LLMs generate low quality content

AI generated content is not just incorrect a lot of the time, it is also of inferior quality to human created content and it is not going to get better. The nature of Large Language Models requires that they are retrained on new data very often. But there is not enough new data to train them on, and so their quality gets worse.

How bad is the quality of generated content?

So bad that the tech companies behind the LLMs do not use their own products for marketing purposes or otherwise. At the 2026 World Economic Forum, Anthropic CEO, the company behind Claude, Dario Amodei said that AI could replace software engineers within 6 to 12 months. Meanwhile, the company is still looking for several software engineers to fill positions within its own company.

And when OpenAI wanted to create a commercial for the Super Bowl in 2025, they too hired human creatives to do the work. And in 2026 they did the same thing again. They “shot the ads with real people, on film” according to Kate Rouch, chief marketing officer of Open AI.

When other big brands did use generative AI to create their ads, there was a lot of backlash. Coca Cola’s 2025 holiday season ad campaign was roasted online, it still required lots of human help and wasted a lot of resources. McDonalds Netherlands also gave AI generated ads a try for the 2025 holiday season, but it was so bad that it was quickly pulled.

Why is AI generated content so inferior?

Large Languages Models (LLMs) are basically really advanced predictive text or image models; like the one on your phone that offers you three options every time you type a word only better trained. Because they are just predictive models, they give incorrect information a lot of the time. It is why when asked in January 2026 if 2027 is the next year, Google’s AI overview said no; and why ChatGPT still says that there are only 2 r’s in the word strawberry. LLMs cannot count, they can only predict the next word you want to see.

50% of internet content is AI generated

Where the predictive text on your phone is only trained on your text input, LLMs are trained on everything on the internet. This is a problem on multiple levels, the first problem was copyright and now the tech companies behind these LLMs are facing lawsuits all over the world. Now the problem is that the LLMs need constant retraining on new content but 50% of internet content is now AI generated.

This means that LLMs are training on their own content instead of original work, which leads to further degradation of AI generated content over time. This is called model collapse or AI collapse.

What is model collapse and what are its consequences?

Model collapse is not a new notion. In July 2024, researchers published a study that showed how AI models collapse, and gave the following definition for the term: “a degenerative process whereby, over time, models forget the true underlying data distribution”. What does that mean?

In the case of LLMs, this means that it will forget the patterns of the original content it was trained on, the human-made content. For a really advanced pattern recognition and predictive text model this is a problem. Because how can it continue to generate human sounding text if it no longer has human-made patterns to emulate?

AI generated content will become more and more homogenised

In October 2025, another study looked into the consequences of model collapse. The collapse was characterised in two ways, first a single model consistently generated similar responses, and secondly different LLMs generated very similar output to each other. With different LLMs generating very similar content, we have to consider the consequences of people constantly being exposed to similar content instead of original ideas.

What is the solution to model collapse?

LLMs will have to be trained on more specific data, instead of everything that exists on the internet. Or platforms will have to find ways to reduce the amount of AI slop that has infested them. Maybe both.

If LLMs are to be trained on specific data to avoid content they generated themselves, the tech companies will have to find a way to choose which content to train on. If they are able to create the technology to recognise their own work, than you will want to make sure you do not have any AI generated content on your website. The other option the tech companies have is through official licensing deals, which means paying human creators for their work.

For now, Youtube has already started removing AI slop channels. They are using AI to remove gen-AI, which appears to be a hit or miss approach. While they have taken down several big AI slop channels, they have also mistakenly demonetised human creators.

Why you should care about AI slop and model collapse

Why does it matter that AI generated content is bad? Because people do not only consume this content for entertainment. More and more people use social media, where AI generated content abounds, to learn or stay on top of news developments. If the content they consume in those pursuits is fake or manipulative, it can have disastrous effects on society.

You and your business deserve good quality content. Yes, hiring a copywriter might be more expensive in the short term and take more time than using generative AI but good things come to those who wait. Gen AI is quick fix, not a permanent solution, it is not a sustainable solution to the need for fresh content that will help you rank better. Human copywriters and creatives are.

LLMs are not actually intelligent, artificial or otherwise. They are simply trained on patterns and can reproduce those patterns. They cannot create something new, they can only regenerate what they have seen before. And you and your business deserve original content, original thought and authentic creations.

Why I care about AI slop and model collapse as a copywriter

As a copywriter, I care about the quality of content people consume, whether that is online or offline. If low quality content becomes the norm, it will affect society in ways we cannot quite predict. Chances are literacy rates will drop, as well as political engagement and real social interactions.

Good quality content stimulates the brain, it calls on our emotions, teaches us something and can give us purpose. That purpose can be all sorts of different things, from taking political action to buying something. If you are in need of good quality copy, you can hire a copywriter like me to create the content for you.