I spent over 15 years being paid, by agencies including TBWA, Ogilvy, Saatchi & Saatchi and VCCP, to write words for Google, the Financial Times, IKEA, Milka and many more brands I can shake a stick at.

None of these esteemed agencies ever paid me to draft ideas and then present them as the finished article. No. They paid me to take a first draft, then hone, nurture and craft it into a polished idea that would be something to be proud about.

The difference between those two things (draft and craft), is exactly what a lot of teams using AI haven’t worked out yet, which is why so few of them have a real set of editorial guidelines for AI-generated content.

A few weeks ago, I was speaking to a senior director at a financial services firm in South-East Asia. We were discussing a study he’d read, which concluded that “using AI makes you dumber.” I pushed back, countering that the report was too sweeping and generalised too heavily.

I gave him a personal example of how AI can make people smarter: I’ve been using a mix of LLMs (Gemini and Claude) to learn Bulgarian and it’s been genuinely brilliant, a proper game changer.

My argument was that it really depends what you’re using it for, in what capacity and with which tool (or tools) you’re doing the work.

He wasn’t wrong though, not entirely. Buried in the study he’d sent was a detail I couldn’t argue my way around: doctors who’d become accustomed to using an AI tool to spot things during colonoscopies got measurably worse at spotting those same things themselves, when the AI was switched off.

Not because they’d forgotten how, but because they’d gradually stopped checking as hard and became reliant and overly-trusting of the tool, like going on auto-pilot.

That’s the exact risk sitting inside teams right now, and it has nothing to do with whether AI can write a decent sentence (it can). The real question is what happens to your own editorial eye after it’s spent a few months watching a machine write a solid first draft.

When Doctors Stopped Checking, Their Skills Stopped Working

This study is real and the details are worth learning about.

It was first published in the Lancet Gastroenterology & Hepatology in August 2025 (Budzyń et al., DOI: 10.1016/S2468-1253(25)00133-5) and followed 19 highly-experienced endoscopists, each with over 2,000 colonoscopies behind them, working across 4 centres in Poland.

Before they started using an AI detection tool, they found at least one adenoma (a polyp that can turn cancerous) in 28.4% of colonoscopies. After they’d been using the AI tool for a while, the researchers looked specifically at the colonoscopies those same doctors did without it, and the detection rate had dropped to 22.4%. In other words, almost one-fifth of their own detection ability had disappeared, the moment the AI safety net wasn’t there.

The two researchers behind the study, Dr Marcin Romańczyk and Professor Yuichi Mori, called this the Google Maps effect: the same reason your own sense of direction gets worse the longer you follow a screen telling you where to turn. It’s not that the doctors forgot how to spot a polyp. It’s that, over time, they’d stopped needing to look as hard, because something else (the AI) was looking for them.

The AI didn’t make these doctors worse at their job, but it did make them stop checking as carefully, which resulted in them getting worse at their job on their own.

Thankfully, this isn’t inevitable, nor is it really about the AI at all. It’s about whether anyone in a team has actually decided where and when holding work up to an agreed-upon set of standards stops being optional.

Why Every Team Needs Its Own Editorial Guidelines For AI-Generated Content

I hear a version of the same problem from almost every team I talk to, which is what inspired this week’s article.

One half of the team won’t touch AI, on principle or because they just don’t like it, they still do everything the same way as before, while watching colleagues move faster. The other half leans on it for more than drafting: research, structure, headline options, and yes, sometimes the actual thinking.

But because nobody has sat the whole team down and said here’s where AI drafts and here’s where you craft, what eventually happens is that two very different standards start operating in the same team, often on the same project, and nobody highlights the problem out loud.

If I’m honest about the real cause, it isn’t fear on one side and laziness on the other, it’s a mix of both, and they tend to feed each other.

The sceptic feels like they’re the last line of defence for quality and humanity, doing everything by hand and resenting the colleagues who don’t. The over-user keeps delegating “just a little more”, not only the typing but also the thinking behind it, because they’ve never been told where the line actually sits.

Neither of these sides are wrong to feel the way they feel, they’re both filling a gap that should have been filled by a standard, with a personal instinct instead.

A team split between AI-sceptics and AI-over-users is a personality problem that needs fixing.

The Argument I Don’t Think Either Side Has Quite Right

The received wisdom tends to land in one of two places:

1. Ban it outright, to protect the craft
2. Or let it draft everything and treat editing as a quick proofread

I think both instincts are wrong, for the same reason.

AI is genuinely brilliant at solving what I’d call the fear of the blank page: those horrible first 10 minutes where nothing exists yet, where you’d do almost anything to avoid starting. This is precisely where AI earns its keep.

But what it should never be allowed to replace is the thinking that happens afterwards.

This kind of thinking is where the truly great ideas turn up, because you don’t know what you actually believe about an idea until you’ve really explored all aspects of it and measured it against your own perceptions, morals and viewpoints. Nor until you’ve read all the footnotes and citations and sat with it for long periods of time, letting your brain truly digest every part of it. This is how you build a 3D map of the idea, allowing you to navigate it fully and find new routes to places you hadn’t yet explored.

