Warning, this article contains human-written em dashes.
I know this isn't going to land well, but you're going to have to bear with me.
AI is very effective at writing.
It's adept at translating between languages, skilled at comprehending and synthesising information, can write functional code quicker than almost any human programmer, and can construct thorough, rational chains of thought to reason through a problem.
What ties these examples together is that the effectiveness of the output matters, but the brilliance of it does not.
Good code does not need to have soul—it needs to be functional, adhere to conventions, and be easy to follow. Similarly, it doesn't matter if the documentation accompanying that code sounds like AI wrote it. Its purpose is to convey information clearly and concisely so that a person can quickly get to grips with how it works.
- A data report needs to convey information clearly and accurately—not eloquently
- Internal company documents need to be concise and informative—not exude personality
- A meeting summary should be accurate and complete—not poetic
Can we set aside all of the AI-hate for a second and just appreciate how incredible it is at this sort of thing?
The good and the bad
There are a lot of tasks out there for which AI-generated writing is perfectly acceptable. No, you know what, I'll go one step further: there are some tasks for which using AI is the most sensible and effective option.
However, there are also plenty of situations where AI should not be used. It cannot bring the same level of taste, judgement, experience and personality that a human can.
It should not be used for anything where the subjective quality of the output matters.
Trust me, I hate seeing AI-generated articles and social media posts as much as the next person. That is absolutely, unequivocally not a good use of AI. But bad uses of AI aren't limited to clickbait stories on Medium.
Don't worry, this isn't going to be yet another anti-AI-generated content story—they're almost getting as tiring as the AI-authored posts themselves.
AI is reshaping brand credibility, too—and not for the better. A 2024 study by the Nuremberg Institute for Market Decisions found that consumers become measurably more sceptical and less engaged the moment they discover content is AI-generated. A recent study by Skyword found that 30% of consumers state they are less likely to engage with or buy from a company if they suspect its content is AI-generated. Customers are sceptical of AI so much so that it can actively harm your brand identity. For public-facing content especially, the efficiency gains are almost certainly outweighed by the reputational cost.
While AI has both some strong and potentially detrimental use cases, knowing when AI should and shouldn't be used is not always so clear-cut.
The grey area
Public-facing AI usage may harm brand identity, but what if you label it as AI to distance yourself from it?
Say you published frequent data reporting and wanted to add an AI-generated summary, labelled as AI.
- The quality of the main report is unchanged.
- The headline summary actually provides additional value for users seeking a brief synopsis instead of digesting the entire report.
- AI can write a very succinct, accurate and comprehensive summary far quicker than a human can, making it a more economical option for the company.
- A reader may have asked AI to summarise it anyway, so by providing your own summary, you can better ensure its suitability.
- The only downside is that it sounds like AI wrote it.
Does that make it a good use of AI? Or perhaps, even if there is real value being supplied, the fact that AI was used diminishes the quality of the product? Maybe, as much as we like to criticise AI use, it does prove itself to be a valuable tool.
There are plenty of situations where using AI to write content for you seems to work. Take LinkedIn, for example. If you've been on there in the past year, you'll immediately know what I'm talking about. On platforms like this, where AI-generated posts are so common, it almost takes me by surprise to see something written by a human. It's what the audience has come to expect. There are plenty of people who have viral posts that have clearly been made with AI.
Maybe the problem is louder than it is widespread, but it seems as though we all reject AI-generated content, yet the posts still go viral.
What about text humanising?
Text humanising and voice replication aren't foolproof approaches. It's great in some scenarios, but it wouldn't work for an article like this one, for example—you'd see straight through it.
Again, before anyone comes for me in the comments, I typed that em-dash myself.
I publish all of my articles on both Medium and my personal blog. On my blog, each post needs some additional metadata to help with Google search results. I use AI to write a title and description for each post. This is a very AI-teachable skill; there are guidelines and rules you can follow to write a good meta-title and meta-description. It's a task that I don't enjoy doing, so I outsource it to AI. However, it is public-facing—this is the text that's going to appear in a Google search result. My personal workaround for this was to train the AI agent that performs this task on my personal brand voice so that the text it generates still sounds like me.
For something like a meta title and description where it's a fairly short and formulaic format, I don't really notice the difference. A human-written meta description and an AI-generated one (especially using my brand voice) are pretty much indistinguishable since there's so little room to play with.
