Used to be, people had to bully or pay someone to do their work. Now, AI promises a victim-free, no-cost (or super-cheap) online robot instead.

Even a shallow dip into the AI pond reveals that a great deal of AI-generated content is pretty bad, but that doesn’t have to mean an AI user is any less interested in quality. How can the human editor help clients who use AI produce an impressive final draft?

The following is a step-by-step guide to editing (but not co-authoring) AI-generated content based on my research into and personal experiences with AI text creators.

Every editor I know respects language. We learn its rules, its preferences, its limitations, and its possibilities. We promote creativity and tend to sneer at trite, clichéd writing, and we get paid to do it.

So it’s with little surprise that I’ve seen colleagues react to AI-generated content with disdain, if not disgust. The idea of some computer program assembling text from various sources of unmediated quality and passing it off as “writing” seems not just worthless, but somehow unsanitary.

But we editors are going to have to suck it up. AI generates product quickly, and in today’s text-heavy market the fastest typist gets hired. Using a website or app to generate text is here to stay.

I resisted the whole thing myself, at first, but with effort I have found that when approached correctly, editing AI-generated copy can be a satisfying challenge with an effective outcome.

I need to address the most heated argument against AI-generated text, which is that students will use it to cheat. Make no mistake, they will, whether AI is marketed to them or not.

But teachers have been watching plagiarists cut-and-paste from the internet for decades, and before that students just copied from books and each other. Dealing with cheaters is part of a teacher’s job, and they will cope with AI too. Leave all that to them.

The AI-generated content getting sent to editors is text that makes money. We’re talking:

• News Articles (Press Releases, Feeds, Clickbait)
• Marketing Materials (Posts, Emails, Websites, Ad Copy)
• Blog and Social Media Posts (Tips, Guides, Testimonials)
• Product Descriptions

This is fast-paced commercial material. The editor is part of a team to take quickly generated but highly flawed copy and help make it impactful and engaging.

Some AI content generators promote “customization,” but it’s pretty limited. When working with AI copy, as with any copy, make sure you and your client agree on what they and/or their company want to achieve. Is this supposed to motivate people to subscribe? Is it selling a service? Is it meant to generate controversy? As for the audience, what grade-level vocabulary are we talking about here? Is this for newbies who need help or for professionals who don’t want definitions for technical terms?

An added issue for AI copy is to see if you can get the exact AI prompt(s) the client used to generate the text can so you can tell when the copy digresses from that prompt. Say the prompt is “America needs gun control now,” and the AI copy includes an explanation of how bullets work. That explanation can probably go.

Is the whole thing you’re editing done by AI, or is there some stuff in there your client wrote? You want to edit it all either way, but it will help you to know which you’re dealing with. You can ask the client, but even without them, spotting AI copy is pretty easy. It will doubtlessly have some or all of the following:

• Frequent repetition of phrases and ideas;
• Breaks in logic;
• A lack of emotional investment or personal expression;
• Pointless digressions;
• No transitions;
• “Facts” that conflict or just make no sense; and
• An excess of exposition with no opinion or controversial content. (ChatGPT in particular does this.)

When it comes to gathering data, AIs have no standards. If it’s out there, they’ll use it, so fact checking is vital. The client, being the subject matter expert, should do this before you see it, but if they don’t, either make sure you’re getting paid to do it (Most editors are expected to fact check only with an extra fee.), or highlight suspicious-looking text and tell the client to double-check.

Recently, CNET pulled back from using AI copy because it found errors in more than half of the AI-generated content it was publishing. What sort of facts need to be checked?

• Names (and the spelling of names);
• Statistics;
• Dates;
• Historical background;
• Quotes and references; and
• Gossip/urban legends treated as facts.

Line editing for AI copy isn’t really different from editing human writing. There are just some extras to watch out for.

AIs claim to be great at spotting plagiarism, but they aren’t. Either they label pretty much anything that’s ever been said before as “suspect,” or you can get a clean pass by taking the national anthem and just changing some words around. So strap your plagiarism goggles on.

And don’t forget about graphics, including photos. Many clients and AIs will grab things off the internet with no regard to copyright issues. Know the basics of fair use and apply them correctly.

Surprisingly, for a robot, AIs can be very weak in mechanics. It’s so bad some articles out there will recommend doing this part of the edit first, but why edit content you’re going to change?

Checking AI grammar and style works the same as for human writing, but you will notice a different pattern to the errors. If you’ve edited ESL/EFL texts, and most of us have, you know that different native language writers make different sorts of mistakes. Asian language speakers tend to have problems with “a” and “the.” Arabic speakers often have long sentences that end in a verb. Spanish and French speakers tend to be fine with grammar and structure but fumble with idioms.

AI copy will make grammar mistakes that defy the rules of usage, particularly in terms of phrases that don’t fit into the rest of the sentence. The sentences themselves often have the same structure over and over, offer erratic word choices, and, as I noted above, repeat stuff more than a freshman at 3 a.m. desperately stretching for the word count.

AI copy also tends to be riddled with clichés. You don’t have to replace these clichés, but you should point them out to the client. Tired phrases are easy to change with just a word or two.

And you know, there’s nothing wrong with using AI to help out, especially if you can use a different AI from whatever created the copy. There’s Word’s spell check, of course, and if you’re dealing with something from ChatGPT, a stop at Grammarly will find some (though hardly all) errors quickly.

Because AI tends to spit out text without consideration of overall structure, you need to spend extra time on the introduction looking for a good hook, see if there’s a solid conclusion or gaps in flow, and do what you can with transitions. Sometimes it’s just a matter of tweaking a few words. Sometimes you need to tell the client to come up with something new.