In other words: new ideas are created by exploring and understanding other ideas. Not by letting someone (or, rather, something) do all the thinking for you.

On 11 September 2026, Terence Tao, arguably the most respected mathematician alive, published a declaration titled “A Severe Misalignment of AI in Mathematics”, signed alongside 25 other Fields medallists (mathematics’ answer to a Nobel Prize), about AI’s rush to solve famous unsolved problems.

Their argument wasn’t that AI got the answers wrong, it was that solving a problem was never really the point of maths in the first place.

The point, they say, was the understanding that comes from a human wrestling with the question, discussing it, refining it, passing it from one mathematician to the next. Hand someone the answer before they’ve made that journey, and you haven’t helped them, instead, you’ve taken something away from them.

Writing has exactly the same problem, dressed differently. The value in a piece of writing is never only in the finished sentence, the majority of it is always in the thinking that produced it.

An editor who lets AI draft the whole piece hasn’t just outsourced some typing, they’ve forfeited the part of their job where they notice the story isn’t quite right, the argument doesn’t actually hold up or there’s a sharper angle in paragraph 3, which deserves to be explored further. That craft is the job. It’s also, not coincidentally, the part that makes a piece worth a reader’s time, rather than a search engine’s crawl.

AI is amazing at elevating a team’s work, but it is not a replacement for it completely.

The Same Line, In A Completely Different Industry

I was recently advising a Global Product and Supply Chain Director at a major manufacturing firm, on using AI to forecast stock levels across several regions.

Done well, it’s a genuine business benefit: fewer regions sitting on overstock, fewer regions caught out by a shortage and a lot less of his team’s time lost to short-notice fire-fighting.

AI drafts that forecast brilliantly. It can work through years of sales history and regional patterns faster than any team could manage by hand. But the decision about what to actually do with that forecast (which region gets priority, which factory adjusts first, how much risk the business is willing to sit with this quarter) was never something to hand to the model. This is the judgement call that needs to be made by people who understand the business, and it’s usually best to make it altogether in a room where you can talk about it in great detail with each other.

The AI drafted the forecast…but the team still had to craft the decision.

This is how it should be, whatever the industry: AI used to assist the humans making the final decisions.

Manufacturing, mathematics, medicine, an editorial desk, the approach should be the same every time: AI drafts, humans craft.

What A Real Editorial Standard Actually Needs

None of this is an argument for banning AI from your editorial process. It’s an argument for writing down, properly, where your team’s line actually sits, so nobody’s left guessing. A working standard needs to answer 5 questions, not vaguely, but in language specific enough that two different people would draw the same conclusion from it:

1. Where AI is allowed to draft: name the actual stages, research summaries, first structures, headline options and be specific enough that nobody has to guess where the boundaries are
2. Where it isn’t: the final line edit, any judgement call, anything carrying your publication’s own opinion, these should all stay with a real person, every time, no exceptions
3. Who signs off (and what they’re actually checking for): not just spelling and grammar, but voice, accuracy and whether the piece still sounds like your brand’s voice, rather than someone else’s
4. How you handle disclosure, and to whom: your internal standards and your reader-facing standards require two very different sets of decisions, and mixing them up is how teams end up arguing about the wrong questions entirely. Set them up clearly
5. What happens when two people disagree: a split team needs one named person who makes the final call when the sceptic and the over-user can’t agree

I’ve built a working template that covers all 5, ready for you to adapt this afternoon rather than starting from a blank page. You can download it here.

You Don’t Have To Choose Between “Made Quickly” Or “Done Properly”

The senior director who sent me the “AI makes you dumber” study is still one of the smartest people I know, and he certainly wasn’t wrong about how unmonitored AI use erodes skill. Where I think his belief was misplaced was “stop using AI completely”. That will never be a permanent solution now. The best approach, in my experience, is knowing exactly where you stop letting AI draft and start doing the job you’re actually paid for.

Do that properly and you get to keep both things: the speed AI genuinely offers on the first (aka: worst) 10 minutes of a piece and the editorial eye that took your team years to build. The work gets better, stays human and is unmistakably yours, not the average of everyone else’s.

If you want the practical detail behind everything in this piece, I’ve written the specifics up as 5 short answers below: what to actually put in your standard, how to edit AI copy properly, how to spot it when you’re reading it, whether to disclose it to your readers and what AI still can’t do no matter how good it gets.

Related questions:
– What should an editorial standard for AI-assisted writing contain?
– How do you edit AI-written copy properly?
– How can you tell if copy was written by AI?
– Should writers disclose that they used AI?
– What is AI genuinely bad at in editorial work?

If you’re watching all of this happen and wondering how to make AI work for your business, book a call with me today – I train and consult businesses and teams to use AI with confidence, skill and intention.