Why human editing still matters in AI-generated content

Generative artificial intelligence (AI) content has improved remarkably since the first platforms appeared a few years ago. Those early versions produced flowery writing with little structure and obvious repetition, while today’s systems generate far more coherent and natural-sounding prose. However, underlying issues, such as generic phrasing and a lack of specificity, still remain.

The growing sophistication of AI content has also made detection far less reliable. When a data scientist fed the US Declaration of Independence into an AI-detection tool, it said 97 per cent of the document was written with AI. It sounds absurd, but the reason is simple: AI has gotten so good at mimicking human prose that detection tools do not always keep up.

“AI has been trained on the formulas of good writing such as the rule of three, juxtaposition and formal cadence,” says Ralph Grayden, an AI writing specialist and co-founder of Antelope Media. “It almost writes too well and is very formulaic.”

Even so, AI-generated writing is not entirely beyond detection. When a book publisher recently pulled a novel from shelves after suspected AI use, it was not a detection tool that raised the alarm. Readers pointed to flat emotional tone, repetitive sentence structures and heavy use of lists of three.

The problem is not that AI writes badly. It is that it has common, recognisable characteristics. The key is learning to recognise these patterns, correct them and add unique information.

How AI is trained

Dr Kavita Ganesan, founder and chief AI strategist and architect at Opinosis Analytics, argues that these systems have absorbed enough books, reports and web content to mistake formulaic language for good writing.

“The knowledge embedded in them is like a PhD in English and other fields such as a PhD in physics, computer science and political science, combined. The problem is, they express it by always reaching for the most obvious word,” she says.

And it is getting worse. AI models are increasingly training on their own output rather than original human prose, creating a feedback loop that makes the writing more generic over time. In fact, Ganesan argues that some newer models sound more generic and have less personality than the older models.

“There is no critical thinking in AI-generated content,” she says. “The models generating output are not really thinking, they are just stating (or rather, generating) what is most plausible given what you have asked for.

It is a useful first draft if you are already a subject matter expert on the topic.”

What to look for

At its worst, AI writing a few years ago might have sounded like this: “As we stand at the crossroads of a transformative moment in human history, organisations that fail to leverage synergistic opportunities risk being left behind in an increasingly competitive landscape”. While that might be grammatically fine, is it something that a real person would say?

Today it is more subtle. “Most people generate something, read it back and think it sounds okay — because it does, technically,” says Grayden. “Spotting the problem requires a specific lens, and developing that lens is where most people get stuck.”

 

"Verify the facts, because if AI does not know something, it will make it up. It is really important to understand that you have got to edit the content and treat everything it produces as a first draft."

 

Ganesan identifies the most common signs: vague generalisations, overuse of power words, repetitive sentence structures, filler words and phrases, and a lack of personal voice or real-world context.

Grayden has a different test. “The one that really stands out is the em dash [—]. Most people would not know how to put an em dash into a document, and yet AI uses it all the time.”

How to fix it

The fixes are not complicated, Grayden says. Stripping out what makes it sound like AI and replacing it with specifics changes the content immediately.

For example, replace filler phrases with a specific detail, a case study or a point of view that reflects actual experience. Instead of “businesses must navigate complex regulatory environments”, say which regulation, which industry and what it actually means in practice.

The other tip is to give the AI tool direction on the type of voice and tone needed. “Most companies have a style guide,” says Grayden. “Feed it to the AI [in line with workplace protocols] before you write. You will get a different result straight away.”

Ganesan agrees that specificity is everything. “The best approach is to tell it exactly how you want it to sound and give it a persona.”

Always check the facts

The other essential step is fact-checking. AI generates false references with the same confidence it generates real ones.

“You cannot just press a button and leave it all to AI,” Grayden asserts. “Verify the facts, because if AI does not know something, it will make it up. It is really important to understand that you have got to edit the content and treat everything it produces as a first draft.”

Routine emails are fine to hand off to AI, as long as they are edited first, says Grayden. Thought leadership, content that affects reputation and considered opinions on major transactions are not.

“AI is going to save you time, but you still need to go back in, check the facts and write it properly. One of the biggest questions professionals face right now is where to draw that line.”

His advice is straightforward: “If your name is on it, you take the lead.”