Aaron Levie says he now receives board decks that were written entirely by AI, "and it kills me."
He said it while deliberately putting aside his day job. Levie runs Box, whose software stores and manages companies' files, and on Sequoia's Training Data podcast one of the hosts had noticed something odd about where AI is welcome. In coding, the host said, using AI is "not only accepted, it's embraced." In content — the kind of documents that sit inside a product like Box — there is "almost this allergic reaction to it," the reaction people have been labelling workslop. Would that still be a thing in a few years?
Levie took off "the Box corporate hat" and answered as a consumer of the stuff. He wished there were a better word for it. Then he tried to explain why the same technology is a gift in one file format and an insult in another.
Code is a utility. A deck is a claim about you
Set aside the top engineers, Levie said — the ones with "deep taste in the code," for whom the judgment involved is as much an art as a science. For most of the world, he argued, code is a utility. You are trying to automate something, or to put an interface in front of someone so they can press a button and move to the next step. The value was always in the thing being accomplished, not in the authorship of the lines.
So if you can tell an agent to generate your whole backend or frontend, "it's not only acceptable, it's preferable," because that work was the thing blocking you from moving forward. He does not think the word slop disappears from software: there is taste in front-end design and in systems, and nobody wants vulnerabilities in their code. But the direction, for him, is settled.
Presentations are different, and the difference is not aesthetic. When somebody hands you a presentation, Levie said, there is still an association attached to it: you are trying to decide whether you can trust that person to execute on the thing, deliver the result, understand the topic. When the deck is generated, that reading falls apart. "I'm losing my ability to sort of know for a fact that, like, how much of the thought process was them versus how much was the AI?"
And then the question he could not settle: "How much should I even care about that? Because I myself am doing the same thing."
The double standard he admits to
Levie described the situation as a collective problem rather than other people's failing. He uses AI for some of his own brainstorms and decisions. When the same kind of output arrives from someone else, his first instinct is: "Hmm, should I trust you?" He thinks society may simply have to grind through this for the next three to five years and come out the other side.
His analogy — which he introduced as one of those "totally busted" ones — is the financial model. Nobody is offended by a spreadsheet. You can see that a macro produced it; you know the person did not personally compute all of it; they are showing it to you and you are discussing it together. So why, he asked, can't the same eventually be true of a strategy deck?
The unresolved part is what the document was ever standing in for. Is content a proxy for how much the person knows? For what you think they can go and execute on? "I think we're just in this very messy period where we have to kind of figure that out."
The temptation to let a detector decide
The conversation then landed on a specific piece of writing: a Wall Street Journal item by the investor Stan Druckenmiller. Levie said he had read the discussion around it and the reaction to it, but had not read the piece itself. Was it very sloppy, he asked.
The host who raised it said no — he had loved it, and treated it as a counterexample, since he normally has an allergic reaction to anything clearly written by AI and did not have one here. Levie went looking for the tells anyway: how many "it's not X, it's Y" constructions were in there? None that the host could recall. But, the host added, it shows up as 100% AI in Pangram. And dashes? He thought there were. "Okay, okay, you can't do that."
Pangram is a classifier: a system trained to sort text into human-written and machine-written, rather than a checklist of suspicious punctuation. The company describes a training process that starts from roughly a million human and AI-generated documents, pairing each human example with a synthetic version matched in style, tone and subject so the model learns differences that come from generation rather than from topic. It then searches larger collections of human text for false positives, adds those hard cases back into training, and repeats. Text is converted into embeddings and passed through a neural network with a human-versus-AI head. What comes out is that system's estimate, not a record of who typed what.
Neither the em dashes nor the classifier's verdict changed the host's view of the piece, which he still rated highly.
Reading everything twice
The cost Levie describes is paid by the reader. He says he is now doing twice the work on posts: reading once for the substance, and once for "the calculation of, like, did the person write it or am I just literally reading, like, a Claude prompt?" Then comes a second judgment on top of that one — whether the answer should raise or lower his estimate of the person, or of the post. "We're in for some weird times because of this."
He extended it to the people with no way out of the assessment. "I would hate to be a college professor. I would just, I would totally quit," he said. "Because you're just like, I don't know anymore what you did."
One host offered the calculator as the closest historical parallel. Levie thought even that analogy was breaking down. The calculator was finite; you still had to piece together many more things yourself. What is arriving now is not one task lifted off you. "At some point, like, this thing is doing, like, at least ten tasks at once."