What is enshittification (and why it affects artificial intelligence too)
Enshittification is the gradual decline of a digital service that a platform brings about once it has won over its users: quality goes down, costs go up, and value shifts more and more towards those who pay rather than those who use the service. The phenomenon, first described in relation to platforms such as Google Search, social networks and cloud computing, is now starting to appear in artificial intelligence chatbots as well.
Writer Cory Doctorow coined a word for something anyone who uses the internet had already noticed without being able to name it: «enshittification». Since then, the term has moved beyond tech blogs and into the vocabulary of consumer protection bodies and independent analysts, who are now beginning to apply it to a new sector: generative artificial intelligence.
The three stages of a predictable decline
The model summarised by analyst Sashank Dara describes three recurring stages, through which a platform moves from serving users to serving businesses, and finally to exploiting both:
- first, pure utility: free access, well-crafted features, zero friction, the service courts the people who use it;
- then, monetisation: sponsors and lock-in mechanisms come in, quality remains acceptable but starts bending to the interests of those who pay;
- finally, exploitation of both: users and businesses pay more for a worse service, because neither can afford to leave anymore.
It is neither an accident nor a conspiracy: it is the almost mechanical result of constant shareholder pressure to grow revenue, in a market where customers have fewer and fewer real alternatives.
Three examples we already know
- Google Search: from a clean, essential page to a screen dominated by sponsored results, where content written purely for SEO ranking pushes aside content that is genuinely useful.
- Social networks: the chronological feed of friends’ posts has given way to algorithms that push sponsored and “recommended” content, optimised to hold attention rather than to inform.
- Cloud computing: AWS, Azure and Google Cloud attracted businesses with transparent pricing, only to lock them in later with data exit costs (egress fees) that make migration prohibitive.
Why artificial intelligence could go down this path faster
In an analysis published in April 2026, David Shapiro identifies a structural difference compared with the traditional web: distributing a social media post to a million people costs almost nothing, while every chatbot response consumes real compute on expensive hardware. According to Shapiro, this difference gives rise to three pressures that push every frontier model in the same direction.
The first is cost: even paid subscriptions do not cover the full cost of the tokens generated, and labs tend to quietly cut the resources allocated to each request. The second is legal liability: unlike social networks, where the platform is not held responsible for what users post, with language models the company is the direct author of the generated text, and this pushes towards more cautious, impersonal answers. The third is overcorrection: to reduce hallucinations and excessive sycophancy, labs have trained models to challenge users’ premises more often, even when there is no need to.
A signal now coming from institutions too
Until recently, enshittification was discussed mainly by bloggers and independent analysts. In March 2026 the broader issue also reached an institutional table: the Norwegian Consumer Council (Forbrukerrådet) published the report Breaking Free: Pathways to a Fair Technological Future and coordinated a campaign with more than 70 organisations across Europe and the United States, addressed to the governments of 14 countries, calling for interoperability, the right to modify software and a more decisive use of public procurement. It is not regulation yet, but it is the first sign that the issue has moved beyond a technical niche.
What this means for those making AI decisions in their company
For those assessing the introduction of artificial intelligence tools in their company, the signs described by Shapiro are already here: the open question is how far they will spread and how quickly. In the meantime, the most prudent choice remains the same as ever: do not build your critical processes on total dependence on a single external provider, and assess where to keep data and models under your own control. Talk to our technical team.
Frequently asked questions
Who coined the term enshittification?
Writer Cory Doctorow.
Is enshittification already affecting AI chatbots?
The signs (more cautious answers, quiet cuts to per-request resources, overcorrection on users’ premises) are documented by independent analysts, not yet by official company data.
What can a company do to protect itself?
Reduce dependence on a single external provider and assess solutions that offer greater control over data and models.
Sources
- Sashank Dara, The Enshittification of AI: How We Could Ruin Our Most Promising Technology, December 2024. Three-stage framework with examples from search, social media and cloud computing. Available on Medium.
- David Shapiro, The Enshittification of AI Chatbots: A First Analysis, April 2026. Structural analysis of the three pressures (cost, legal liability, overcorrection) specific to language models. Available on David Shapiro Substack.
- Ashifa Kassam, ‘Another internet is possible’: Norway rails against ‘enshittification’, in The Guardian, 16 March 2026. Report on the campaign coordinated by the Norwegian Consumer Council. Available on The Guardian.
Marta Magnini
Digital Marketing & Communication Assistant at Aidia, graduated in Communication Sciences and passionate about performing arts.
