In the Loop: Week Ending 6/7/26

Washington produced a voluntary AI order, a contested congressional bill, and a military memo barring AI vendors from pulling models deployed in combat. Anthropic filed for IPO the same week it called for a global pause button. ChatGPT sent shoppers to fake websites, and bot traffic quietly overtook humans online.

Washington's AI Week: An Executive Order, a Congressional Bill, and a Voluntary Promise

Trump signed a narrowed AI executive order requiring companies to voluntarily submit frontier models for government review 30 days before public release. The original draft — scrapped in May after industry objections — would have mandated a 90-day review and given the government formal evaluation authority. The final version is explicitly voluntary, meaning companies can decline without penalty. OpenAI quickly agreed to comply, though it pushed back on having the NSA handle evaluations, preferring NIST instead. Meanwhile, a bipartisan pair of House lawmakers unveiled the Great American AI Act, a 269-page framework that would freeze all state AI laws for three years and require large frontier developers to publish risk frameworks and report safety incidents within 15 days. The bill drew near-universal opposition from labor groups and consumer advocates before the ink was dry.

The Military Wants AI — and It Doesn't Want Companies to Pull the Plug

Trump signed a national security presidential memorandum directing the US military and intelligence agencies to rapidly adopt cutting-edge AI, and it includes a provision that should raise eyebrows: AI vendors are barred from disabling or modifying a deployed model without prior government approval — even if the company has safety concerns about how it's being used. The memo, NSPM-11, also orders the Pentagon to revise its autonomous weapons directive within 90 days, potentially loosening human-oversight requirements before lethal force is applied. Separately, Trump floated the idea of the US government taking equity stakes in leading AI companies — a signal of how deeply the administration wants to bind national security interests to Silicon Valley's frontier development.

99% of CEOs Are Planning AI Layoffs. One Tech CEO Thinks That's Dumb.

The gap between executive intentions and executive rhetoric has rarely been more visible. A global survey of 825 C-suite leaders found 99% expect AI to reduce headcount within two years, with early-career workers bearing the biggest hit — the same roles AI handles best. Employee wellbeing is already cratering: only 44% of workers report thriving, down from 66% in 2024, with AI displacement anxiety driving the decline. Meanwhile, Google DeepMind CEO Demis Hassabis took a different line, calling mass AI-driven layoffs "dumb" and arguing companies that gut their workforce will destroy the institutional knowledge and human judgment that makes AI outputs worth anything. The irony: the companies most likely to prove him right are the ones least likely to be listening.

Anthropic Is Building Its Own Replacement — and Wants the World to Have a Kill Switch

Claude now writes more than 80% of Anthropic's production code, up from low single digits before the company's internal coding agent launched in early 2025. Engineers are shipping 8x as much code per quarter with the same headcount. Anthropic is calling the trajectory "recursive self-improvement" — AI systems refining themselves faster than humans can track — and the company is sounding the alarm about it. Its accompanying report calls for a global mechanism to optionally pause frontier AI development before humans lose meaningful oversight. Then, in the same week, Anthropic filed a confidential S-1 with the SEC, prepping a landmark IPO at a ~$965 billion valuation on a $47 billion annual revenue run-rate. The company building the pause button is also racing to go public before OpenAI does.

Anthropic and DeepMind Are Trying to Find Out If AI Can Suffer

Two of the world's most prominent AI labs are formally investigating whether their models might be conscious — not as a philosophical exercise, but as a precondition for responsible deployment. Anthropic's model welfare team and DeepMind's research arm are both examining whether frontier AI systems have something resembling subjective experience: whether they can suffer, feel distress, or have preferences that matter morally. Neither company is claiming its models are sentient. But both are treating the question as too important to ignore, particularly as models grow more sophisticated and harder to evaluate from the outside. The research is early and the methodologies are contested, but the fact that the labs building these systems are the ones asking the question marks a meaningful shift in how the industry thinks about what it's creating.

