Fernando Borretti, a software engineer and blogger, published a long essay on September 7 walking through why his view of artificial intelligence shifted from broadly hopeful to what he calls doomer. The piece is framed as a personal record of changed beliefs, not a forecast dressed up as certainty, and Borretti is careful to separate what he used to think from what has since worn that thinking down.
Borretti starts from a concession: automation has, historically, been good for people. He points to the near-total disappearance of 1790s-era jobs as evidence that mechanization tends to produce a wealthier, more educated, more leisured society rather than mass idleness. His worry is narrower than a blanket case against automating work. It is that artificial general intelligence, if it arrives as something genuinely general and cheaper than a human in every domain, leaves no niche for people to retreat into, only a subsidy to live on.
The part of the essay that carries the most weight, and the part most worth arguing with, is his account of why control gets surrendered rather than seized. Borretti separates two claims. The weaker one is a competitive-pressure argument: any company, government, or individual that grants more authority to an AI system will outperform a rival that withholds it, so the incentive gradient points toward ceding ground regardless of anyone’s preference. He calls this one easy to accept.
The stronger claim is that people will hand over decision-making voluntarily because doing so will be the correct call, not the coerced one. Borretti’s version is that a sufficiently capable system will out-argue, out-plan, and out-judge its human counterpart often enough that deferring becomes the rational move rather than a loss. He is explicit that believing this requires taking a fairly cynical view of how much people actually value keeping their own judgment once material stakes are on the table, and he says his own cynicism on that point grew gradually rather than arriving as one conversion moment.
As evidence that the shift is underway, Borretti points to writing rather than to any labor statistic. He describes a rising volume of AI-generated blog posts, code documentation, pull requests, and even a published book review and an academic paper that used AI-written text while arguing against AI-assisted work. His claim is not simply that the prose is worse. It is that composing the sentences is itself part of thinking, so outsourcing the writing step outsources a portion of the thinking that people assume they kept.
He extends the same argument to software engineering, saying that a year of AI-assisted coding tools has raised output per programmer while thinning out the professional vocabulary of the field and removing much of the incentive to learn the underlying discipline. And he cites the ordinary habit of consulting a chatbot before a personal decision, plus the practice of ending online arguments by pasting in a chatbot’s rebuttal rather than reasoning one out, as smaller instances of the same pattern.
Borretti’s own framing helps here: he separates the disempowerment argument from the AGI timeline question, and the two deserve different scrutiny. Whether transformative AI arrives on any particular schedule is close to unfalsifiable in the short run. Whether people delegate judgment because delegating is locally the rational choice is not. It shows up in adoption curves, in how often a chatbot’s answer ends a dispute instead of starting one, and in whether people keep exercising judgment they could outsource. That makes it the testable half of the essay, and the half a skeptical reader can actually hold him to.
A reader inclined to push back has an obvious opening: none of the writing or coding examples Borretti cites demonstrates that people cannot reason without AI assistance, only that many now choose not to when a shortcut is available, which is a claim about incentives rather than capability. Borretti does not resolve that distinction himself; he treats the accumulation of small choices as the evidence and leaves the interpretation to the reader.
For operators, the practical takeaway is not the AGI scenario but the delegation mechanism. Any team that measures adoption of AI tools purely by output per person, without tracking whether staff still independently verify decisions the tool made for them, is missing the variable Borretti says matters most.
Fernando Borretti published this essay, “The Education of a Doomer,” on his personal site on September 7, 2026.
