The 90% Test: Why Good AI Writing Still Needs a Human Editor

AI has made it much easier to produce decent writing. Give it a reasonable brief and, within seconds, you can have 1,500 words that are grammatically correct, neatly structured and, at first glance, perfectly publishable. A few years ago, getting to that point might have taken a writer several hours. Now it can take a few minutes.

I don't think there's much value in pretending otherwise. I use AI myself for research, organising information, testing ideas and getting through some of the more repetitive parts of the writing process.

But there's something I've noticed the more I've worked with AI-generated content. Getting to 90% has become much easier. It's the last 10% that causes the problems.

That doesn't mean the human part of the process only begins once AI has produced a draft. Research, briefing, source selection and verification can all require human judgement long before the editing starts. The 90% is really about something else: AI can get a piece of writing surprisingly close to looking finished, and the remaining gap is often where some of the most important decisions are made.

Polished writing isn't necessarily finished writing

Bad AI writing is easy to spot. It repeats itself. It uses strange phrases. It makes sweeping statements. Every section has three bullet points. The introduction spends several paragraphs telling you how important the subject you're already reading about is. That stuff is relatively easy to fix.

The more interesting problem is AI writing that is actually quite good. The grammar is fine, the structure makes sense and nothing immediately jumps out as wrong. Read it quickly and you might think there isn't much left to do. That's where I think people get caught out.

Polished writing and finished writing aren't necessarily the same thing. A paragraph can be perfectly well written and contribute almost nothing to the article. A sentence can sound authoritative while slightly misrepresenting the source behind it. An introduction can be engaging without answering the question the reader came for. A piece can contain all the right keywords, headings and information while still feeling like something you've read twenty times before.

These aren't always obvious problems. They're judgement problems.

AI is very good at producing something plausible

One of AI's biggest strengths is also one of the reasons its writing needs proper review. It is very good at producing something that looks right.

If there's a gap between the information available and the information needed to complete a sentence, AI doesn't necessarily stop and tell you there's a problem. Sometimes it bridges the gap, and the result can sound completely reasonable.

That's relatively harmless if you're writing about the best way to organise your desk. It's a different matter when you're writing about immigration rules, tax, corporate structures, investment programmes or anything else where somebody may actually rely on what you've written.

I've spent a large part of my career writing about subjects like these. The difficult part isn't making the sentence sound professional. It's knowing when a sentence needs checking.

I've seen this repeatedly when working with immigration, tax and corporate content. An AI-generated explanation can read perfectly well while missing a qualification, using an outdated requirement or presenting a general rule as though it applies in every case. Unless you know enough about the subject to question it, there may be nothing in the writing itself that tells you something is wrong.

Sometimes the problem is a date. Sometimes it's a qualification that's been omitted, or two sources that describe the same rule slightly differently. And sometimes a technically accurate statement becomes misleading because an important condition has disappeared during simplification.

AI can help enormously with the research, but it doesn't remove the need to understand what you're looking at.

Editing AI-generated content isn't just proofreading

This is where I think the idea of "humanising AI content" can undersell the job. If all we're doing is swapping a few robotic phrases for more conversational ones, we're not adding much. Editing AI-generated content should be much more demanding than that.

I find myself asking questions like:

Is this actually saying anything?

A paragraph can be clear, accurate and completely unnecessary.

Would the client really say this?

Something can sound natural in isolation while being completely wrong for the person or business whose name is attached to it.

Does this claim need checking?

Not because it looks obviously false, but because it looks just plausible enough that nobody might think to question it.

Are we answering the reader's question?

Or have we spent 300 words providing context because that's what articles are apparently supposed to do?

Is this specific enough to be useful?

"Businesses should develop a comprehensive strategy tailored to their unique circumstances" sounds perfectly professional. It also tells you almost nothing.

Have we repeated ourselves?

AI is particularly good at expressing the same idea several different ways without making the repetition immediately obvious.

And perhaps the most important one:

Would anybody notice if this paragraph disappeared?

If the answer is no, I usually delete it.

That isn't really humanisation. It's editing.

