Why I do Not Use Generative AI for Editing

I have used AI in work unrelated to editing, but in my view there are some serious problems with using AI for editing work. I have been testing out and using generative AI for a few years now, and the criticism posted here is the result of a sincere effort to make AI work for me, not a simple hand-waving dismissal of new technology.

My main reason for not using AI for editing is the issue of privacy loss. The policies of each company may vary slightly, but in general, using AI to edit someone’s writing means uploading it to the web and leaving a permanent copy of it there, to be used to train AI systems and putting it at risk of being stolen or leaked.

AI has difficulty following the simplest directions even when clearly spelled out. To try to prepare a reading activity for a reading course, I directed a well-known and widely trusted AI system to copy paragraphs from an article and reproduce them from the source without images. (Of course, I had permission from the publisher to do this.) On one occasion, it presented the sentences out of order. On another, it paraphrased random sentences from the article. I then changed the prompt to emphasize not to do those things. The results improved temporarily, but it soon went back to manipulating the text in ways it was not directed to. I understand that this partly comes from the way generative AI is designed, and it can be a huge benefit in some ways. For editing, though, it is inconvenient to have an assistant that constantly tries new things out and learns by tweaking its environment.

AI is error-prone. Developers are constantly trying to improve the accuracy of its responses, but the errors continue to appear. In one instance, I asked an AI system to select words from an article and produce simple definitions and sample sentences for each. One time in four, the system included words that did not even appear in the article. From time to time, the system froze in the middle of executing a request, requiring me to resend the prompt and start all over.

The quality of the output of an AI system, depending on the intended content, is often mediocre. Whenever it is asked to write, the sentences are typically short and lacking details or a clear context. Even when directed to add specifics, the results are not much better. When producing sample sentences with AI for an EFL class activity, I regularly had to rewrite most of them.

Different workarounds have been posted on the web, and I have tried them. Yet with all of the monitoring required to be able to use AI effectively and safely, it is much more time-consuming to prepare the detailed instructions necessary, have it perform the tasks, and then have to recheck the entire document to be sure it did a proper job and did not introduce new errors. I have found the process more efficient by doing without generative AI.