The Facts Behind the September 2026 AI Panic

Susan Frew • September 17, 2026

Share this article

Real quotes, real sources, zero hype. A five-minute read.

Quick take

  • A researcher named Jacob Coxon resigned from Anthropic on X the weekend of September 7, 2026, warning that neither Anthropic nor OpenAI is "acting responsibly."
  • Anthropic's own Alignment Science lead, Evan Hubinger, publicly agreed, putting his own estimate at more than 10 percent risk within the next decade, from a future system, not today's models.
  • A Google DeepMind researcher resigned days later with a similar warning.
  • Anthropic CEO Dario Amodei published a new essay, "We Must Pace the Frontier," comparing his proposal to Cold War arms treaties and calling for U.S. coordination with China on AI's pace.
  • Elon Musk called the original warning "a setup," and Mark Zuckerberg said market competition already handles AI safety, no coordinated slowdown needed.

No video with this one. Getting the sourcing right took the week instead. Every claim below links straight to where it came from.

I got more calls that week about AI than I had in the two years I'd been doing this work. Friends, family, people at events, all asking some version of the same question: are we all about to get killed by AI. So instead of the usual tactic, let's just walk through what actually happened, with sources, so you can decide for yourself instead of reacting to a screenshot someone sent you.

The Timeline of the September 2026 Resignations

The weekend of September 7, a researcher named Jacob Coxon posted his resignation from Anthropic on X. He'd spent about three years total doing pretraining research at OpenAI and then Anthropic. His words: "Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives." He said the people building this technology "earnestly believe that it could kill us all by the end of the decade." By the following Tuesday the post had over 100 million views and was national news.

Here's the part that made it harder to wave off: Evan Hubinger, who actually leads Alignment Science at Anthropic, a company that would very much prefer this story go away, backed him up publicly. "Jacob is correct here, we really do earnestly believe AI could kill all humans," he said, putting his own number at more than 10 percent within the next decade. But he was careful to add that the risk from today's models is low. What worries him is a future, more capable system that can improve itself.

Five days later, on September 15, a researcher at Google DeepMind resigned with a similar warning. So this stopped being an Anthropic story and became a three-lab story in about a week.

In the middle of it, on September 12, Anthropic CEO Dario Amodei published a new essay called "We Must Pace the Frontier." He compares his proposal to old Cold War arms treaties: "capping the number of missiles limited the potential for destruction while preserving each country's deterrent." His ask isn't a unilateral U.S. slowdown, it's the U.S. keeping its lead in AI while trying to coordinate, even informally, with China on the pace of development. China's state media called it a "Cold War script." You can read the whole thing yourself, it's short: darioamodei.com/post/we-must-pace-the-frontier.

And here's the part I think is actually the most useful thing to know: the loudest pushback didn't come from some rival camp of AI optimists nobody's heard of. It came from two of the biggest names in tech, and neither of them agreed with the alarm. Elon Musk called Coxon's post "a setup" and a "psy op," publicly doubting it was even genuine, not agreeing with it. Mark Zuckerberg said trust and alignment are becoming competitive advantages on their own, so market pressure already handles this, no coordinated slowdown required.

So no, this isn't "everyone in AI secretly agrees we're doomed." It's a real, specific, named disagreement, and now you know exactly who's on which side of it and why.

What I'd Actually Do With This News

  1. Separate two different questions. "Could AI eventually be dangerous" and "should I stop using ChatGPT this afternoon" are not the same question, and this news is about the first one.
  2. Read Amodei's actual essay, not just my summary of it or someone's screenshot of a headline. Two minutes, linked above.
  3. Notice this is a specific, named disagreement, not an industry consensus. Coxon, Hubinger, and a DeepMind researcher on one side, Musk and Zuckerberg publicly pushing back on the other.
  4. If having real guardrails matters to you, that's a conversation for your reps and your industry associations, not just your group chat.
  5. Keep working. Keep using the tools. Keep asking questions like this one. That's the whole idea behind Recoded, staying clear-eyed instead of frozen.

Try It Yourself

Skip the hot takes for two minutes and read the first few paragraphs of Dario Amodei's actual essay, linked above. Primary sources beat screenshots every time.

This is exactly what Recoded was built for. Not panic, not denial, staying informed enough to make your own call instead of outsourcing your opinion to whoever posted first.

Where I land, for what it's worth: I don't think heavy-handed government control of AI is the answer. I do think real guardrails matter. A global pause sounds nice until you remember it only works if every single country signs on, including the ones we have no reason to trust to keep their word. Amodei's Cold War framing actually makes more sense to me than either extreme: sit down with the countries that matter and negotiate real limits, the way we eventually did with nuclear weapons, instead of a pledge that only binds the people already trying to do this responsibly.

If sorting out how much of this is signal versus noise, for your business and your own peace of mind, is something you'd rather talk through with other people instead of doom-scrolling alone, that's a lot of what we do in my group coaching program. No pressure, no countdown clock, just an open door.

Grab a seat in the next cohort

Stay informed, not afraid.

— Susan


Recent Posts

Susan Frew video on five tech outages in eight days
By Susan Frew • November 19, 2026
Microsoft 365, the FAA, Chase and Apple all went down in eight days, and none were hacks. Here's the small business lesson and a free AI prompt for your Plan B.
By Susan Frew • November 12, 2026
Your next customer is probably already in your files. Export your list, let AI find who went quiet, and send a personal note in 15 minutes. Free prompts inside.
Title card: AI Is a Great Stylist. It's a Lousy Friend.
By Susan Frew • October 29, 2026
My AI gave me a 10 out of 10 on an orange jacket. Here's why that worries me, what the JED Foundation research shows, and a free 20-minute AI stylist handout.
Title card: Four AI Assistants Are Already Living Inside Your Browser
By Susan Frew • October 22, 2026
ChatGPT, Claude, Gemini, and Copilot already live in your browser. Here is how to use them on your open tabs, plus official links and a free cheat sheet.
Title card: This Week's AI Drops, Sorted by What You Pay For
By Susan Frew • October 8, 2026
OpenAI, Google, and Anthropic all shipped in eight days. Here is who gets Dots, Gemini 4 Argon, and Claude Sonnet 5.5 on each plan, plus a free cheat sheet.
Title card: I Hired 11 Employees. None of Them Eat Lunch.
By Susan Frew • September 30, 2026
I built an AI org chart with 11 named agents. Two are already live, and a human still checks everything. Here's the six-line job description and a free worksheet.
Title card: Set It Once. Your AI Remembers Everywhere.
By Susan Frew • September 11, 2026
Google expanded Gemini custom instructions across Workspace. Tell it who you are once, and it uses your voice in Gmail, Docs, Drive, Chat, Sheets, and Slides.
Title card: I Let Claude Click Around Without Me
By Susan Frew • September 3, 2026
Claude in Chrome can now act on its own, with a safety check on every click. Here's how to try it carefully, flip the toggle, and keep control of the keys.
Title card: One Click, One Video, Zero Extra Software
By Susan Frew • August 27, 2026
Google built a Record button into Slides through Google Vids. Talk through a deck, fix flubs in the transcript, and send a shareable video with no extra software.
Title card: Desktop? App? Browser? So Confusing. Are You Using the Wrong Version?
By Susan Frew • August 13, 2026
Claude and ChatGPT each show up in at least four places: web, desktop, mobile, and your browser. Here's how to pick the right one, plus a free one-page map.
Show More