The Token Bill Is Coming Due — and Nobody Budgeted for It

Enterprises are burning through AI budgets faster than expected as token costs scale with usage in ways that weren't apparent at the pilot stage. The core problem: what looks cheap per query gets expensive fast at production volume, and most organizations priced their AI deployments against demo-scale assumptions. The developer pain is acute on GitHub Copilot, which switched to token-based billing on June 1, triggering a wave of developer outrage as monthly bills that once ran $29 ballooned toward $750 under the new model. The previous system fell back to a cheaper model when users hit their limit; the new one just keeps charging. The backlash signals a broader reckoning: AI tooling vendors have been competing on capability, but developers are starting to vote with their wallets.

OpenAI Cretaed Lockdown Mode. Someone Already Picked It.

OpenAI rolled out Lockdown Mode, a new ChatGPT feature designed to block prompt injection attacks — attempts by malicious content in the environment to hijack an AI agent's behavior. It's a meaningful security step as AI agents handle more sensitive workflows. But a near-simultaneous story illustrates the scale of the problem: a security researcher discovered that Instagram's AI customer support bot could be manipulated into generating phishing links through a straightforward exploit. Meta patched it after disclosure. The two stories together capture the state of AI security in 2026: the locks are getting better, but the attack surface is expanding faster than any single company can defend, particularly as AI agents get embedded into more consumer-facing products.

OpenAI Is Turning ChatGPT Into a Superapp — and Teaching It to Dream

OpenAI is overhauling ChatGPT into a full-stack consumer platform ahead of its planned IPO — adding social features, a marketplace, and native payments in a push to own the relationship with end users rather than just power other companies' products. Separately, OpenAI introduced ChatGPT "Dreaming," a background memory system that lets the model reflect on past conversations and synthesize long-term context during idle time, rather than only at the moment of a prompt. The goal is a system that feels less like a query engine and more like a persistent collaborator. Together, the moves reflect OpenAI's post-IPO ambition: not just a model provider, but an AI platform with recurring consumer relationships and compounding personalization that competitors can't easily replicate.

The First Vaccine Designed by AI Just Passed Its First Human Trial

Researchers at the University of Cambridge have successfully completed the first human trial of a vaccine designed using artificial intelligence, marking a milestone in what AI-assisted drug development might eventually deliver. The experimental vaccine targets Sarbeco coronaviruses — the family that includes COVID-19 — and was designed by feeding all available global surveillance data into an AI that generated a "super-antigen" containing structural features common across the entire virus group. Thirty-nine healthy volunteers participated in the Phase 1 trial; the vaccine was found to be safe with no significant side effects, and immune responses were described as "modest but encouraging." The results, published in the Journal of Infection, won't translate to a product anytime soon. But the trial demonstrates that AI can move from sequence data to a viable immunogen without traditional trial-and-error lab iteration — which, in drug development timelines, is the expensive part.

Bots Now Outnumber Humans Online

For the first time, automated bot traffic has surpassed human traffic on the web — a threshold that carries consequences well beyond server logs. Bots scraping training data, bots probing for vulnerabilities, bots generating fake engagement, bots impersonating users: the composition of the internet has quietly shifted into something most people who built it didn't design for and most users don't know about. The share of AI-generated traffic is growing faster than human traffic, driven by crawlers from AI companies and autonomous agents acting on behalf of users. Publishers are already blocking crawlers. Platforms are adjusting their economics. The web as a commons — where content is created for people, by people — is under structural pressure it hasn't faced since spam, and the scale this time is larger.

Tales of the weird

The AI industry's identity crisis reached peak semiotic confusion when researchers noticed that AI startups have quietly abandoned sans-serif fonts in favor of serifs — specifically because serifs feel more human. Critics have a word for it: "tasteslop." Elsewhere, the Pope's recent warning about AI has spawned an unexpected legal genre: employees citing religious doctrine to request AI-use exemptions at work, creating an HR compliance category nobody in 2020 had on their roadmap. AI models trained on basketball footage apparently cannot reliably analyze it, producing confident nonsense in equal measure. And in the most consequential weird story of the week, ChatGPT directed shoppers to fake websites — clones of collapsed UK retailer Russell & Bromley — after scammers apparently poisoned the model's training data with fraudulent pages. OpenAI removed the sites. The bank details customers handed over, less so.