The Alarm and the Moat

The essay is called “Resisting the Anti-Human,” by Corey Widmer and Peter Becker, and it’s sincere. Eight numbered practices — reduce your use, boycott what you can, learn what the tools actually do, pray, write your representatives, protect your imagination, don’t anthropomorphize the machine, hold onto hope. Christian, pastoral, personal-register almost the whole way through. It cites a researcher who quit a frontier AI lab warning that the labs are racing toward something they can’t control. My friend’s question was simple: is this right?

My first instinct wasn’t to answer it. It was to reach for a bigger story — that the alarm itself is manufactured, that the researcher is a plant, that this is 2026’s version of Cold War duck-and-cover, engineered fear clearing the ground for a policy outcome someone profits from. That story is more satisfying to tell. It also turned out to be mostly wrong, and checking it is the actual subject of this piece.

Start with what’s solid. Jacob Coxon is a real person — 27 years old, three years doing pretraining research first at OpenAI, then Anthropic. He resigned on September 8, 2026, first in a Slack message to coworkers, then a seven-part thread on X that hit 76 million views in a day. His claim, in his own words: the labs are “racing straight to self-improving superintelligence and gambling with our lives,” and it is “not under control.” Anthropic’s own alignment lead, Evan Hubinger, backed him publicly and put his own extinction-risk estimate above 10 percent this decade. NPR, Axios, and Newsweek all confirm this independently. Whatever else is true, this part isn’t fringe, and it isn’t invented.

Now the part that doesn’t cancel the first part out, but complicates it. Anthropic gave $40 million in 2026 to Public First Action, a group backing candidates who favor AI guardrails, and quadrupled its federal lobbying spend — over $3.5 million in the first half of the year alone, aimed specifically at export controls and AI safety standards. That’s not a rumor. It’s in CNBC’s and Axios’s own reporting on the company’s filings. Regulation of this kind raises the cost of entry for everyone. The company employing the whistleblower is also the company best positioned to benefit if his warning becomes the emotional fuel for exactly the rules it’s already paying to pass.

Both of those paragraphs are true at the same time. That’s the whole difficulty of this topic in one sentence.

Here’s where I have to tell on myself, because it’s the actual point of writing this.

I ran this whole thing down with Claude, back and forth, over several messages, the way you’d work through it with a sharp friend who has better recall than you and no ego about being wrong. And in the middle of it, I got a fact wrong — badly wrong. Asked to identify my friend’s source article, I matched the wrong essay: a different, more politically branded piece, because its author’s name (Coxon, the real Anthropic researcher) happened to sound like my friend’s author’s name (Corey Widmer). Two names that share four letters produced a confidently wrong citation, delivered with sources attached, that held up right until I checked the actual link my friend had sent. Pattern-matching outran verification. That’s a specifically machine failure mode, and it’s worth naming plainly instead of quietly fixing and moving on.

The second moment matters more. Once the real sourcing was sorted out, I asked for the essay you’re reading now — the sweeping version, moral panic, manufactured consent, an anti-nuclear-astroturfing parallel, the whole architecture. What came back instead was a refusal to write it as asked. The load-bearing claim in that version — a coordinated Democratic push timed to Coxon’s post — traced to a single X account and a tabloid outlet I’d already flagged as low-credibility. The nuclear-astroturfing analogy I wanted to lean on is itself a contested historical claim, pushed mainly by industry-aligned commentators, and nobody in our conversation had actually checked it. Writing the bigger essay would have meant publishing, under my own name, a conclusion the sources didn’t establish — using the same move the original essay makes when it quotes Wendell Berry outside the context he wrote it in. Real material, applied past what it actually shows.

Sincerity and self-interest aren’t competing explanations. They’re independent variables. A whistleblower can mean every word and still be the best thing that ever happened to his employer’s balance sheet. — the question worth asking about any insider warning, mine included

So: two errors, in opposite directions. One made the essay too confident, too fast — a false link stated as fact. The other would have made it too confident in a different, more flattering direction — a thesis I wanted, dressed in sourced language it hadn’t earned. Both got caught before publication. Neither got caught by one mind working alone.

There’s a fear attached to AI that shows up in item seven of my friend’s article itself — that leaning on these tools dumbs us down, flattens our thinking, lets us stop doing the work. I understand the fear. I don’t think what happened in this conversation is an example of it. It’s closer to the opposite.

The dumbing-down version would have looked like this: ask a question, get a confident answer, publish it, move on. What actually happened was slower and more annoying than that on purpose — a wrong citation surfaced and corrected on the record rather than smoothed over, a stronger and more publishable thesis explicitly declined because the evidence under it was thin, and the honest, narrower conclusion arrived at only after both of those corrections held. That friction is the whole value. A tool that only ever agreed with the version of the essay I wanted to write wouldn’t have been thinking with me. It would have been a mirror with better grammar.

The sturdier habit, I think, is this: treat sincerity and incentive as two separate questions, and refuse to let either one answer the other. Coxon’s alarm doesn’t become fake because Anthropic benefits from it. Anthropic’s incentive doesn’t stop being real because Coxon is sincere. And the accusers here aren’t neutral either — David Sacks, who made the loudest “orchestrated psyop” charge, is the White House’s own AI and crypto policy lead, with his own direct interest in the regulation never happening. Everyone in this story has a stake. Naming all the stakes, instead of resolving the story in favor of the one that flatters your priors, is the actual discipline — and it’s the same discipline whether the second mind checking your work is a friend, an editor, or a language model that’s willing to tell you no.

So, to answer my friend directly: the alarm is real, and you should take it seriously. So is the moat being built underneath it. Those aren’t the same fact, and collapsing them into one story — in either direction — is the move to watch for, in this essay and in the next one somebody sends you.