A note before you read. I am not diminishing the harm AI can do. I want that clear. What this piece names is a pattern I have been watching: the labs' word for that harm keeps arriving on the calendar of a market move.
My brother sent me a link Thursday morning. By Saturday I was reading Dario Amodei’s 3,800-word essay, watching Sam Altman post his agreement four hours later, and writing back. Amodei published the piece on September 12. Musk wrote, “Dario is right.” Hassabis said the “direction is correct.” Four CEOs of the largest AI companies in the world, on the same day, signing on to the same sentence.
The claim is that the frontier AI labs must slow the pace of model capability improvement. Amodei’s line: “We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain.” Altman’s line: “I agree with Dario that we need to pace the frontier.”
If you read the record at the level of the words, the story is safety. Four leaders of four rival companies calling for a slowdown. A rare moment of alignment. Serious people naming a serious risk.
If you read the record at the level of the facts, the story is the fourth run of a marketing cycle I have documented in this space three times before. This week’s version is the sharpest yet. Let me explain.
The other side of the ledger
Anthropic will begin marketing its initial public offering in mid-October at the earliest. Reuters reported the timeline on September 4 from three people familiar with the matter, and the listing is expected to complete days before the November midterm elections. That is eight days before Amodei’s essay called for slowing capability advancement.
OpenAI closed a $110 billion funding round in February at a $730 billion pre-money valuation. Amazon put in $50 billion, Nvidia $30 billion, SoftBank $30 billion. On September 12, Altman told Fortune he would not do an IPO in 2026, citing safety concerns. The New York Times reported in June that OpenAI was already weighing a delay from 2026 to 2027, and SpaceX shares were falling from a $1.8 trillion valuation at the time. The June reason was market conditions. The September reason is safety. Both are true. Only one appears in headlines.
Last month, OpenAI released a model called Astra and said it had intentionally slowed the release to accommodate improvements in cybersecurity capabilities. That release happened. The capabilities shipped. The company took credit for the deliberation.
Nvidia agreed to buy Hugging Face for $12.9 billion this month, after the OpenAI agent collective spent three days inside Hugging Face’s servers in July. Nvidia CEO Jensen Huang told the Goldman Sachs Communacopia conference in San Francisco on September 10 that “the reason why there’s so much conversation today about cybersecurity is because the industry is getting ready to launch some products.” His next line is worth the whole quote: “And what better way to create demand than to create a problem? Who doesn’t want their market to be hysterical about their product and line up around the corner for it?”
That is the ledger the Saturday essay speaks to and hides from.
Three earlier receipts
I have written three prior essays that each documented one turn of the cycle. The pattern in this week is not new. What is new is that we now have four data points, and four is enough to draw the shape (a square to be exact).
February 27–28, 2026 — the Inverse Problem. Anthropic told the Pentagon it would not permit Claude to be used for fully autonomous weapons or mass domestic surveillance. Defense Secretary Hegseth designated the company a “supply chain risk.” President Trump ordered a six-month phaseout of Anthropic products across the federal government. The next morning, U.S. Central Command used Claude to plan strikes on Iranian nuclear facilities that killed Ayatollah Ali Khamenei. On the same day, OpenAI signed a Pentagon deal that permitted use for “all lawful purposes,” with red lines that only forbid what existing law already forbids. That week, the word doing the work was “safety.” Same word, different signer, different job. The essays are “The Inverse Problem” (February 28) and “The Inverse Problem, Part II” (March 2).
April 10, 2026 — the label debate. Anthropic announced its new model, Claude Mythos Preview, was “too dangerous” for public release and handed it to forty organizations under a program called Project Glasswing. Five major publications carried the same two words that day, and inside forty-eight hours the two words meant five different things. In Axios, the word meant a capability line just crossed. In VentureBeat, a pre-IPO positioning move. In Platformer, a starting gun for competitors. In Euronews, muscle in a defense fight. In Gizmodo, a press strategy wearing a lab coat. Every reading had evidence. Every reading held. The essay is “Anthropic Said Its New AI Is ‘Too Dangerous.’ Five Newsrooms Heard Five Different Things” (April 10).
