We all marvelled at the limitless possibilities of ChatGPT when it first appeared. “You mean it can write sentences that sound pretty much like a human wrote them?”
Fast forward three years and the humble copywriting profession was as dead as the dodo. We all rebranded ourselves as “digital marketers” and scrambled for new sources of income.
After all, AI could write the generic SEO blogs that many companies relied on to stay at the top of a search page ranking. Website copy, social media, emails. All the formulaic chaff of the marketing mill was handed over en masse to our new AI friends. Hey presto.
The writing plateau
Something unexpected has happened in 2026, however. Despite a continuous rollout of ever smarter models, writing doesn’t seem to be improving in the same way.
In fact, it seems to be getting worse.
This is not to discount the fact that AI models have continued to advance in many capabilities. Some frontier models have become so terrifyingly adept at cyberattacks that even the “let it rip” ethos of the Trump administration is being reversed in favour of regulation and restraint.
I have close friends who are software engineers. The most sceptical of these has told me they’re genuinely awestruck by the coding and problem-solving abilities of Claude Fable. They’ve become true believers in an eventual AI takeover of the software world.
When it comes to writing, we seem to be at a plateau. After ingesting every written word in human history and parroting it back to us, AI models are beginning to struggle with originality.
The em dash was an early bugbear for many grammar nazis, but these were excised by some tweaking or massaging of the models (no one 100% knows how an AI model is trained, as they’re more like a brain than a piece of software).
Yet a number of AI-isms still persist:
- Constant bullet points (often with emojis in social posts).
- Title Case Headings And Subheadings (An American convention).
- Correlative conjunctions (“not only x, but y”).
- Auto-generated questions at the end of every post.
The list goes on, but for human writers and editors, we spend an inordinate amount of time unpicking these telltale signs. You can set rules to prevent the AI from falling back on these habits, but the results still seem to be lacking a certain sparkle.
It just never seems to nail the tone I need. I struggle to parse the nonsense it concocts. It gets tied up with one idea and won’t let it go despite how many times I call it an idiot.
In the pursuit of writing a genuinely good article, I found myself wasting my day by “teaching” the AI how to write when I could have just written the damn thing myself.
When we understand how these models are trained (or grown), it’s not a surprise.
Billions of people are using them to generate endless streams of generic crap, and only a tiny percentage of us are giving the models any corrective feedback. The end result is a feedback loop of AI slop training on more AI slop.
Now every piece of writing looks identical. Every tagline is unimaginative and uninspiring. Humour and irony are nearly non-existent. There are no real stories, just facts.
And none of us want to read it anymore.
The worm has turned
A couple of weeks ago, LinkedIn added a new option to the three-dot menu on every post. It reads “Seems like AI slop”.
Tap it and the post disappears from your feed, and your judgement gets fed into LinkedIn’s ranking models. Chief product officer Hari Srinivasan has said slop is a top priority for the company, and that human feedback is the most reliable training signal for removing it.
In the same announcement, LinkedIn retired its own “enhance your post” AI writing tool. A social media platform that spent the last two years promoting a button to generate AI content has now taken that button away and given readers a button to punish it.
A week before that, Substack had taken a similar action. It now runs posts, notes, replies, and comments over 100 words through Pangram, a detection model, and shows readers an estimate of how much was machine-written.
Chief executive Chris Best was blunt about their motivation. He said Substack “doesn’t want to turn into LinkedIn”, and that platforms rewarding fakeness start a race to the bottom. Substack also added a “How I make this” disclosure, where writers explain their process in their own words.
Are these measures going to have the intended result? As with anything in the world of media and technology, probably not.
But they do illustrate the changing public sentiment towards AI-generated content. Neither of these global platforms would be making significant changes unless they’d been hammered by feedback from their users.
Spotify, YouTube, Meta, and TikTok have all launched new measures for labelling AI content, some more effectively than others. Again, none of these measures happen because of Daniel Ek or Mark Zuckerberg’s altruism. They’re trying to get out ahead of the AI slop debate before their users vote with their subscriptions.
What this means for your brand
Content marketing is a slog. It seems like minimal rewards for constant effort. You post a blog or a carousel, and it gets two likes from your own employees.
So why bother writing it yourself or paying someone to? I totally get it.
The problem is, you’re now better off posting nothing at all.
We all know AI slop when we see it, so your “Five Reasons to Secure Your Data Before Rolling Out AI” blog is not fooling anyone. Ditto for the AI-generated image you’ve slapped on the top of the blog. Combine an AI-generated blog and image with an AI-generated caption for posting it on LinkedIn, and you have the perfect storm of slop.
The negative sentiment towards this stuff has moved faster than most marketing teams have realised.
- Klaviyo surveyed 8,000 consumers across eight markets, including Australia. When people notice AI-generated content in brand marketing, a third of them say they trust the brand less.
- Gartner found that half of consumers would rather give their business to brands that keep generative AI out of customer-facing messages entirely.
- DoubleVerify surveyed 22,000 consumers across 22 markets and found 42 per cent would think worse of a brand that ran low-quality or uncanny AI creative.
For B2B, the credibility cost is heavier than it is in consumer marketing. 64% of B2B buyers say they trust thought leadership more than product sheets when assessing a vendor’s capability. If your thought leadership reads like a machine wrote it, you’ve lost one of your most effective channels.
Finding a balance
I use AI every day. Claude helped me to research this article. It helps me stay on top of my workload by managing mundane tasks like surfacing information buried in emails and transcribing meeting recordings. It saves me hundreds of hours of drudgery.
I’m not suggesting we do away with AI-generated writing altogether. It has its place. If you want a generic explanation of your technology solutions, by all means, use AI. It’ll do a great job.
But if you want your customers to take your brand seriously, it’s time to reconsider the amount of AI-generated thought leadership you post.
AI transcription tools have never been better. Start by recording a conversation with your internal experts. Get their unvarnished ideas and unique tone of voice on a page verbatim. Edit out the ums and ahs, and you’ve got a real thought leadership article.
Better yet, get them on video. Cut it into a 60 second clip and you can prove there’s real human brains behind your business.
For the love of god, though, please stop posting 100% AI-generated blogs and social posts. If the LinkedIn slop police don’t stop you, your lack of credibility will.
This article first appeared on LinkedIn on 12 August 2026.
