Has the enshittification of AI begun? What it means for your business

OpenAI has announced that ChatGPT will start showing visual ads beside the images it creates. It’s the latest step in a year that has seen the world’s most popular AI assistant, used by more than a billion people a week, become an advertising business.

We’ve seen this pattern before. Social media, search and online shopping all started out as great deals for their users, then changed once the business model shifted towards advertisers and shareholders. The author and technology activist Cory Doctorow has a name for it: ‘enshittification’. In this blog, our CEO, Gareth Murphy, explores whether it has started to happen in AI and what it means for how your business adopts AI tools.

What is enshittification?

At first, a platform is good to its users, often running at a loss to build an audience. Once users are locked in, it shifts value to its business customers, such as advertisers and sellers. When those businesses are locked in too, it squeezes everyone to return as much as possible to its shareholders. The product still works, just less well for the people using it.

The pattern only works when leaving is harder than staying. It also happens in small steps that each feel too minor to complain about.

Lessons from Amazon’s marketplace

Amazon is the clearest example. It won shoppers with low prices and fast Prime delivery, then won sellers with access to that audience and tools that made selling easy. Fulfilment by Amazon handled storage, delivery and returns, while the Buy Box, the main ‘Add to Basket’ button on each product page, let a small seller with a good price win the sale ahead of bigger rivals. Sellers built their businesses around Amazon, and Marketplace Pulse, an e-commerce research firm, estimates it now handles around 70% of US online marketplace sales.

Over time, those same tools became a way to squeeze sellers. Using Amazon’s fulfilment became the price of the Prime badge and advertising became the price of being seen. Marketplace Pulse estimates that a typical seller now pays Amazon more than half of their revenue in fees, fulfilment and advertising, up from 40% five years earlier. Shoppers haven’t escaped either, with Prime Video showing ads by default since 2024 unless members pay extra.

The US Federal Trade Commission alleges that Amazon pushes sellers down its rankings if they sell more cheaply elsewhere. Amazon denies the allegations and the case is due to go to trial in March 2027. In the UK, Amazon gave commitments to the Competition and Markets Authority in 2023 over how it treats sellers, and in February 2026 the Court of Appeal cleared the way for sellers and consumers to bring claims against Amazon UK.

The lesson for businesses is that no single step looked outrageous at the time. Each new ad slot, fee rise or Buy Box tweak was defensible on its own. Amazon’s advertising is also already arriving inside AI, through a pilot that lets its advertisers extend their campaigns into ChatGPT.

Is AI following the same path?

ChatGPT’s ad business launched in early 2026 and, by OpenAI’s own projection, was earning at a rate of around £750 million a year by August. Ads appear on the free plan and on Go, the cheapest paid plan, while the more expensive Plus and Pro plans remain ad-free. Advertisers now have self-serve tools, can pay only when an ad leads to a sale and can use new formats such as Sponsored Agents, which let users chat with an advertiser’s own AI.

In September, OpenAI reportedly stopped accepting ads from rival image and audio tools, the same categories where it had just launched its own products. OpenAI hasn’t confirmed this, but it’s the kind of self-favouring Amazon is accused of.

The financial plan points the same way. OpenAI hasn’t published a forecast by plan, but internal projections reported by The Information show where subscribers are expected to go over the course of 2026. Most are forecast to move to Go, the cheapest plan and the only paid plan with ads.

Projections for 2026. Source: internal OpenAI projections reported by The Information, converted to pounds.

The longer-term picture goes further. By 2030, OpenAI expects around £75 billion a year from advertising, roughly 36% of its revenue. On our own rough estimate, that would make ads worth more than twice all consumer subscriptions combined. With heavy losses and around £1 trillion committed to computing power, OpenAI has every incentive to show more ads, target them harder or find new formats. We’ve converted US dollar figures at around 75p to the dollar.

Targeting is where AI differs from search. Where personalisation is switched on, ads can be matched to past chats and ChatGPT’s memory as well as the current conversation. Advertisers don’t see the conversations, but people tell chatbots things they would never type into a search box.

There’s a fair case that this is a business model rather than decay. Free users have gained, with limits on text chats removed in August alongside a new default model. In Doctorow’s terms, advertising is currently paying for more generosity to users, not less.