You could apply this same brand voice mimicking technique to something like data reporting—and I have. Public-facing data reporting uses specific language, has a somewhat repetitive structure, and a distinguishable style. I've found that the more something exhibits a deliberately impersonal style, the easier it is to replicate with AI and the harder it is to spot. The same is true with some forms of journalism and news reporting. If you are writing a news story to be unbiased and deliberately hit both sides of the argument without inserting your own opinions, that's a voice that you can replicate fairly well with AI.
Where it gets a bit trickier is when you're trying to replicate personality through humanising AI-generated text or mimicking a brand voice. This isn't something I've found to be possible. Experienced AI users are adept at detecting AI-generated content, even when 'humanised'. So even if you did manage to fool most of the audience, it only takes one sceptic to cry "AI" for everyone else to get involved, too.
To really stop people from being able to tell, you'd have to completely rewrite the whole piece to remove the deeply ingrained AI-writing patterns. At that point, you may as well have just written it yourself in the first place.
Human in the loop
Even in all of these AI-favourable use cases, human judgement is still not redundant.
For data reporting and journalism, we can implement error and accuracy checking routines into the content workflows to spot errors and hallucinations. While you can get pretty close to perfect, you still can't guarantee 100% accuracy. There are some scenarios where this is okay, and naturally, some where you do need 100% certainty. In this case, you would still need a human (or some more deterministic tool) to ensure that what you're saying is correct
Human-in-the-loop is the notion of keeping a human involved in the decision-making process. It could be approving every step of the way, having a high-level oversight at key stages, or just a final review before publishing. The requirements of the project will determine the appropriate choice.
As someone who builds AI-content workflows, I personally wouldn't completely outsource quality control to AI. I don't yet have the confidence in the tools to do exactly what they have been told to do, adhering to the guidelines supplied and with complete accuracy. It can get you 80%, maybe 90% of the way there, but at least that final 10% should rest with you. The only way to build a reliable workflow that will actually work for you in practice is to accept that a human will still be involved in the process—at least a little. Once you've accepted that fact, you can start designing it in and start seeing the results.
I help small businesses and creators build content workflows that actually hold up in practice, where human-in-the-loop mechanisms are designed in rather than bolted on. I'm currently taking on a select number of clients. If that sounds useful, get in touch.
If this is you, I understand, but also: please stop.
I opened this article with the idea that AI is an effective writer but not a beautiful one. I don't believe this idea is unique to me; far from it, in fact. I think it's almost a universal conclusion for people learning AI to reach.
For all of us, it's a journey that begins with initial curiosity that blossoms into fascination. You start to experiment with AI and learn more about how to use it effectively. Somewhere along the way, you get a bit carried away and start applying AI to everything.
When you're at this point, its potential seems unlimited—there is no task it cannot accomplish, and nothing it can't do quicker and more effectively than a human can. But it doesn't last forever. Eventually, the facade starts to crack, and you realise it isn't quite as clever as it seems.
The more you write with it, the more you notice the sing-song rhythm of its paragraphs. The more you chat with it, the more you realise how sycophantic it can be and how much of a tendency it has to become so focused on something while ignoring the bigger picture. The more you try to treat it like a person, the more you start to realise how much humanity it lacks. It's absolutely incredible, yes, but far from human—at least for now.
Before there was so much hate for AI-generated content online (and before I had reached this stage of my own AI progression), I'll admit, I used AI to write content for the internet. I hold my hands up; I was part of the problem. But I almost think it's a necessary step everyone needs to take to get to the other side. Not publishing AI-generated content online necessarily, but pushing the limits of what it should be used for.
You need to be inspired by it and to have that awe in what it's capable of. Put it on a pedestal, at least for a little while. You'll realise eventually that it doesn't belong there, but you'll have learnt more about its capabilities and limits than you would have otherwise.
You have to live in the illusion for a little while before you can see through it.
Today, I still use AI to help with my writing, but now it's more the way I'd collaborate with an editor. It helps me review my writing and only suggests high-level changes to stop me going off on a tangent and ensure everything flows properly. That's the balance I've ended up with as AI has slowly settled into my workflow over the past few years.
So, as I say, I completely understand the urge to share the amazing things that AI can create for you, but please, be better than me. AI has its place in the world and on the internet, but it's not in posts like this. Save it for what it does best.