Sometimes the best edit is deleting something good

This was true long before generative AI existed. One of the harder things about professional writing is learning that a sentence doesn't have to be bad to deserve deleting. You can like it. It can be well written and contain useful information. And the piece can still be better without it.

AI makes this harder because it can produce an enormous amount of perfectly competent material very quickly. When you've generated 2,000 words in thirty seconds, there's a natural temptation to treat those 2,000 words as the thing you're now responsible for polishing. I think that's backwards.

The draft is raw material. Maybe 1,600 words deserve to survive, or maybe only 900 do. The most useful sentence in the entire draft might be buried halfway through section four and actually belong in the introduction.

Editing isn't about protecting what already exists. It's about working out what the finished piece needs.

AI hasn't made research less important

If anything, I think it has made research more important. AI can speed up the process of finding information, comparing sources and understanding unfamiliar subjects. That's valuable.

But speed creates its own temptation. When research took longer, you were forced to spend time with the material. Now it's possible to go from question to summary to finished-looking article very quickly.

The danger is mistaking access to information for understanding it. For straightforward subjects, that may not matter much. For complicated ones, it matters enormously.

If I'm writing about a visa programme, for example, I don't just want a list of eligibility requirements. I want to know where those requirements came from and whether they're current. I want to see whether the official source says the same thing as secondary sources, whether an apparent rule is actually a general rule with exceptions and whether something that was true six months ago is still true today. And I want to know whether the information answers the question the reader is likely to have.

AI can help me get there faster. It can't make those decisions for me.

Voice is harder than removing AI phrases

There's another part of editing AI-generated content that has very little to do with factual accuracy: voice.

You can ask AI to be conversational, tell it to avoid corporate language, give it examples and tell it not to use certain words or sentence structures. All of that helps.

But someone's voice isn't just a collection of stylistic rules. It's what they choose to talk about, what they find funny, how quickly they get to the point and which details they think are important. It's also what they refuse to say because it sounds ridiculous, where they're comfortable being uncertain, where they have a strong opinion and the things they notice because they've spent years doing the work.

That's why editing somebody's writing is often easier after you've worked with them for a while. You start recognising the sentences they'd never say.

AI can imitate patterns. Understanding the person behind them is harder.

This doesn't make AI the enemy

I don't subscribe to the idea that using AI somehow makes writing less legitimate. Writers have always used tools. Search engines changed research, spellcheck changed proofreading, word processors changed editing and SEO tools changed how we understand search behaviour. AI is another substantial change, and pretending it isn't useful doesn't achieve much.

The question is what we hand over to it.

I am perfectly happy letting AI help me organise a large amount of information. I'm happy using it to test whether I've missed an obvious counterargument, and asking it to help compare material or identify repetition.

What I'm less interested in outsourcing is the decision about what the finished piece should actually say, because that's the part the client is paying for.

The 90% test

So this is the test I've increasingly started applying to AI-assisted writing. Imagine the draft is already 90% there. It's clean, readable, structured and optimised. Nothing is obviously broken.

Now ignore all of that and ask:

What would I change if my name were going on it?

That's where the real editing starts. A claim might need another source. Three paragraphs might become one, or the entire introduction might go. The article might need an example from somebody who has actually done the thing being discussed, or the conclusion might just be repeating the introduction and can disappear completely. Sometimes the piece is technically excellent but still doesn't contain a single idea worth remembering.

And sometimes, after going through all of that, the draft really is good. Great.

The point isn't that AI writing must be bad. The point is that "AI wrote it well" isn't a sufficient reason to publish it.

The easier writing becomes, the more judgement matters

AI has lowered the cost of producing competent words. I think that's broadly a good thing, but it also means competent words are becoming less scarce. Anyone can generate a polished article, produce ten LinkedIn posts or create a professional-looking service page.

The interesting question is increasingly what happens after that. Who checks the claims, removes the paragraphs nobody needs or spots that the article is answering the wrong question? Who knows enough about the subject to recognise when something is slightly off? And who is willing to look at 1,500 perfectly acceptable words and decide that 500 of them shouldn't be there?

That's the last 10%. It might also be the part that matters most.