People have had it with the AI slop. I have been hearing the frustration for months. But the root cause is misunderstood. It’s not low effort; it’s low bandwidth.

In defense of slop

Alright, I’ll cop to it: I’m not really here to defend AI slop, per se. But I do want people to start to think differently about why this is happening, and encourage a more empathetic mindset.

We hear a lot from the people who bear the burden of receiving AI slop. These are the leaders within organizations. They have never been more inundated with information, yet most of it is less tailored to their understanding than before.

We hear a lot less from the senders.

All slop is not created equal

“If I had more time, I would have written a shorter letter”

Blaise Pascal

I have a strong suspicion that the receivers of AI slop can’t really evaluate the level of effort that went into its production. We all agree that the least sympathetic scenario is that the sender gave 2 sentence prompt to produce something that will take an hour to digest. But unless the people in your organization suck, the “slop” received is more likely an artifact of trying to quickly compress an incredible amount of context down into something transferable.

An anecdote

I hope you’ll excuse my rambling, but I found it easiest to get this off my chest by dictating it, rather than presenting it as writing:

I’ve been thinking about this a lot in the places where I feel like I get dinged for AI writing coming through in stuff that I send to people.

The problem is, if I’m doing something like super complicated with the help of AI—and that could be like the coding itself or it could be technical deep dive—there’s a reality that, with the pace that we’re expected to go, I’m just not going to understand everything down to like the root principles. I wish I could, but I’m constantly working on things that I’ve never seen before and I have no idea how deep the rabbit hole goes.

And so I kind of have to heuristically decide how deep I want to get with it. You know, like how much I want to truly, truly understand on a line by line level every byte that’s going through the system. And usually the answer is like I just can’t afford to get that deep.

So then when it comes time to communicate with people, I have the choice between:

• Do I put this into my own words with my own like approximate understanding of the situation?
• Or, do I let the AI explain it in its nuanced way, with all of the caveats and Claudish and other tics it has, knowing that it’s definitely going to be more precise than me?

I feel like it’s a catch-22. Sometimes, I try to get around this by putting something into my own words. Then I ask the AI to validate that what I’m saying sounds right. Then the AI will come back and say, “well, this is close, but it’s not quite right for these like seven load-bearing reasons”. And I can then choose how much I want to iterate on this editing, in hopes that it may eventually be both accurate and in my voice. But the more I listen to the AI, the more it starts to sound like a level of comprehensiveness that the AIs produce, which is annoying to people.

And so I think the core challenge really is that there are ways of abstracting information enough to get points across to people who don’t have time to do like the full deep dive. It’s just that this process is not simple and we can’t automate it with AI.

It takes a lot of effort to be able to like in order to truly do a good job of it. You have to start with an understanding that really gets down to the deepest level. Then you know what the right choices are of what details to abstract out. I just find that there’s never the time to do any of this and so I’m pretty much just fucked whatever my choice is because people are going to be mad it’s taking too long or they’re going to be mad that it’s not precise enough or they’re going to be mad that it it sounds overwrought. So it’s just a hard game to win.

I kind of suspect that this is what your report is kind of going through and i think what i why i kind of I think why I empathize with him is because I’m in his same situation.

And I just wish that other people had more empathy for why this is happening. Like it’s not just people throwing slop over the wall. I hear that all the time from other engineers, from my manager, from, you know, the discourse. And I think folks just are not fully appreciating the constraints that people are under who are trying to do their best to communicate all this stuff out.

The bandwidth is different now

For those of us now spending most of our days interfacing with chat bots, work has fundamentally changed. We now process immensely more information:

• Rounds and rounds of questions and followups with agents
• Parsing through large code diffs in pull requests
• Understanding and triaging review agent feedback
• and so on

This pushes most of us to the brink of our mental capacity to digest information. Periodically, we need to figure out how to transfer the most important context we have gained to other people.

But here’s the thing: all of those other people are also at the edge of their capacity, because they are doing the same thing. This is a hugely underrated challenge.

The end result: patience and empathy are lost, and anything that doesn’t hit the bullseye is “slop”.

So, what now?

I wish I had a solution for the bandwidth problem, but I don’t. It’s a really tricky problem. I do have some ideas.

On the sending side:

• Tell your agents to always label writing they generate as “agent-produced” so that people aren’t guessing.
• Communicate whether you are optimizing for speed of handoff, accuracy, and/or human perspective.

On the receiving side:

• Look for the positive intent and presume that what you’re seeing is actually the result of a lot of thought and work.
• Don’t expect one-shot perfection — ask questions for things that aren’t immediately clear.