I resigned from Anthropic today. I spent the final three years doing pretraining investigation at the two OpenAI and Anthropic. Neither business is acting responsibly. They are racing direct to self-improving superintelligence and gambling alongside our lives. More thoughts below.
— Jacob Coxon (@hilbertspaess) September 9, 2026
This is how Jacob Coxon stated goodbye to the AI industry, on X, final week.
And in a follow-up post:
“The group construction AI earnestly accept that it could slay us all by the end of the decade. This is not a promotion stunt.”
Jacob Coxon is not a star. He is not a CEO, not a celebrated scientist, and does not clasp a PhD (yes, in USA you can be an “AI researcher” without a PhD). He joined Anthropic at 27 years old, chiefly because of its safety reputation, and he likes to resolve problems and “make a earth a improved place”.
He is, in another words, an engineer. Like most of us.
But he did item most of us never do: he remaining two months before his equity vested, and gave it up to speak.
“I no longer have item to acquire by juicing up Anthropic’s valuation. … I remaining before any of my equity vested.”
Note that he does not call Anthropic “evil”. He equal told the Wall Street Journal that their safety activity is sincere. His allegation is colder than that: “if you’re under force to race, you have to cut corners”.
You could contend the reply was oversized. One resignation letter, additional than 115 myriad views, a rage on social networks between engineers, entrepreneurs, or merely humans, and lawmakers abruptly calling for regulation.
But that is exactly the point: whenever the departure of a sole engineer alarms the entire world, it tells you what the earth already suspects: whenever the group inside the machine say this one is dangerous, afterward the noise is additional a indication than a bad buzz.
The insiders confirmed it themselves. Evan Hubinger, an alignment squad guide at Anthropic, wrote:
“Jacob is accurate here—we really do earnestly accept AI could slay all humans! I personally think it is >10% inside the next decade. I accept Anthropic is trying its best, but we do not yet have a scheme to resolve alignment for superintelligence and are not plainly on track to.”
We have seen this movie before #
Five years ago, another employee looked at her own company’s inner investigation and faced the identical choice.
Facebook’s own slides, revealed by the Wall Street Journal’s “Facebook Files”:
“We create build depiction issues worse for one in three adolescent girls.”
and:
“Among teens who reported suicidal thoughts, 13 percent of British users and 6 percent of American users traced the desire to slay themselves to Instagram.”
Frances Haugen did not shrug. She gathered the documents, resigned, and testified before the US Senate.
Coxon in 2026, Haugen in 2021. Different companies but identical move: they refused, they left, and they warned the outside.
The form is the point. In the two cases the business knew.
And, in the two cases, person inner had to decide whether knowing was enough.
The disease #
It is tempting to decrease this to a few evil companies. It is additional uncomfortable, and additional accurate, to appearance at the civilization feeding them: ours.
Engineers today are asked to adore efficiency and to treat norm as a luxury. A large 2026 study of tech workers summarizes the circumstance in one respondent’s sentence:
“I can do more, faster, but not better.”
The identical study names the condition: smiling exhaustion.
People are shipping again, compensation has climbed, and there is no off-switch. More than fractional of tech workers (53%) would steer a newcomer distant from their own field, and they stay. Another report on application delivery established that 67% of developers now expend additional period debugging AI-generated code. We obtain the slop, we obtain the exhaustion, we obtain observing the norm of our own activity slip, all of that for the compensation.
This is not constricted to AI labs. It is the identical reflex that makes engineers rationalize surveillance products, addictive algorithms, defence contracts, and agency activity alongside a shrug: “if I refuse, person worse volition do it”.
I wrote before that comfort is the enemy of skill, but I was too gentle because, at the end, we can see that comfort is the enemy of responsibility.
Trading morale for prosperity is not a occupation decision but a transaction.
It is a transaction anywhere you market the part of you that was expected to care. Do that frequently enough, and it does not arrive back.
That is not burnout, but it is motionless a disease. Although, akin all disease, it spreads: archetypal to your team, afterward to your product, afterward to the group using it.
The accusation #
So, why are you motionless operating for them?
If the answer is the salary, say it louder, and do not power yourself to create excuses.
If the answer is “someone worse would substitute me”, say it too, and louder too.
If the answer is “I can alter things from the inside”, afterward display me what you changed. Haugen tried the inside. Coxon tried the inside. At the end, unfortunately, the inner did not move.
We live in a society, and every act we obtain carries a responsibility.
You can disregard it, collect your compensation, keep your caput down, and inform yourself you are fair an engineer construction features… But then, do not regard yourself a contributor to humankind. You are fair a individual thinking concerning itself, nothing more.
The group who refuse are not heroes but human, and good group refuse to do bad things.
