Tips for using AI in writing and editing

It's been a long time

After my last post, where I was enthusiastic about starting again in the text-only web, two and a half years have passed. Among other things, I got anxious about posting too often. Here, daily would be too much. And I am typically so overwhelmed with work that I have time for little else.

Well, there's time to try again. And maybe once a week is the right cadence if I can manage it.

I've worked in the publishing industry as an editor and copywriter for about 20 years. Much of that has been as a full-time freelancer. A year ago, I had two months almost without work at a time when I was already financially strained, and it made me afraid about the future in a way I haven't often felt. It made it worse that some of the old contacts I reached out to looking for jobs said things like, 'I'm not surprised you're quiet; isn't everyone getting AI to do that these days?'

The fear that I'll be replaced isn't exactly gone, but things have turned out a bit differently than I expected. Among other things, since then I've had two of my busiest months ever. And AI is a part of that. People are writing (and 'writing') more text, and that text contains new sources of error and stylistic deficiency that the AI creates. That means more work for editors, not less.

The tips

Yesterday an academic researcher wrote asking for advice on what AI tools they could use, and I felt happy to give some tips.

I was going to paste the exact notes I sent to her, but I'll summarise instead.

The big one: don't let AI work directly on your text. Use it as a conversational partner. Give it drafts and ask questions about things like errors of fact and reasoning that you may have missed, how you could fix sentences you're not happy with. Upload collections of source documents and ask it questions about what's in them. After your own, human editing passes, ask for a final review of remaining issues. Get a list and implement them yourself.
I mostly use Claude now. It is better at following instructions and avoiding hallucination than ChatGPT is. Its use of full-context retrieval makes it superior at working with uploaded documents.
Best practice can be to turn off any options like 'improve the model for others', to minimise the risk of sensitive inputs being used in training data. At least one organisation I work with has made this part of their AI use policy.
Don't assume that an AI tool has accurate knowledge about editorial style (eg, about what is in the Chicago manual or what a particular journal requires). Chatbots will frequently make incorrect assumptions about the style used by specific guides and publications, and suggestions about grammar can be wrong or unaware of stylistic nuances.
Working on text or code in small chunks can be more reliable than working on whole files, and gives you a better chance to spot errors in the output.
AI tools can be super useful in writing boilerplate code or troubleshooting coding or other technical errors. If you use LaTeX (for example), it really is a great tool for solving compilation errors or working out how to code for a desired document appearance. Chatbots are particularly good at reading error logs.
Be very wary of citations that AI tools have touched. I once edited a 90-page report where nearly every citation was at least partially fabricated.

Thoughts welcome.

Write a reply