Too many slop grenades these days. LLMs are cool and all, but does every Slack message need to be a bulleted list where the first sentence of each item is bold?

I recently read one too many blog posts with the phrase, “it’s not X, it’s Y” as a one-line paragraph and finally snapped. It is now my life’s mission to purge selfish LLM usage from the internet.

Define “selfish”

Selfish LLM usage: A situation in which someone uses an LLM to save their own time at the expense of other people’s time, resulting in a net productivity loss.

Opinions vary on the usefulness of LLMs, but hopefully we can all agree that using them selfishly is cringe.

Nevertheless, people use AI selfishly all the time. This is not surprising: many people are under heavy time pressure and some people aren’t even aware they’re doing it. But the world would be a better place if people stopped.

But it’s not polite to call someone out for blatantly using AI when their brain would have sufficed. The trick is subtlety. I propose a method of subtly shaming people with plausible deniability:

1. Secretly make fun of people via emojis
2. Other things (less fun)

Dog whistle emojis

React to suspected AI content with an emoji that only makes sense to likeminded people. Here are some ideas with variously intense connotations:

Passive-aggressive emojis

Respond to selfish AI-generated content with these emojis if you want to keep a low profile while secretly making fun of the author.

Fully aggressive emojis

If you don’t mind being fired and/or punched in the face, respond to slop with the following emojis:

Positive emojis

Reward human effort with compliment emojis:

Yes, the amphora emoji is also on the passive-aggressive list because irony, but obviously it works for human content too. By reacting to both kinds of content with it, you gain plausible deniability while still subliminally encouraging people to think about the importance of humanity and artistic expression.

Emojis don’t cut deep enough

I admit the dog whistle emoji thing is at least 20% sarcastic.

If you want to be less culty, probably the most neutral and unambiguous emoji is robot face 🤖.

Semi-jokes aside, here are some better things you can do to shape the world to your liking:

Positive reinforcement

If someone sends rough (but clearly human) thoughts, engage with it positively. Be extra nice and chill.

Ask genuine questions that force the slop author to reflect on what they produced and answer directly from their human brain. The questions should be short and simple. Match the author’s effort level.

Slauthor?

• How did you come up with this idea?
• I really liked [interesting detail], how did you learn to [verb] like that?

Epmathy

I spelled this heading exactly as intended.

Keep in mind that the average person over-relying on AI for common tasks is potentially under immense time pressure and/or sufficiently out of their depth that they lack the ability to distinguish good from bad content. They almost certainly are not using AI out of malice.

Turns out the average person does not like slop, so it’s not hard to ringlead your friends into improving the situation.

For example, my software engineering team at work has been fatigued by AI writing cluttering up our pull requests, so we added a temporary PR checklist to help us get in the habit of making the code easier to review:

## PR Checklist

<!-- All PRs -->

- [ ] Description is human-written, concise, and useful

<!-- Medium complexity PRs also include -->

- [ ] Screenshot/video attached for frontend changes, or real
      examples of backend behavior
- [ ] Code comments on tricky parts are human-written and
      concise (no AI wall of text)
- [ ] Tests verify important functionality and are
      straightforward (not brittle, noisy, implementation-quirk
      tests)

<!-- Large / risky PRs also include -->

- [ ] Assumptions, weak points, follow up, and open questions
      called out
- [ ] GitHub comments signpost areas most important to the
      reviewer
- [ ] Scope is as narrow as it reasonably can be (or noted why
      it isn't split)

This also led to me creating a comment guard script for Claude Code, which prevents comments longer than two lines long. Highly recommend.

Conclusion

When discussing this post please include 🏺