top of page
  • Writer's pictureEffy Lindström

Tips for AI-Assisted Writing: Six Dos and Don’ts

When I was in the sixth grade, I brought a Christmas gift to school that won me a full day’s worth of clout in an otherwise geeky and largely overlooked existence. The item in question was a miniature American Girl “laptop” that served as an address book, with slots for around two dozen names and phone numbers, rendered in a seven-segment display.


So awe-inspiring was this technology that it induced every popular girl in the glass to buck OPSEC and give the art freak their hitherto closely-guarded digits. Naturally, I was on cloud nine.


Now, roughly two decades later, students can generate an entire research paper from ChatGPT. Good news for lazy pupils, and bad news for those hoping to impress Regina George with the latest gadgetry from Mattel.


Of course, the halls of learning haven’t been the only area forever changed by the advent of AI. Most notable of all has been the neural network-sized crater left in the world of professional writing. Would AI soon replace us? Or would it revolutionise our workflows, making our field more profitable than ever?


So far, results have varied. But for those who are aiming for the latter, I do have a few tips and tricks, gleaned from a year’s worth of near-daily tinkering with AI on the job. And just like my American Girl laptop, I am all too happy to share my findings with the class.


Behold: what passed for advanced technology in the noughties.


Do use AI for preliminary research.


Back in the days of yore–that is, pre-autumn 2022–one of the most time-consuming aspects of writing an article was the research phase. This was particularly true for covering current events, as online news tends to take the form of succinct, digestible snapshots rather than comprehensive timelines.


When ChatGPT quantum-leaped onto the scene, however, everything changed. Suddenly, one could obtain a primer on everything from the Cretaceous period to famous rivalries in hip hop. These overviews could be long or short, take multiple formats, and be adjusted for specificity. One could request, for example: Give me a bulleted, chronological breakdown of Azealia Banks’ messiest feuds–and don’t leave out her beef with the entire country of Sweden! (That last bit is optional, yet well-worth the clickhole.)


Where things got dodgy was developing stories, as ChatGPT’s training data cut off at 2021. This meant that if you prompted it to recap Charles III’s first acts as king–well, as far as the bot was concerned, we were still in the New Elizabethan age.


Granted, there was a workaround, as one could manually feed the AI recent articles and ask it to provide a summary. Though I still sometimes use this technique (for reasons I will explain later), it is no longer a necessity as of September 2023, when OpenAI bestowed upon the chatbot the power to browse the Internet.


For content writers the world over, these developments have significantly truncated the research phase. There is one caveat, however…


Se upp, Sverige! Azealia är på krigsstigen.


Don’t use AI for fact-checking.


There can be no doubt that ChatGPT is a powerful tool for taking the first steps in research, especially if you’re going into a topic blind.


But as with anything with the machine learning model, it’s important to be mindful of its limitations.


In my experience, ChatGPT’s factual errors range from small details (like the number of times Nixon visited the Soviet Union) to complete fabrications. Though the former is far more common, the latter does occasionally–and alarmingly–strike.


Take, for example, the time I asked ChatGPT to summarise the events surrounding a recent chemical spill. For reasons unknown, the digital assistant took it upon itself to include elaborate and highly specific details of the local environment’s ecosystem, health risks to the community, and cleanup efforts from public officials. There was just one problem–the information on the ground was scarce, meaning all of these intricacies were complete bunk.


This brings us to why I still, from time to time, feed the bot a selection of articles about a given topic rather than asking it to browse the Internet. That way, if it generates an anecdote or statistic that sounds a little outlandish, I can easily search the source articles for verification.


Bottom line: Always remember that when you ask ChatGPT to pull information from the net, it could have sourced those “facts” from anywhere. So you must, must, must do your own fact-checking.



For the record, Nixon visited the Soviet Union three times before his first presidential visit.


Do use AI to organise information.


ADHD generation, rejoice! ChatGPT has an oft-overlooked ability that can easily bypass your executive dysfunction: namely, the organisation of information. 


This, in my opinion, is the most useful aspect of the language model. As we will touch on in the next section, ChatGPT is not skilled at rendering prose, but it does tend to do a good job of highlighting the most salient facts of a given topic and ordering them in a sensible manner.


