LLM Burnout: A Practical Reset (When You’re Sick of AI Slop)

Here’s the endgame: you keep using LLMs, but you stop feeling nauseous every time you open a chat window. You cut the parts of your workflow that turn you into a rubber-stamp reviewer. You keep the parts that make you faster. No cold turkey. No dopamine detox. Just a workflow reset – about 20 minutes to set up.

If you clicked this because Alec Scollon’s post “I Think I Have LLM Burnout” just hit the Hacker News front page and something in it made your eye twitch – same. Let’s fix it.

The key takeaway

LLM burnout isn’t addiction and it isn’t the model’s fault. It’s what happens when your job silently mutates from doing the work to reviewing AI-generated versions of the work. The cure isn’t quitting – it’s rebalancing what percentage of your day is spent reading LLM output versus producing your own.

If you skip the rest of this article, do this one thing: pick the single task you use AI for most, and ban AI from it for one week. That’s it. The rest is optimization.

What actually triggered this trend

The trending post is by a developer describing a very specific feeling. His job changed from designing and writing code to designing code, describing the design to an LLM, reviewing code the LLM produces, and then finally writing code – a second job that appeared without anyone announcing it. He’s not anti-AI. He’s productive with it. The problem, as Scollon writes on alecscollon.com, is repetition: LLMs write in the same style and make the same kinds of mistakes, and dealing with the same thing over and over grinds you down.

The top HN comment made the deeper diagnosis: what used to be varied, skillful craft became standing in one spot on an assembly line, because LLMs took the creative and variable part of the work and left the repetitive QA rubber-stamping. That’s the burnout. Not too much AI – too much AI reviewing.

Method A vs Method B: Detox vs Rebalance

Two schools of thought are fighting for your attention right now. Only one matches what people are actually describing.

Method A comes from the addiction framing. Bloomberg’s May 2025 piece by Parmy Olson profiled a young worker going cold turkey – the flagship example of this approach. It works for compulsive use. It does not work if your problem is being sick of AI prose. Going cold turkey doesn’t un-teach you the patterns you’re now allergic to. You’ll come back in two weeks and see exactly the same stylistic tics, because the models haven’t changed.

Method B is what the actual post recommends between the lines. It’s what we’ll walk through.

The 5-step LLM burnout reset

Do these in order. About 20 minutes end to end.

Step 1 – Audit where your day actually goes

One workday, tag every AI interaction with one of three labels: Create (you wrote a prompt to make something new), Refactor (you fed it something to change), or Review (you read output someone or something else produced). Count them at end of day. If Review tops 50% of your interactions – there’s your burnout source. The assembly-line QA diagnosis from the HN thread isn’t abstract; it shows up in the numbers.

Step 2 – Reclaim one whole task from AI

Pick the task where you feel the most “ugh” opening the chat. Emails. Commit messages. Standup notes. Whichever one made you click this article’s title.

Ban AI from that one task for seven days. Not all tasks. Not all AI. One task, one week. Targeted at the specific well that’s poisoning your relationship with the tool – not a scorched-earth abstinence campaign.

Step 3 – Rewrite your Custom Instructions to fight sameness

The repetition problem is real but partially fixable. Open your ChatGPT Custom Instructions (or Claude’s system prompt / Projects setup) and paste something like:

Do not use these words: delve, leverage, robust, seamless,
streamline, tapestry, landscape, empower, unlock, harness.
Do not start responses with "Great question" or "Certainly".
Do not end responses with a summary paragraph that restates
what you just said.
Do not use em-dashes as a stylistic tic.
When listing items, vary the format - do not default to bullets
for everything.
If you don't know something, say so in one sentence.

Scollon’s caveat: even with personalization features, some idiosyncrasies seep through. Custom instructions won’t make the model sound like a human – but they’ll cut the specific patterns that make your eye twitch noticeably. Worth 90 seconds of setup.

Step 4 – Cap your session length

Odd parallel from the coding-agent world: every model tested eventually melts down. One Medium case study documented Claude surrendering after roughly 64 iterations on a nested Podman debugging session – literally writing “I’m sorry, but after 3+ hours we’ve hit a wall.” The model’s fatigue becomes your fatigue by proxy.

If a single chat has more than 30 back-and-forths on one task, start a new session with a fresh summary of what you’ve learned. You’re not obligated to ride a chat until it collapses.

Step 5 – Add a “human input” quota

For every hour spent reading LLM output, spend 15 minutes reading something a human deliberately wrote – a book chapter, a well-edited newsletter, a peer-reviewed paper. Calibration, not wellness. You lose the ability to detect AI patterns when your entire reading diet is AI patterns.

The fastest tell that your reset is working isn’t that you use AI less – it’s that you notice yourself editing AI output instead of accepting it. Still accepting first drafts? You haven’t recalibrated yet.

A quiet observation

There’s something strange about needing a viral blog post to give this feeling a name. People have been experiencing it quietly for months, assuming it was personal – a moral failing, or a sign they were becoming a Luddite. Turns out it was a workflow shape common to millions of jobs, and nobody had labeled it yet. Whether “LLM burnout” is the label that sticks or gets replaced by something more precise in six months, I don’t know. But it’s real either way.

Edge cases the other tutorials skip

The polluted-search problem. Cutting back is harder than it looks partly because, as Scollon notes, useless AI-generated articles clutter the search results – so falling back to “just Google it” often delivers more AI slop, laundered through a domain name. Workaround: use search operators to restrict to sources with human bylines (site:news.ycombinator.com, site:reddit.com), or bookmark 5-10 trusted sources and browse them directly for a week.

The withdrawal isn’t imaginary. Productivity anxiety during ChatGPT outages? Real, documented in a CHI 2025 diary study on LLM withdrawal in knowledge workers (arXiv preprint, CHI 2025 – note: verify URL independently, as preprint IDs can shift). If your one-task ban in Step 2 produces mild panic, that’s normal – not evidence you should quit the ban.

Overreliance is separately measurable. A CHI 2025 paper by Spatharioti et al. found LLM-based search changes decision speed and accuracy in ways users don’t notice. You may be worse at your job with AI in specific domains you can’t feel. Step 1’s audit is how you find out – the numbers don’t lie to you the way your intuition does.

FAQ

Is LLM burnout the same thing as ChatGPT addiction?

No. Addiction is compulsive use despite consequences. Burnout is exhaustion from over-exposure to a specific pattern. Completely different problems with different fixes.

Should I quit AI entirely if I’m feeling this way?

Probably not – and here’s what most detox guides miss. If you quit entirely, you come back in two weeks and see the exact same stylistic patterns, because the models haven’t changed. What works instead: a diet change. Someone using AI for 30 minutes of brainstorming a day will burn out much slower than someone spending 6 hours reviewing agent-generated pull requests – even if the total “AI exposure” is technically smaller for the second person. The type of use matters more than the volume.

What if my job is literally to review AI output?

Then this is a management conversation, not a personal-productivity one. Bring the HN thread and Scollon’s post to your next 1:1. The assembly-line-QA framing lands with non-technical managers in a way that “I’m burned out” doesn’t – it describes a structural role change, not a personal complaint. That’s a different conversation, and a more winnable one.

Your next 5 minutes

Close this tab. Open your Custom Instructions. Paste the block from Step 3. Then pick the one task for Step 2’s seven-day ban and write it on a sticky note where you’ll see it. That’s the whole starter kit – everything else is refinement.