Two claims sit uneasily side by side in the same book, and once you notice the seam you cannot unsee it.
In The AI Con, Emily Bender and Alex Hanna call large language models "inherently unreliable, being designed to make shit up." A Magic 8 Ball, Bender reportedly told an audience at Harvey Mudd in late 2024, is about as good.
Then they devote a whole chapter to Clarkesworld, the science-fiction magazine forced to suspend submissions because it was drowning in stories generated by ChatGPT.
Hold both thoughts. The tool is useless garbage. The tool is so useful that it broke a magazine's intake system. As Francesco D'Isa puts it, if the output were truly garbage, literally nobody would want it.
The tool is useless garbage. The tool is so useful that it broke a magazine's intake system. If the output were truly garbage, literally nobody would want it.
That contradiction is the small crack. What it reveals is structural, and it is the subject of this essay.
D'Isa opens with a telltale detail. 2026, he writes, is the year AI's fiercest critics began writing their critiques with the help of AI—"clumsily, barely polishing the output." He notices the seams the way you notice a bad wig.
The easy move here is the sneer. Gotcha, hypocrite. You denounce the tool in prose the tool clearly wrote.
But the sneer is the least interesting thing in the room. The interesting thing is that the critic still bothers to hide it—still runs the output through a light rinse of self-editing—when everyone already assumes he didn't.
That assumption is the story. And the numbers behind it are no longer marginal.
The PagerDuty Shadow AI Survey, conducted across four countries at large firms in mid-2026, reportedly found two-thirds of office professionals had used AI tools at work while believing it wasn't permitted. A clear majority—88 percent, per PagerDuty—had fed work information into public tools like ChatGPT, Claude, or Gemini.
ShadowLock's synthesis of Gartner, Microsoft, Cyberhaven, and IBM data is blunter: shadow AI is universal in 2026. Every organization that has measured has found it. Three to five unsanctioned tools per employee per month is, per ShadowLock, the typical pattern.
So when a critic hides his use, he is not hiding it from a naive audience. He is performing abstention for a room that has done the math and already discounted the performance.
Concealment, in other words, has stopped working as communication. It persists as ritual.
Here the survey data turns from interesting to damning, because the concealment is stratified—and it runs the wrong way.
You might expect the powerful to flaunt their tools and the rank-and-file to sneak around. The opposite holds. In the PagerDuty data, 81 percent of employees believe leadership operates under a different AI rulebook than everyone else.
And leadership does. Senior leaders (77 percent) are, per PagerDuty, more likely than mid-level managers (66 percent) to believe they know AI better than their own tech teams. Senior leaders are, per PagerDuty, more likely both to feel they know AI better and to perceive a double standard.
The double standard scales with size: at larger organizations by both revenue and headcount, 85 percent of employees, per PagerDuty, perceive the two-tier system.
This matters because purist refusal has always drawn its authority from a certain moral altitude—the person who abstains speaks down to the people who use. The stratification numbers invert that geometry.
The loudest position of principled distance may correlate with a higher rate of hidden dependence. When the executive who lectures about restraint is quietly relying on tools they publicly disavow, "I don't use it" stops functioning as a claim and becomes a tell.
The audience for purity has evaporated. Not because it was argued out of existence, but because everyone in the room now assumes the purist is lying—and is usually right.
Which returns us to the book, because The AI Con is not a weak argument. Its chapters on labor are well documented; its analysis of AI hype as ideological cover for gutting public services is, D'Isa grants, its strongest work.
The trouble is a load-bearing beam that the rest of the structure quietly leans on while the authors deny it exists.
To indict AI for stealing jobs, the book must show the tools working well enough that studios swap concept artists for Midjourney, newsrooms sack editors for synthetic copy, and Clarkesworld buckles. Every page of the labor chapter presupposes a technology good enough to be adopted at industrial scale.
To indict AI as a con, the book must show the tools not working—"designed to make shit up," useful to almost no one, especially in the Majority World.
These require incompatible premises. The book alternates between them according to rhetorical convenience.
There is an escape hatch, and the authors reach for it: the problem is us. We project meaning where there is none; we are fooled by fluent synthetic text.
But follow that. If AI genuinely does not work and the entire species is too credulous to notice, then the problem is the species, not the software—and it is rather pointless to write a book warning a species that cannot read the warning.
Concede instead that some users can tell useful output from useless output—that they select, discard, and correct—and the usefulness of the tool becomes an empirical fact, and "it doesn't work" collapses.
This is not a personal failing on Bender and Hanna's part. It is a rhetorical structure, and the structure cannot survive contact with the saturation numbers.
You cannot simultaneously tell a workforce where 88 percent have already pasted their emails into a chatbot that the chatbot is a Magic 8 Ball. They have used the Magic 8 Ball. They kept the answer.
The stochastic-parrot posture worked when the tools were rough and the audience inexperienced. That audience no longer exists. Denialism did not lose an argument; it lost its listeners.
So what survives after denialism dies? One position, and it is the one that gives up the secret exemption.
The writer of "Why I Avoid AI Tools" makes no claim that the tools don't work. He concedes the opposite, and openly. "I know that my approach makes me less efficient," he writes. "I could have prompt-engineered hundreds of books in the time it took me to craft this essay."
That single sentence does what the polemic cannot. It prices the refusal instead of smuggling it.
He is not saying the tool is useless; he is saying he would rather be slower and remain himself. He chooses inconvenience as a form of cognitive sovereignty, and he pays for it in the open, in books unwritten.
Notice what this costs him rhetorically: nothing. Because he claims no hidden purity, there is no tell to catch, no seam in the wig, no gap between the posture and the paste history.
The coherence comes precisely from the admission. He has removed the thing the survey data punishes—the secret exemption—and what remains is a position you can disagree with but cannot expose.
That is the shape of the future for anyone who wants to criticize this technology and be believed. The critique has to be conducted by admitted users, or by openly costed abstainers—people who say I use it and here is what worries me, or I refuse it and here is what that refusal costs me.
What no longer travels is the third posture, the one that denounces from a height while the output arrives clumsy and barely polished.
Not because hypocrisy is new. Because the cover of secrecy has collapsed, and everyone—including the abstainer—already knows the neighbor, the colleague, and the critic are quietly using the tools they denounce.
The Magic 8 Ball, it turns out, gave a usable answer. The problem was never the ball. It was pretending you never shook it.