What Is AI Genuinely Bad At In Editorial Work?

AI is genuinely bad at original judgement: knowing what’s actually true rather than merely plausible, holding a point of view it would defend under pressure, and noticing when a story or argument is wrong in a way no source states outright. It produces fluent text, not understanding, and editorial work depends far more on the second than the first.

It Can’t Tell True From Plausible

AI generates the sentence that’s statistically likely to follow, not the sentence that’s been checked against reality. That makes it fluent and occasionally wrong in ways that read exactly as confidently as the parts that are right. An editor’s job includes catching the difference, and no amount of AI improvement removes the need for a person to do that checking, because the tool that wrote the sentence has no independent way of knowing which category it falls into either.

It Doesn’t Have A Point Of View It Would Defend

Good editorial writing usually argues something a reasonable person could disagree with. AI-generated drafts tend to describe a debate rather than take a side in it, because there’s no professional reputation or personal conviction sitting behind the words. A human editor’s opinion, formed through actually doing the work over years, is exactly the ingredient that can’t be generated, because it isn’t a pattern in existing text. It’s a judgement someone is willing to be wrong about in public.

It Can’t Do The Thinking That Produces The Insight

On 11 September 2026, mathematician Terence Tao published a declaration titled “A Severe Misalignment of AI in Mathematics”, co-signed by 25 other Fields medallists (mathematics’ equivalent of a Nobel Prize), warning that AI companies racing to solve famous unsolved problems were missing the actual point of the discipline. Solving the problem was never really the goal, they argued. The goal was always the understanding a mathematician builds by wrestling with a question themselves, refining it and passing that understanding on to others. A correct answer handed over before that process happens isn’t a gift, it’s a shortcut past the part that mattered.

Editorial work has the same structure. The value in a piece often isn’t only the finished argument, it’s the noticing that happened while writing it, the moment a writer realises the obvious angle is wrong and finds a better one three drafts in. AI can produce a fluent version of the obvious angle very quickly. It can’t do the part where someone sits with a bad draft long enough to realise it’s bad, and finds something true instead.

It Doesn’t Know What Your Readers Would Notice

An experienced editor knows things about their own readership that were never written down anywhere for an AI model to learn: which claims this particular audience will instinctively distrust, which tone lands as confident and which lands as arrogant for this specific publication, which topic was covered badly by a competitor last month and needs a different angle to avoid looking like an echo. That knowledge lives in a person’s accumulated experience of the actual readership, not in any dataset, and it’s the part of editorial judgement that ages a piece of AI-drafted copy the fastest when it’s missing.

What This Means Practically

Use AI for what it’s genuinely good at: getting past the blank page, summarising research, generating options to react to. Keep the judgement calls, the fact-checking, the actual argument and the point of view with a person, every time, because those are the parts AI is structurally unable to do, not just currently bad at.

This Gets Worse, Not Better, With A Better Model

It’s tempting to assume these gaps close as AI models improve. They don’t, because they’re not capability gaps in the way a spelling mistake is. A more advanced model produces more convincing fluency, which makes the absence of genuine judgement, a real point of view and independently verified truth harder to spot, not easier. That’s precisely the risk Tao and his co-signatories were warning about in mathematics: a system that gets better at producing convincing answers doesn’t get better at replacing the understanding a person builds by working through the problem themselves. The editorial equivalent is a more fluent AI draft that’s easier to mistake for finished work, which makes a human editor’s judgement more necessary as the tools improve, not less.

This is one of five questions answered in the full guide, AI Should Draft, Humans Should Craft: Where The Line Sits In An Editorial Team.

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.

Reference: Tao, T. and 25 co-signatories (2026). A Severe Misalignment of AI in Mathematics. Published 11 September 2026: terrytao.wordpress.com