Here’s what’s on our plate today:

• 🧪 The office with no veto: why AI's safety chiefs keep resigning.

• 📰 Altman and the decel debate, Fender's CEO on AI music, PR firm's fake AI publicists.

• 💬 Prompt of the Day: design an AI safety role that actually has teeth.

• 🗳️ Poll: why do AI safety chiefs keep resigning?

Let’s dive in. No floaties needed…

The Laboratory

TL;DR

• The broken pairing: Institutions split power so no part overwhelms the rest, but AI safety offices got the responsibility without the budget or standing to use it.

• Four exits, one shape: Leike, Sharma, Hitzig, and a federal director all left roles where someone else controlled shipping speed, compute, and mandate.

• Rewritten mid-tenure: CAISI's chief quit three months in, after a June order moved model testing into classified channels and halted public evaluations.

• Cheap to hollow out: Prevented harm never shows on a balance sheet, so stripping a safety office costs nothing today and bills someone else later.

• The stakes: As AI reaches hospitals, banks, grids, and militaries, thousands will inherit the same job design, and warnings will keep arriving in the form of resignation letters.

The office with no veto: why AI's safety chiefs keep resigning

Institutions that run the modern world tend to rest on the same underlying idea: no single part of a system should grow powerful enough to overwhelm the rest. Power is divided on purpose, whether between the branches of government or among the officers of a company, so that consequential decisions require the consent of more than one person. That arrangement only works when those entrusted with responsibility are also given the means to exercise it: a budget, a staff, and the authority to refuse something and be heard. Responsibility and authority are meant to arrive together, because neither can accomplish much on its own.

The principle sounds straightforward enough that it is easy to mistake it for a law of institutions rather than an aspiration. In practice, responsibility and authority are often separated, leaving people accountable for outcomes they cannot meaningfully shape. Few places illustrate that contradiction more clearly than the offices created to keep artificial intelligence safe. Between 2023 and 2025, nearly every organization building or overseeing advanced AI created an office whose purpose was, in effect, to put the brakes on.

OpenAI built a 'superalignment' team, a group working on how to control AI systems smarter than the people who make them, and set aside 20% of its computing power for the effort. Anthropic, founded by researchers who had left OpenAI over safety disagreements, built its own safeguards team. The U.S. government went furthest by creating an AI Safety Institute inside the Commerce Department to test the most powerful models before the public could use them. Each of these offices carried the responsibility required by the design. The question was whether they also carried the authority.

A resignation with no reason attached

The most recent of those people lasted three months. Chris Fall ran the Center for AI Standards and Innovation (CAISI), the federal office that tests major companies' unreleased AI models for security risks, and he resigned on July 20 without any public explanation from him or the Commerce Department. Arvind Raman, who directs the standards agency housing CAISI, has taken over the office on an interim basis, returning the body to a condition it knows well.

Its first director, Elizabeth Kelly, departed in February 2025 after the incoming administration revoked the executive order that had created her institute, and the chair remained vacant for over a year before Fall arrived. America's main AI oversight body has had a permanent leader for roughly half of its existence.

The same exit, four times over

Nothing on the record links Fall's departure to a safety dispute, and his motives remain unknown, so the resignation matters less for its cause than for the pattern it joins. Jan Leike, who co-led the OpenAI superalignment team, quit in May 2024 and wrote that "safety culture and processes have taken a backseat to shiny products", after his team had reportedly struggled to get the computing power it had been promised, and OpenAI dissolved the team days later.

Mrinank Sharma, who led Anthropic's safeguards research, resigned in February 2026, writing that he had "repeatedly seen how hard it is" to let the company's stated values govern its actual decisions. In the same week, OpenAI researcher Zoë Hitzig left over the company's move into advertising in ChatGPT, and at Elon Musk's xAI, at least 11 engineers and two co-founders walked away, while former employees said Musk was pushing to make the Grok chatbot "more unhinged". Four organizations produced four distinct immediate triggers, and beneath them lies a single repeated structure.

When only half the arrangement shows up

That structure is the broken pairing described at the start, in which an office receives everything it will be blamed for and has no means to act. A safety chief answers for the harm a model might do, while the decisions that determine that harm, covering which products ship, how quickly, with what guardrails, and on what budget, belong to executives or politicians whose incentives reward speed. Leike was accountable for controlling future superintelligent systems and could not reliably obtain computing time formally reserved for the job. Sharma built safeguards at a company that sells itself on safety, yet still saw its values erode in its operating choices. Kelly's entire institution rested on a single executive order, so one signature in January 2025 removed the ground beneath her.

Fall inherited a stranger version of the same problem because his mandate changed underneath him while he held it. A June executive order moved the testing of AI models' hacking abilities into a classified process run by national security agencies, and CAISI was ordered to stop publishing the results of its model evaluations, a change that unsettled even the administration's own Senate allies.

Someone hired to run a public science office found himself running an annex to a classified system, and within weeks the post was vacant again. Anyone sitting in that chair has three available moves: accept shrinking influence, fight and risk removal, or leave. The record since 2024 shows which one keeps getting chosen, because resignation is the only lever the job reliably retains.

Why the seat keeps getting built this way

Understanding why organizations keep building the role this way means asking what the appointment does for them. Naming a safety chief reassures regulators, recruits, and the public that somebody is watching, and that reassurance holds whether or not the watcher can block anything. The act of watching itself shows no visible return because harm prevented never appears on a balance sheet or in a headline. When money or politics gets tight, the safety office becomes the cheapest thing to hollow out, since stripping it of power costs nothing today and the bill, if it ever arrives, lands later and on somebody else.

Governments hollow out the same way, except they redefine an institution rather than close it. The U.S. oversight body has been renamed and rescoped twice in three years. It became a pro-innovation standards center in June 2025, when Commerce Secretary Howard Lutnick announced that "Innovators will no longer be limited by these standards", and parts of its work moved into classified channels this year. Companies feel the same instability from the other side, and OpenAI agreed in June to limit the release of its GPT-5.6 models to a set of trusted partners at the government's request, under a review process the industry has found hard to read. An office whose purpose is rewritten faster than anyone can learn to perform it is bound to become a revolving door. Eventually, it stops mattering who occupies the position because the job itself never stays the same long enough for anyone to succeed.

Prompt of the Day

Bite-Sized Brains

• Altman and the decel debate: Sam Altman is wading into the "decel" debate over slowing AI down, a sign the argument over pace has moved from the fringe to the industry's center.

• Fender's CEO on AI music: Fender CEO Bud Cole pushed back on AI-generated music, arguing that real instruments and human craft still anchor what people actually want to play.

• PR firm's fake AI publicists: A PR firm was caught staffing its roster with fake AI-generated publicists, the latest case of synthetic identities quietly slipping into real business.

Tuesday Poll

The Toolkit

• Synthesise: No-code platform for building and monetizing AI-powered digital products without writing a line of code.

• Runway: AI video and image generation tool used by filmmakers and marketers to produce cinematic content from prompts.

• Tabnine: AI coding assistant that runs privately on your stack, useful for teams that can't send code to public AI services.

Rate This Edition