The LLM Doom Loop

Your old procrastination habit, now wearing a respectable badge.

Here's the cycle, and you'll recognize it. You give the model a vague instruction. While it works, you tab away -- to the feed, the inbox, the group chat -- for a few minutes that become thirty. You tab back, notice it misunderstood you, write a slightly better prompt, and tab away again. Repeat until the afternoon is gone and the actual thing is no closer to done.

People have started calling this the LLM doom loop, and the name is good because it's honest about what it is. It isn't a new problem. It's the oldest one, with better cover.

The loop didn't change; the alibi did

Strip the tooling away and look at the shape. A trigger -- the hard, ambiguous part of the work comes into view. An automatic move -- you do something that feels adjacent to progress instead. A payoff -- the relief of not sitting in the difficulty. A cost -- the work, still undone, plus the hours. That's a loop, and it's the same one that used to run while you waited for a build to compile or a page to load. The waiting was always just permission to leave.

What the model added isn't a new behavior. It's a story. "I'm waiting for the AI" reads as productive in a way that "I'm staring out the window" never did. The machine is genuinely working, so your avoidance has, for the first time, a respectable badge to pin on it. You're not procrastinating; you're orchestrating. Except the brain you're scrolling away from the work with doesn't know the difference, and neither does the work.

This is worth being precise about, because procrastination was never really about laziness or time management. The account that holds up frames it as a way of managing a feeling: when a task carries some low-grade aversiveness -- it's ambiguous, it might expose you, it's just hard -- you reach for whatever repairs your mood right now and pay for it later. The model is the most frictionless mood-repair device ever placed next to your real work. It hands you a legitimate reason to look away every sixty seconds.

Why it's stickier than the old version

The old procrastination had honest edges. Staring out the window felt like staring out the window; the guilt arrived on time. The doom loop hides its edges. Because output is technically accumulating -- prompts sent, drafts generated, tabs busy -- the day reads as work, and the reckoning that used to keep you honest gets deferred. You can run this loop for weeks and call it a workflow.

The micro-structure makes it worse. The model produces a steady supply of thirty-to-sixty-second gaps -- long enough for the hand to find the phone, not long enough to do anything real. You're not taking breaks from focus; the focus never assembles in the first place. It's fragmentation sold as multitasking.

The move out

You don't fix this with a better prompt, and you definitely don't fix it with another tool. The loop is upstream of the tooling.

The interruption is small and unwelcome, which is how you know it's the right one. When you hand the model the vague instruction, notice that the vagueness is the tell: a precise instruction means you've already done the hard thinking; a vague one means you're outsourcing the part you're avoiding and calling the wait "productivity." Name that, once, in the moment -- I'm leaving because this part is uncomfortable, not because the machine needs me to -- and the badge falls off. What's left is the old loop, which you already know how to interrupt, now that it isn't pretending to be a workflow.

The machine will keep working while you decide. That was always the problem.

Sources

• On procrastination as short-term mood regulation -- looking away from an aversive task to repair the feeling now, and paying for it later: Sirois & Pychyl, Social and Personality Psychology Compass (2013).