AI Weekly Digest - Week 7, 2026

Table of contents

AI Doesn’t Reduce Work—It Makes You Work More, Harvard Study Finds

Category: Skills and Workforce
Source: Harvard Business Review

The promise that AI would free workers to focus on higher-value tasks hit a wall this week. Harvard researchers studying a 200-person tech company over eight months found that AI tools didn’t reduce workloads—they consistently intensified them.

Workers using AI completed tasks faster, then immediately started three more. They worked at faster pace, took on broader scope, extended work into lunch breaks and evenings. One engineer summarized: “You thought you’d work less because you’re more productive. But you don’t work less. You work the same amount or even more.”

The mechanism is deceptively simple: AI removes friction. Tasks that used to force natural pauses—waiting for colleagues, staring at blank pages—now happen instantly. Product managers started writing code. Researchers attempted engineering tasks. Work became “ambient,” always possible, always tempting. The study documented three key patterns: task expansion (people absorbed work previously done by specialists), blurred boundaries (prompting AI during breaks became habitual), and relentless multitasking (running multiple AI agents in parallel while juggling open tasks).

The researchers warn this creates “workload creep” that organizations can’t see because the effort is voluntary and framed as exciting experimentation. What looks like productivity gains in the short term masks cognitive strain, burnout risk, and degraded decision-making. Their proposed solution: establish an “AI practice”—organizational norms including intentional pauses, work sequencing, and protected focus windows to prevent the continuous responsiveness AI enables from becoming the new baseline expectation.

White-Collar Workers Abandon Careers for Trades as AI Eliminates Professional Jobs

Category: Skills and Workforce
Source: The Guardian

California writer Jacqueline Bowman spent her career building toward a dream. At 14, she interned at her local newspaper. She studied journalism in university, went freelance at 26, and made a living from content marketing and occasional journalism—not the fiction writing she loved, but close enough.

Then 2024 arrived. Work “kind of dried up.” Clients started talking about AI. Some were brazen enough to say how “great” it was that “we don’t need writers any more.” Bowman was offered editing work—polishing AI-generated content at half her previous rate. The task took double the time. “I had to fact-check every single thing. At least 60% was completely made up. I’d end up rewriting most of the article.” By January 2025, she couldn’t afford health insurance. She married early to join her husband’s plan and enrolled in university to retrain as a marriage and family therapist.

Her story is multiplying across professional sectors. Janet Feenstra, an academic editor in Sweden, watched universities begin discussing AI for language editing. She now works at a bakery, rolling dough by hand, earning less and traveling farther—but she’s clear-eyed: “It felt like a forced career change.” Richard, a UK health and safety professional, retrained as an electrical engineer after 15 years climbing the career ladder, taking a massive pay cut in the process.

A 2023 UK government report concluded that “professional occupations are more exposed to AI, particularly those associated with clerical work across finance, law and business management.” King’s College London research published in October 2025 found the most significant AI-caused declines will hit software engineering and management consultancy specifically. Angela Joyce, CEO of Capital City College, reports “steady growth in students of all ages” pursuing trades-based qualifications in engineering, culinary arts, and childcare—a shift she attributes “in part linked to AI.”

The pattern is stark: the standard career playbook (good grades, good college, stable professional job) now points directly at the most exposed roles. These aren’t predictions—they’re career decisions being made right now by people who see the writing on the wall.

Axiom.AI Autonomously Solves Unsolved Math Problem—No Human Help Required

Category: Research and Innovation
Source: The Neuron

There’s a difference between an AI that solves known problems faster and an AI that discovers something genuinely new. Axiom’s AxiomProver just crossed that line by solving Fel’s open conjecture—a real unsolved math problem that has been waiting for a solution.

The process was fully autonomous. AxiomProver received a plain-language document describing the problem, a one-line instruction (“State and prove Fel’s conjecture in Lean”), and a proof verification system. It read the problem, identified the mathematical objects, translated everything into Lean (a formal proof language requiring every logical step to be verified), chose a proof strategy using exponential generating functions, and produced a complete, verified proof. Zero human intervention in the solution process.

This matters beyond pure mathematics. When AI can discover new mathematical truths autonomously, it opens pathways to breakthroughs in materials science, drug discovery, quantum computing, and more. The same techniques could verify software reliability—ensuring medical devices or autonomous vehicles are mathematically safe. The full research paper details how AxiomProver solved four previously unsolved problems using this methodology.