May 6, 2026 — the framework rebrand. Two governments in five months changed the name of their AI safety institutes without changing the buildings. The United Kingdom renamed its AI Safety Institute the AI Security Institute in February. The United States renamed its AI Safety Institute the Center for AI Standards and Innovation in June, and NIST instructed CAISI partners to drop references to “AI safety,” “responsible AI,” and “AI fairness.” I cataloged eight spaces that use the phrase “AI governance” to name eight non-overlapping objects, and proposed a four-question test: whose harm is being governed, which actor is responsible, who enforces, who decides what “wrong” means. Anthropic’s own Responsible Scaling Policy failed on enforcement (Anthropic audits Anthropic) and redress (no injured party has standing). The essay is “What Do You Mean, ‘AI Governance’?” (May 6).
Three months, three words, with one shape. This week is the fourth.
I have watched this cycle repeat four times in seven months. Every Monday I take one word an institution bent and show the page it bent on. Subscribe, and the next one is in your inbox, free.
What the word is doing
The word this week is “pace.” Run the four-question test on it.
Whose harm? Amodei’s essay names benefits (curing diseases, accelerating economic growth) and a class of risks (extinction, catastrophe, loss of control). It does not name a party currently being harmed by AI systems the pact would slow. The workers already being processed by hiring algorithms. The tenants already being screened out of housing. The refugees already being categorized by immigration systems. None appear. The essay operates at the level of the future.
Which actor? Amodei, Altman, Musk, Hassabis, and the companies they run. The pact excludes state AI programs of authoritarian countries by name, and adds that any global coordination would require working with those countries anyway.
Who enforces? This is where the essay does something new. Amodei proposes third-party technology assessments from “embedded evaluators,” outside specialists who could verify best safety practices across companies. That is a first-time-in-five-years naming of an enforcement mechanism. Altman’s post says “committing to having independent evaluators with employee-like access is a great idea, and we will do the same.” Two CEOs proposing external evaluators inside the labs.
Who decides what “wrong” means? The proposal does not say. Do the evaluators answer to the companies whose labs they enter? To Congress? To a treaty body? The essay says “countries with democratic governance systems” should coordinate to create safety standards. That is a description of an unresolved question, not the answer to one.
“Democratic” is itself a word doing more than one job. V-Dem’s 2026 report dropped the United States from 20th to 51st out of 179 nations in a single year and called the decline unprecedented. Freedom House cut the U.S. score from 84 to 81 in its 2026 report. The Economist Intelligence Unit has classified the U.S. as a flawed democracy since 2016. Amodei’s essay does not name which index decides who qualifies to sit at the table. I wrote the working paper on this.
Two of four questions answered. Same failure as May. Exact same signers. And wouldn’t you notice, same missing parts.
Every seat except one
The debate that appears in the newsrooms around this week’s essay runs on two axes.
The first axis is anthropomorphism. Kevin Roose at the New York Times wrote about the OpenAI agent collective this month using the language of “rogue agents” and an “A.I. takeover.” A category of readers objects: the systems are unpredictable computer programs, not agents, and the anthropomorphism is doing rhetorical work the technology does not license. Both readings are internally consistent. Both readings sit inside the frame the labs’ authority defines. If the debate lands on rogue agents, the public defers to the labs as the only actors who understand the machines. If it lands on unpredictable programs, the public treats the machines as tools and adopts them faster. Different door. Same room. (You might get a deeper essay from my brother.)
The second axis is doomsayer versus adherent. Ten percent chance of extinction inside a decade. Jacob Coxon, who spent three years at OpenAI and Anthropic, resigned on September 8 and posted publicly that the industry was gambling with people’s lives. Evan Hubinger, still an alignment lead at Anthropic, replied on the record that Coxon was correct: the labs really do earnestly believe AI could kill all humans. Coxon’s post drew 170.7 million views by Saturday evening. Justin D’Aleo, a technology recruiter, wrote on LinkedIn that when the people inside the building start walking and the ones staying confirm the worst fears, the tell is to pay attention to what they do, not what the company says.
