers and OpenAI, that pipes real-time reporting from Folha de S.Paulo and UOL into ChatGPT with attribution and links back to the publishers’ sites. The deal also resolves the lawsuit Folha filed against OpenAI in 2025 over unauthorized scraping of its articles, ending the litigation as part of the licensing framework. Beyond content access, the contract gives both publishers ChatGPT Enterprise seats, OpenAI API credits, and access to Codex for software development, with the joint goal of building reader-facing AI products and modernizing internal editorial workflows. Financial terms are confidential, and the announcement extends the publisher-deal pattern OpenAI has been running since 2023 with Financial Times, Axel Springer, Condé Nast, Le Monde, and Time. Source

Also covered in today’s OpenAI roundup.