LLM etiquette | Paperless

etiquette

LLMs

writing

Or, how to treat other humans as such, in an age of generated text.

0. Respect the readers

If you read nothing else, read this:

Don’t waste readers’ time.

Don’t treat the LLM as a source of indisputable facts.

Don’t pretend that you wrote text you didn’t write.

1. Respond as you

Manually rewrite LLM responses in your own words before putting them where<br>other humans can see them. You have a personal way of writing, and familiarity<br>is important when communicating with humans.

You might worry that your own writing has more spelling mistakes than the LLM<br>response. That can be fixed by sending the text through any spell checker at<br>your disposal - any decent text editor does this out of the box.

You might worry that your own writing is less eloquent. Don’t. LLM output is<br>extremely verbose, and tedious to read. Unless you’re deliberately trying to be<br>prolix, or to obscure what<br>you’re actually saying, your own words are better. They will also feel<br>familiar to anyone who knows you. An unfamiliar style can cause readers to get<br>cognitive dissonance, subconsciously biasing them against whatever the text is<br>saying.

You might worry that your writing is bad. Is copy-pasting LLM responses going to<br>make it better? And how good is LLM writing, really? The LLM will scramble what<br>you meant to say, wrap it up in a pretty package of mostly correct grammar and<br>spelling, using many more words than necessary, and in a tone which is<br>completely different from what people expect from you. That’s not good writing,<br>that’s overly verbose, generic slop.

If nothing else, do it because a lot of platforms (like<br>Stack Overflow and<br>Wikipedia)<br>enforce this, and will react appropriately if you post LLM content as yourself.

2. Fact check

Treat every statement in the LLM response as something which needs verifying.<br>The LLM is implemented to produce a “reasonable” response to your query, not to<br>fact check them, look for<br>falsified data, or ignore<br>bad interpretations. That job is yours. And if<br>you don’t do it, people will naturally ascribe any faults to you, not the LLM.<br>If you don’t want to be known as a liar, do your due diligence before repeating<br>something produced by a machine.

LLMs are great tools for researching a topic, but no amount of flagship models<br>and clever prompts is going to remove the need for thinking. Unless, of course,<br>you’re suffering from<br>AI mania.<br>In that case I’m truly sorry, but I’m also not going to read your slop.

3. Cite

Sometimes, posting the LLM response verbatim is warranted. Obviously when<br>discussing LLMs, but also to inform readers of sources of your own bias. Since<br>LLMs aren’t deterministic, it’s important to include all the information someone<br>else would need to get as close as possible to the response you got: the model<br>name, version, context, and prompt. You want to make checking your work as<br>easy as possible.

Acknowledgements

Thanks to Borja, Kit, and Sarah for their reviews.

Comments? Feedback? Corrections? Send an email.

Version history

No webmentions were found.