However, regulation is making a difference in the UK. UK and EU users must give explicit consent before their past chats and memory are used to personalise ads, and Meta can’t use AI chats for ad targeting here at all. Switching to a different AI assistant still only takes minutes, and good alternatives exist.

Our view is that ChatGPT is at the very start of Doctorow’s second stage. The service is still good, but the business behind it now depends on advertisers. The real test is what happens when the interests of users and advertisers start to pull apart, and whose side the platform takes.

Could AI models get worse?

For businesses, a quieter risk is the quality of the models themselves. Language models learn from human writing, and the web is filling up with AI-generated content. Researchers from Oxford and Cambridge showed in Nature in 2024 that models trained repeatedly on other models’ output gradually lose touch with the real world, a problem known as model collapse. Later research suggests this can be avoided if fresh human data keeps being added, so it’s a risk to manage rather than a certainty.

Good human-written data is also becoming more expensive. Reddit now licenses its data to AI companies and Cloudflare blocks AI crawlers by default on new sites. Creators deserve to be paid, but these moves favour the providers with the deepest pockets and could leave cheaper and free services learning from a thinner diet.

There’s a commercial risk too. The most capable models tend to appear on premium plans first, and users rarely know which model is answering them or when it changes. When OpenAI launched GPT-5 in August 2025, it removed older models overnight and only brought back the popular GPT-4o for paying users after a backlash. On the other hand, today’s free models outperform the paid ones of two years ago. The question is whether that continues once growth slows.

The role of open-weight models

Chinese labs including DeepSeek and Alibaba’s Qwen team have released open weight models this year, meaning anyone can download them, run them on their own systems and adapt them. On standard tests, the best now sit within a few points of the leading US models at around a tenth of the price or less, although many headline scores are self-reported.

This gives businesses somewhere credible to go if a closed platform gets worse. It also makes it harder for closed providers to charge a premium for intelligence alone, which pushes them towards advertising and lock-in. For UK organisations, data sent to Chinese-hosted services may be stored in China, so running open models on your own or a Western cloud provider’s infrastructure is the safer route, with the right technical capability and governance in place.

Signals to watch

If AI does follow Amazon’s path, the changes are likely to look like new plans or technical updates rather than a worse product. Signs to look out for include:

• Sponsored content appearing inside answers rather than beside them
• More ads per conversation, or a narrower definition of sensitive topics
• UK ad personalisation moving from opt-in to opt-out
• Quiet cuts to free or cheaper plans, such as shorter answers or lower usage limits
• A widening gap between the models available on different plans
• Conversation data being used for more than ads
• Features that only work inside one provider’s ecosystem, making it harder to leave.

Protecting your business through people, process and technology

The risk isn’t that your AI tools get worse overnight. It’s building important work on a platform whose incentives drift away from yours and only noticing when it would be expensive to leave. That’s why, at Fluid, we approach AI adoption the same way we approach any change: by looking at people, process and technology, in that order.

People: Business and Enterprise plans are ad-free, but many staff still use free personal AI accounts for work. A clear acceptable use policy, backed by business licences and training, closes that gap and reduces the shadow IT risks that come with unapproved tools.

Process: Treat AI providers as strategic suppliers. Ask how they make money, how your data is used and how easily you could move. Contracts should set out how your data can and can’t be used, whether commercial content can ever appear in the service and how much notice you’ll get of significant changes.

Technology: Keep the option to leave. Hold your prompts, knowledge and data where you control them, and avoid building critical processes around one provider’s features. Where it makes sense, design your systems so you can switch between models without starting again.  This is where a solution-agnostic approach pays off. Being solution-agnostic means we don’t start with a preferred product, platform or vendor in mind, and we don’t earn anything from recommending one over another. We start with your goals, your people and your processes, and only then look at which AI tools genuinely fit. If a provider’s pricing, terms or quality change, you aren’t tied to it, because your approach was built around your business rather than someone else’s sales target.

Our IT Transformation Director, Jason Lock, explains: “AI needs to be aligned to the business strategy and there needs to be a clear understanding of the technology and the data. These then need to be overlaid with the organisation’s culture and governance.”

Doctorow’s point is that platforms don’t get worse because the people running them are villains. They get worse because the incentives reward it and nothing stops them. That isn’t a reason to hold back on AI, but it is a reason to adopt it with your options open.

If you want to make the most of AI without tying your business to a single provider, get in touch today and find out how we can support your AI strategy.