There are two ways you can go about this. First, you can ask the AI to create an outline of your article. Though this is straightforward enough, concrete learners (like me) will benefit most from asking ChatGPT to generate an article according to their specifications, and then drawing inspiration from the way the chatbot summarises and structures the information.


Just remember, though…


Do not use ChatGPT’s wording.


While it may sound obvious, writers should always put AI-generated text into their own words. That means that you take cues from the way ChatGPT sections articles and arranges facts–but never the way it crafts language.


This is for a couple of reasons. First, doing so may be bad for SEO. While it’s important to stress that Google doesn’t directly ding web sites for using AI-generated text, our Californian overlords do penalise low-quality, untrustworthy material. And if you aren’t fact-checking and rewriting ChatGPT’s output, your content will probably meet that description.


This ties into the second reason: ChatGPT’s writing style leaves much to be desired. It’s stiff, at best, and unbearably corny at worst. It also frequently lacks a natural sense of flow and transition. Whether you are a junior writer or seasoned professional, I can all but guarantee that your own wording will be superior, so don’t give in to the temptation to cut corners.


Just write as confidently as you can, and follow the next tip if you find yourself stuck.



Do use ChatGPT to assist with word choice.


One of the most frustrating aspects of writing is drawing a blank when you don’t know what word or phrase to use, as this can seriously disrupt your flow whilst “in the zone”.


In the past, I have gotten around this by typing three underscores whenever I can’t decide on the appropriate word choice, and then revisiting the blanks once the first draft is completed. This allows me to maintain a reasonably speedy pace–without interruptions–-during my first go-through.


I would recommend this method to any writer, particularly now that AI has made the second part so easy. Simply feed one of your uncompleted sentences into the chatbot with the following prompt: “Without regenerating the rest of the text, please fill in the blank, and give me multiple options.”


Let’s break that down. The reason I ask ChatGPT not to regenerate the rest of the text is because I typically feed the AI the entire surrounding paragraph for context. The additional information makes the suggestions more accurate–but I don’t need to see it again (especially when my internet connection is running a bit slow).


The request for multiple options is due to the fact that ChatGPT rarely gets it right the first try, and users are sometimes limited to a certain number of messages per hour. It’s more expedient, therefore, to ask for an array of suggestions–or “a dozen” for good measure. Usually, the AI identifies a suitable word or phrase within those parameters.


Oh, and it’s good to say “please,” because research shows that polite encouragement yields better results. Does this mean language models dream of electric sheep?


Well, not quite yet, as the development of sentience will probably come several steps after it gains the ability to edit. Speaking of which…


Android tripe, anyone?


Don’t rely on ChatGPT for editing.


If ChatGPT’s writing skills are insufficient, its editing abilities are far worse–a lucky break for those of us who have built our careers off the latter.


Admittedly, I was curious to see if it could do my job. Which is why I have fed unpolished text into it using a multitude of different prompts, all with the same result: soulless, long-winded tosh.


There is one chief reason for ChatGPT’s poor editing skills: It cannot restrain itself from significantly rewriting the text, rather than taking a surgical approach to the bits that need fixing. And when it does so, it employs the same uncanny valley-inducing writing style–blegh.


So please, hit up your friendly neighbourhood, flesh-and-blood copyeditor if you require a second pair of eyes. ChatGPT cannot refine your copy; it can only transmute into a sterile, hydrogenated cube of textual slop. 


Final Thoughts


Having led multiple writing teams, I can say with certainty: It is glaringly obvious when you use unedited AI-generated text, particularly for those of us who have taken the time to learn its many “tells”.


Perhaps one day this will change, and I will find myself eating unemployment-flavoured crow. But until then, use AI for what it’s actually good at–like summarising and assembling information. When utilised in these ways, the machine learning model transforms from a blunt instrument to the most useful tool in your kit, saving you time and increasing your output.


And of course, I will continue to post updates as ChatGPT inches closer and closer to transcending the boundaries of human intelligence.  Let’s just hope that when it does, our digital breakup is more like Her than Ex Machina





30 views0 comments

Comments


bottom of page