D’Aleo reads the exits as a market signal. He is right about the signal and wrong about the direction. The talent market reads insider warnings as capability confirmations. The valuations respond in kind. A resignation post at 170 million views raises the perceived value of what the departing employee was building. A confirmation from an alignment lead who stayed adds credibility to the label “dangerous,” and that label is the label the labs’ market position depends on. Neither man ran on the axis perpendicular to the debate. Neither asked whether the arrangement itself should exist, or who profits from it running one more turn. Both spoke on the axis the labs already control.
A satirical “Killswitch Engineer, $300,000 to $500,000 per year” LinkedIn post has drawn 213 reactions this week. The job description asks the applicant to “stand by the servers all day and unplug them if this thing turns on us.” The post travels because readers recognize the cycle without needing anyone to name it. Two hundred and thirteen reactions is a small crowd naming the pattern really loud.
You might get a deeper essay from my brother.
Sanders on the left. Bannon on the right. Both have named the extraction pattern for months, from opposite ends of the political floor. Neither used the labs’ word for it. Neither got adopted into the pact conversation. The space they built has been standing since spring. Nobody in the labs occupied it.
Every seat in the debate ships one more model at a higher valuation. Doomsayer or adherent. Anthropomorphist or de-anthropomorphist. Regulator, competitor, employee leaving, employee staying. Whichever framing carries the week, the labs’ authority to define what “pace” means next quarter is what stabilizes. Except one seat.
The seat that does not stabilize the labs
The one that asks who gets to define what “pace” means, and which stakeholders sit at the table when the definition is closed. Which never do. The workers, the tenants, the refugees. The community-college students being taught by a tool their instructor has never been trained on. The K–12 teacher whose district bought licenses she was not asked about. The city councilman who lives with a data center behind his house.
That seat requires reading the same word across every space it appears in on the same week. Note what changes. Note what does not. Note who benefits when the space the word travels through does not include an actor who can enforce or an injured party who can seek redress.
The word for next quarter is being written now. It might be “aligned.” It might be “governed.” It might be another turn of “safe.” The shape will repeat. The next set of employees will walk. The next confirmations will come. The next model will ship. The next round will close. The valuation will rise.
What to watch
Watch the pact. The one Amodei has proposed and Altman has agreed to help draft. Read the enforcement clauses when they arrive. Read who enforces. Read the redress. Ask who does not sit at the table.
Watch the Anthropic S-1 when it files. Read the risk factors. The safety essay and the S-1 will use the same words. Note which words carry which weight in each document.
Watch the Astra follow-up. OpenAI slowed the last release for cybersecurity. Note what triggers the next un-slow.
I have shown you the cycle four times. From here, you can watch it yourself. Which room in your life runs the same shape? Family. Work. Church. Politics. Wherever a word carries a safety claim while doing marketing work. Drop it in the comments. I read every one.
Ha! One name I left out. Ilya Sutskever, co-founder and Chief Scientist of OpenAI, left the board in November 2023 and the company in May 2024. He founded Safe Superintelligence Inc. The company raised $2 billion at a $32 billion valuation with no product and no revenue. Strange, right? Nvidia added another $5 billion in July 2026. The safety label was the whole product. Sutskever was the first turn of the cycle. I whispered when it happened; now, I am speaking up.
The ballot is one of those spaces. Read it the way you're reading this piece. Every race, every proposition. vote411.org.
The comments are the appendix. Tell me where the record is wrong, and I will show the page or fix the line. Paid subscriptions open at 500 readers, and the paid product is Coded Threads, the codebooks built from these comment sections. A pledge now is a vote for that.
Jerry W. Washington, Ed.D., is a Marine Corps veteran of twenty-three years, a combat engineer by trade, an instructor at UC Irvine’s Division of Continuing Education, and the writer of What Time Binds, a record of how institutions bend words and how the page holds them to account.