Now’s a good time to look over the formatting. Would a table, chart, or other graphic be a good suggestion? How about some bullet points? If you’re doing APA, you’ll need to make sure those headings are in order, and no matter what style the client is using, you need to check for format consistency.

OK, you’ve finished the line edit, but there’s always more work.

While the webmaster/blogger/publisher should be left to do the meta work on search engine optimization, the editor can contribute to SEO in the text itself. AI copy clients will be particularly interested in this. As Denis Manyinsa notes, “Publishers who use AI to write risk having their site penalized by search engines. According to Google, sites with lots of unhelpful content are less likely to rank well in Search.”

There are tons of articles out there on SEO, which I encourage any digital text editor to read. Here are some common tips:

• Be mindful of keywords. If the document is talking about the new trend of protecting personal data with a digital wallet, consistently call it a “digital wallet,” not “digital token keeper,” whether you think that’s a better name or not. And the more on-trend word in 2023 is to “gatekeep” data, not “protect.” (Google is our friend.)
• Make hyperlink text reflect the destination. If the link goes to an article on digital wallets, don’t make the link’s text “like this” or “see more here.” Make it “as this article on digital wallets explains,” or something like that.
• If the text is for a casual reader, consider replacing “extrapolate” with “conclude” and breaking up that fifty-word sentence.
• Pay attention to that title or headline, which is optimally around six to eight words. Because common words are used in search engine queries, titles/heading with 20–30% common words such as “why,” “so,” and “is” rate a higher SEO. The same amount of uncommon words also helps. You’re going for a balance. Also, positive titles/headlines get more traffic than negative.

I’ve been mentioning it all through this article, and now I’m making it a step in its own right: tell the client when something is missing, but don’t write it yourself, just as you wouldn’t write it for a human client.

AI copy can be so painfully, obviously lacking something that the urge to just jump in there and add it yourself is strong. But this is outside an editor’s purview and might irritate your client. So double-check you didn’t overstep.

When something is missing (e.g., more examples, a demonstration, a definition, personal experience), tell the client about it and move on. Considering the limitations of AI, good comments might include:

• An analogy might be useful to explain how this works. (AIs have no skill with figurative language.)
• A tasteful joke might allow the audience to rest for a moment before the next PowerPoint slide. (Most AIs’ attempts at humor are cringe.)
• An emotional appeal here might move your reader to action. (AIs don’t do emotional writing.)

Look at what you’ve done and ask yourself the ultimate new question for AI-generated copy: does it sound like a person wrote it?

I whipped up some AI-generated copy with the prompt, “Describe to me a world where every day is like Mardi Gras.” It made 410 words, so here’s an excerpt:

The spirit of camaraderie and inclusivity infuses every interaction. Strangers become instant friends, and the boundaries between social classes, backgrounds, and cultures blur as people unite in the spirit of revelry. Random acts of kindness and infectious laughter fill the air, as the city becomes a playground of unadulterated joy and genuine connections.

In this Mardi Gras utopia, food takes center stage as a carnival of flavors and aromas. Streets are lined with food stalls offering a tantalizing array of delectable treats. The air is thick with the scents of freshly fried beignets, succulent jambalaya, spicy gumbo, and sizzling crawfish boils. Exotic cocktails and vibrant libations flow freely, bringing an added touch of indulgence to the celebrations.

I know the vocabulary is impressive, but does that sound human to you? Read it more than once if your initial answer is yes.

Note the total lack of nuance with “every interaction” and “instant friends,” where all “boundaries blur,” and there is “unadulterated joy.” Indeed, “every” shows up seven times in those 410 words. And all the language is turned up to 11, with “delectable treats” and “vibrant libations,” which manages to sound both odd and clichéd.

This is more than just travelogue hyperbole. This is 100 percent wonderful, an impossible and inhuman description of wind-up toy revelers without identity like a synchronized CGI background.

So, let’s take our good attitude and assess our task. The prompt makes it clear the client wants something festive, but not so relentlessly positive that imagining “every day” feels like sensory overload.

Now we fact check, and we see this is impossible. I love Mardi Gras, but there’s no human world where “the boundaries between social classes, backgrounds, and cultures” just go away. And if you drink the way Mardi Gras tourists drink, you not only add more than a “touch of indulgence” to your parade, your new “friends” are going to ditch you when you start puking in the gutter.

There’s no overt plagiarism in the piece, though it is full of tired phrases like “the spirit of camaraderie” and “random acts of kindness,” while the “air is thick.”

For grammar, the comma after “fill the air” needs to go, and it should be “spirits of camaraderie and inclusivity” because those are different things. We’ve also got two mixed metaphors with “takes center stage as a carnival of flavors” and “random acts of kindness . . . fill the air.” Carnivals don’t take the stage, and acts don’t fill the air.

Because I’m just looking at an excerpt here, some of my ten steps don’t apply, but I must remember to add no content. Can we keep to our role as editors and make this sound like human-generated copy? I’ll do the first paragraph. Try doing the second on your own.

The spirits of camaraderie and inclusivity prevail as people readily make new friends. People seem less interested in the boundaries between social classes, backgrounds, and cultures and more in uniting to have a good time. Acts of kindness and infectious laughter are common as the city becomes a playground for those seeking joy and human connection.

Well, it won’t win a prize, but it sounds a lot less like a robot. How did you do?

All things considered, an AI is just another bad writer. Take some extra precautions, be prepared for new types of ineptitude, and the end product can still make your client come back for more.

Of course, your best move is to hire a great editing service, such as Edit911. We’ll fix everything for you, so you sound like you and not a robot.