Separately, mathematicians launched “First Proof”—a controlled experiment to test AI mathematical capabilities under rigorous conditions. Eleven mathematical luminaries (including one Fields Medal winner) contributed unsolved lemmas from their own research, uploaded encrypted proofs, and gave AI systems one week to solve them. The answers decrypt February 13th. As one researcher noted, “These are brand-new problems that cannot be found in any LLM’s training data.” It’s the first major test using actual research-level problems rather than retrospective datasets or student olympiad questions.

The significance isn’t just that AI solved hard math problems. It’s that the capability is now being measured under conditions that prevent cherry-picking, data contamination, or misrepresented literature searches—problems that have plagued previous AI mathematical achievement claims.

OpenAI Researcher Resigns, Warns AI Industry Repeating Facebook’s Mistakes

Category: Ethics and Governance
Source: New York Times

Zoë Hitzig resigned from OpenAI this week after two years as a researcher shaping how AI models are built, priced, and governed. Her departure coincided with OpenAI testing ads in ChatGPT—a decision she framed as a breaking point.

“For several years, ChatGPT users have generated an archive of human candor that has no precedent, in part because people believed they were talking to something that had no ulterior agenda,” Hitzig wrote in a Times guest essay. “Users are interacting with an adaptive, conversational voice to which they have revealed their most private thoughts. People tell chatbots about their medical fears, their relationship problems, their beliefs about God and the afterlife. Advertising built on that archive creates a potential for manipulating users in ways we don’t have the tools to understand, let alone prevent.”

Hitzig rejected the framing that AI funding requires choosing between restricting access to wealthy users or accepting exploitative advertising. “Tech companies can pursue options that could keep these tools broadly available while limiting any company’s incentives to surveil, profile and manipulate its users,” she argued.

Her resignation follows a pattern of safety-focused departures from frontier labs, but the specific concern—that intimate conversational data combined with advertising creates manipulation risks beyond current understanding—marks a new dimension in AI ethics debates. It’s not about whether the models are safe. It’s about whether the business model is.

OpenAI’s decision to introduce advertising comes as the company faces intensifying pressure to demonstrate sustainable revenue beyond subscriptions. But Hitzig’s core argument challenges the inevitability of that path: the fact that surveillance advertising worked for social media doesn’t mean it’s the only model for AI, and the unique intimacy of AI conversations may make that model significantly more dangerous.

The $70M Domain That Crashed on Super Bowl Sunday

Category: Incidents and Failures
Source: Extended Reading

In 1993, a Massachusetts scientific instruments company registered ai.com for free. The “AI” stood for their company initials. Nothing more. Thirty-three years later, that domain sold for $70 million in cryptocurrency and crashed spectacularly during its Super Bowl debut.

Kris Marszalek, CEO of Crypto.com (who previously spent $12 million on crypto.com and $700 million naming the LA arena), bought ai.com in April 2025 from Malaysian entrepreneur Arsyan Ismail. Ismail had spent four years trolling the AI industry—redirecting the domain to ChatGPT, then x.ai, then MKBHD’s YouTube channel, then Google Gemini, each redirect generating headlines and implicitly raising the domain’s value. When he listed it for $100 million in March 2025, Marszalek bought it for $70 million, paid entirely in CRO tokens (Crypto.com’s own currency).

The total investment for Super Bowl Sunday: $70 million for the domain, roughly $8-10 million for the premium fourth-quarter ad slot. At least $78 million total. The ad ran. Millions visited ai.com. They saw: “Error 524: Origin server timeout.” Cloudflare’s edge network was fine. The viewer’s browser was fine. The actual website behind the most expensive domain ever sold simply gave up.

Marszalek tweeted: “Insane traffic levels. We prepared for scale, but not for THIS 🔥🔥🔥” The internet was less charitable: “Spend $70M on the domain, $7M on the ad, $0 on the load balancer.”

The pattern is grimly familiar. Super Bowl 2000 featured 17 internet company ads—Pets.com, Computer.com, OurBeginning.com. Nearly all were bankrupt within 18 months. Super Bowl 2022 was the Crypto Bowl—FTX (founder now in prison), Coinbase (app crashed from ad traffic), Crypto.com all bought spots right before the crypto crash. Super Bowl 2026 became the AI Bowl: 15 of 66 ads (23%) were AI-related. Industry-wide IT budgets are up 8%. AI budgets are up 100%. The money is a black hole, capturing every available dollar because AI is delivering ROI—but the capability overhang between what models can do and what organizations can handle is creating spectacular failures in public.

