In his recent conversation with Shane Parrish, Tobi Lütke was asked what AI had made worse at Shopify. He said the failure case of lazy work is “not lack of output. It’s actually over output.” Internally they call the result slop grenades: messages their authors haven’t properly read and inflated emails the recipient has to feed back into an LLM to compress. His solution sits on the sender’s side: “If you’re already using LLM, just use it to synthesize your point simply.”
By the way, this is the same conversation in which he enthusiastically describes Shopify’s use of AI agents, and I don’t see a contradiction here. You can think a tool is enormously useful and still expect people to take responsibility for what they produce with it. A position that, judging by the reactions, is apparently unthinkable on the internet.
A few months ago, I wrote about AI slop eating organizations from the inside. My argument was that we should value raw clarity over manufactured professionalism, and stop sending polished pages when three bullet points would do. I still believe that. But I feel somewhat defeated, and less convinced it’s a workable defense.
Length is only part of it. The messages aren’t garbage, either. Often there’s a real point in there, or genuinely useful information. It’s just buried. The sender skipped the work of deciding what deserves to be there and left everyone on the receiving end to do it instead.
The introduction nobody needed. The background we already share. The historical explanation that is, admittedly, peripherally interesting, but has no bearing on the decision we need to make. Somewhere in the middle sits the actual request or the point. Even useful information has to pull its weight. If it makes the main point harder to find, its inclusion has a cost. “But it’s relevant” isn’t enough when everything else competing for the reader’s attention is relevant too.
It’s a false optimization. You saved yourself the effort of cutting the message down, then distributed that effort across everyone you sent it to. Nice productivity gain. For you. Josh Bernoff called the rule this breaks the Iron Imperative, years before anyone could generate a page in thirty seconds: “Treat the reader’s time as more valuable than your own.”
The obvious answer is to enforce better standards. Leaders could send things back, ask for the actual point, and make it clear that generating a document doesn’t mean the work is finished.
Alas, that only works when it’s held firmly from the top. Somebody has to actually do it, then do it again next week, and the week after that. Most leaders don’t have the stomach for it, or they simply don’t care enough. They have more important things to do, and most don’t see how people write to each other as a critical part of the organizational culture. And the example flows from the top as much as the rule does: if your CEO doesn’t care, why should you? Some organizations will do this well, and Lütke may well run one of them. That kind of leader is a rarity, though, especially when producing something that looks finished is so much easier than deciding what to delete.
Amazon’s six-page memo, read in silence at the start of the meeting, has lasted as long as it has because the writer pays: there’s no padding six pages that the people making the decision will read line by line. People who worked there have told me a six-pager could get polished for days, sometimes weeks, before it went anywhere near the room. The constraint exists to force that effort. They nailed it long before the LLM era, but Amazon is the famous exception rather than the rule.
And even a company with an excellent writing culture still receives messages from clients, vendors, and everyone else outside its control.
So I’m no longer willing to make other people’s restraint my main defense. I’d rather prepare for the slop to keep coming.
For routine work communication, that means giving up on reading everything in its original form. Email, Slack, meeting transcripts, and call notes would feed an assistant that knows what I’m working on and what I’m responsible for. It could be personal or sit on top of shared organizational context, within the appropriate access boundaries. AI Chief of Staff.
I don’t particularly want a shorter version of every email. Twenty summaries are still twenty things to read. I want to know what changed, what needs my decision, which commitment is slipping, and whether today’s confident update contradicts what we agreed last week. The useful output might combine five messages from three channels. Or it might tell me that none of them changes anything I need to act on. The selection follows my priorities and existing context, rather than the amount of prose someone decided to send.
There are limits, of course. A summary can’t recover thinking that never happened, and a neat digest can hide disagreement as well as a bloated email can. I’d want uncertainty kept, conflicting accounts flagged, and the originals one click away.
It also costs me what I defended in March. Some messages carry their meaning through effort: the review a manager actually sat down to write, the thank-you note. Through the same layer, the handwritten one and the generated one arrive looking identical.
And it makes things slightly worse for everyone else. While I read raw, the sender still got something back: a curt reply, a “what do you actually want”, or silence. That social filter was the last one left, and a digest removes it. The sender never learns the grenade landed, so they send more. Here comes the avalanche. My bet: within a few years, LLM-generated messages will outnumber personally written communication by orders of magnitude.
Here’s a crazy thought, though. What if that’s the direction of travel? What if in a few years no one writes or reads messages by hand? In that scenario, some of this additional context might actually be useful. Take a typical generated pull request description:
what the change is
the actual change, buried in the middle of a paragraph
a description of every file touched
the history of the module
the tickets that affected it before
None of it is invented. Most of it just doesn’t matter to me right now, and all of it costs attention to get through. But for an agent reading on my behalf, the arithmetic flips. Low-importance context is cheap for a machine to hold, and the module’s history or the old ticket might indeed be what lets it tell me that this change quietly undoes something we decided last spring. If nobody reads the original anymore, the background stops being addressed to me and becomes context for whatever reads on my behalf, provided it’s good enough to use it, which it is getting closer to.
The padding that stays worthless is the invented kind: a page a model expanded from three bullet points still contains three bullet points. And yes, the arrangement is ridiculous: three bullet points expanded into a page, sent to a machine that extracts three bullet points. I used to call that loop a reason to stop. Now it may just be how we cope. I don’t have to approve of the waste to stop volunteering my attention for it.
I’d still rather you sent the three bullet points. I’m just no longer building my working day around the hope that you will.