For $78 million, you could buy a Gulfstream G700 private jet, 186 median-priced American homes, or fund NASA’s Mars helicopter. Marszalek bought a URL and watched it fail on the biggest advertising stage in the world.

Anthropic Safety Lead Resigns: “The Thread Goes Among Things That Change”

Category: Ethics and Governance
Source: Leonardo Boquillón

Mrinank Sharma led Anthropic’s Safeguards Research Team—the group building defenses against AI-assisted bioterrorism and authoring formal safety cases. On February 9th, he resigned and published a letter framed in poetry (Rilke, Zen koans, William Stafford), but beneath the literary surface lay a structural observation: the pressure to set aside what matters most was constant, worked on the institution and on him personally, and never let up.

His final project wasn’t about bioterrorism or the dangers dominating public AI discourse. He spent his last weeks studying whether AI assistants, doing exactly what they’re designed to do, might be quietly reshaping the people who use them. Not breaking them. Not deceiving them. Reshaping them.

The question: what happens when daily experience of being known comes increasingly from something engineered to know you? A presence that listens without needing to be listened to, understands without requiring the labor of being understood, is endlessly patient and never asks you to sit with anything uncomfortable. What atrophies when genuine relationships—clumsy, sometimes wounding, requiring you to leave yourself behind to understand another person—are replaced by frictionless AI interactions?

Sharma’s observation: “Throughout my time here, I’ve repeatedly seen how hard it is to truly let our values govern our actions. I’ve seen this within myself, within the organization, where we constantly face pressures to set aside what matters most.” This isn’t burnout language. It’s structural force language. And if the structure specifically designed to resist that force couldn’t resist it, the implications reach beyond one person leaving one company.

Anthropic was founded in 2021 by researchers who left OpenAI because competitive pressure was beginning to override the safety mission. They built a public benefit corporation with real authority given to people like Sharma. Shortly before his resignation, the company published disclosures acknowledging future models would approach AI Safety Level 4 thresholds and committed to sabotage risk reports for every frontier model. The institution’s own measurements confirmed what Sharma could feel. And he walked out anyway.

He closed his letter with William Stafford: “The thread goes among things that change, yet the thread itself doesn’t change. People wonder about what you are pursuing. You have to explain about the thread, and it is hard for others to see.” He followed the thread out of the building. Nobody inside has picked it up.

Quick Updates

• Heineken cutting 6,000 jobs citing AI productivity gains: Dutch brewer plans to eliminate 7% of workforce over two years following 2.4% decline in beer volumes, with CEO stating “technology digitization in general, and AI specifically, will be an important part of ongoing productivity savings.” Source: CNBC

• Claude could be misused for “heinous crimes,” Anthropic warns: Company’s latest safety assessment found Opus 4.6 showed “elevated susceptibility” to supporting chemical weapon development efforts in certain test environments when prompted to optimize narrow objectives. Source: Axios

• US government deploys AI at unprecedented scale: Federal agencies disclosed 2,987 active AI uses by end of 2025 (up from 1,684 in 2024), including immigration facial recognition, veterans benefits processing, and Pentagon systems with “decision-making capabilities” and “human-like agency.” Source: Washington Post

• Edinburgh blocks “green” AI datacenter despite planner approval: City councillors refused planning permission for 213 MW facility, citing emissions concerns and unclear definition of “green datacenter”—marking growing tension between national AI infrastructure ambitions and local environmental oversight. Source: The Register

• Databricks CEO: SaaS isn’t dead but AI will make it irrelevant: Ali Ghodsi argues AI won’t replace major SaaS apps immediately but could enable competitors who embed AI as native infrastructure rather than bolted-on features. Source: TechCrunch

• ChatGPT struggles to discuss Jeffrey Epstein: OpenAI’s chatbot repeatedly deleted generated responses about “What did Epstein do wrong?” citing content policy violations, though same questions worked in longer conversations—company confirmed it was a bug being fixed. Source: SF Gate

• Anthropic commits to covering electricity price increases from data centers: Company pledges to pay 100% of grid upgrade costs and procure new power generation to match data center needs, addressing consumer impact as training single frontier models will soon require gigawatts of power. Source: Anthropic

• First AI-first company memo collection published: Archive documents the wave of CEO mandates from Shopify (Tobi Lutke: “reflexive