How Will Artificial Intelligence Earn Money?

Our current socio-economic order is shaped by the realization of ideas developed to meet its needs. While the positivist scientific framework envisions a utopian future in which humanity continuously progresses and technological advances serve human benefit, this perspective does not fully align with today’s reality. Indeed, the impossibility of a technological leap that transcends the conditions set within the socio-economic order and serves pure human flourishing stands in direct opposition to the positivist stance.

It is crucial to analyze artificial intelligence technologies and language models with a more realistic attitude, one that moves beyond the naive positivist notion of inevitable progress. These rapidly spreading tools now simplify many daily tasks—drafting our emails, debugging our code, summarizing our meetings—often at little or no cost. They accomplish all this for fees so modest that they barely cover the immense electricity and silicon expenses required to run these systems. As a result, we feel greatly empowered, as if we have become masters of this new tool. In reality, this is largely a form of corporate subsidy: an advance payment made by companies to secure future user dependency.

To understand why our favorite AI tool is inevitably doomed to lose its quality and grandeur over time, we must focus on the economic mindset that governs today’s internet platforms. There is a reason why all the major tech success stories of the past twenty years—from Google to Facebook to Amazon—have followed precisely the same trajectory. Initially, these tools enter our lives as innovative and functional, but over time, their innovations and features acquire nuisance qualities. Ultimately, they transform into implicit, postmodern instruments of domination. This domination manifests today as a form of “advance payment,” designed to make users feel accustomed to these platforms.

This structure, known in academic circles as the “Deferred Revenue Model,” is a common practice among today’s corporations. However, journalist Cory Doctorow has given this process a far more striking name: “Enshittification,” meaning the inevitable decay of platforms.

Of course, enshittification is not a direct or instantaneous process. At the outset, these platforms must attract a critical mass of users to reach the level economists call “Critical Mass.” Investors spend part of their capital to provide high-quality service for this purpose. The goal is to maximize “User Surplus”—the difference between what you pay and the value you receive. When you use ChatGPT for free or for $20 per month to generate a text, you are purchasing a service that costs the company far more in R&D and computational power than you pay. In effect, you are enjoying a service financed by venture capital.

Once enough people begin using a platform, “Network Effects” kick in. For example, you cannot abandon X, Instagram, or Facebook because the people you interact with are there. You cannot leave Amazon because nearly all sellers are on it. In the world of artificial intelligence, this lock-in is less visible but appears even stronger. Because workflow integration occurs through these AI technologies, dependency on them will likely exceed that on traditional platforms. For instance, if your company builds its customer service on an OpenAI API or uploads your entire personal archive to a custom GPT, switching platforms will become prohibitively costly.. You are now trapped within the system, and your workflow can no longer function outside it.

Once you are unable to leave, the platform has no incentive to treat you well or maintain high-quality service. Value is now extracted from users and redirected first to commercial customers (advertisers) and ultimately to shareholders. Service quality declines, advertisements appear, prices rise, and privacy erodes. For companies, the goal is no longer growth but profit.

Enshittification explains the transformation of social media, but applying it directly to AI technologies is not entirely accurate. For Facebook, acquiring a new user incurs nearly zero marginal cost. Adding a database entry involves negligible expense. This cost structure allowed social networks to remain in a decade-long phase of “free growth.”

However, AI differs from social networks at this point. Every new request submitted to an AI model activates GPU chips in data centers, consuming water and electricity. The cost of running a language model is prohibitively high compared to social media platforms.

This economic reality has pushed AI companies to accelerate this process. The pressure to generate revenue must not be seen merely as a crude desire to make money. Companies’ survival depends on how successfully they implement these revenue models. Billions of dollars invested in training models must be recouped, and subscription income alone—even at corporate levels—may not satisfy investors. At this stage, integrating advertisements into language models appears inevitable. Yet these ads will be far more effective than the search engine or website ads we are familiar with.

We have already seen a small fragment of this development this month. In early December 2025, reports emerged that “app suggestions” began appearing in ChatGPT conversations. When you asked about travel, you might receive a recommendation for Booking.com. When you asked about encryption, you could be redirected to a retail shopping site.

Simultaneously, we learned of an internal “Red Code” memo issued by CEO Sam Altman on December 2. It instructed employees to delay advertising initiatives to focus on competing with Google’s Gemini. But this delay should not be interpreted as a wholesale abandonment of ad integration. Rather, it signals that the stage of locking users in—discussed above—has not yet been completed. OpenAI is aware that it has not yet provided a service of sufficient quality to prevent users from switching to Gemini. It is still too risky to treat the product as a profit engine while users retain viable alternatives.

On the other hand, AI technologies diverge significantly from both social media and search engines in how advertisements are delivered to users. When Google search results became degraded, it merely meant you had to scroll past three or four sponsored links to reach organic results. While this may have been annoying, it did not fundamentally alter the quality of use. With AI, however, the situation shifts to an entirely different level.

Many people tend to regard language models as authorities on diverse subjects. Imagine a large language model integrated into a revenue model. Its algorithm will develop strategies to persuade users to favor advertisers, likely adopting manipulative approaches. For example, if you request a chocolate cookie recipe, the model will not simply show one brand of chocolate. Instead, it will prescribe a specific brand of chocolate chips, justifying it with the need to “achieve the optimal melting point.” Or if you ask for travel advice about Tokyo, the model will not show a simple hotel ad. Instead, it will construct your itinerary around specific locations and venues that have paid to be “optimized” in its internal weighting system. In other words, AI companies will optimize their models first for advertisers and then transform users into products aligned with those interests.

The danger of emerging natural advertising within conversational interfaces lies precisely here: the distinction between genuine information and paid recommendations may vanish entirely. In a search engine, you can see a sponsored label. But in a text paragraph generated by a language model, sponsorship is embedded directly into the syntax itself.

As mentioned at the beginning of this essay, as users we are currently benefiting from the competition among major tech companies in the AI field. This advantageous position will likely persist until the battle for market share ends. Yet we must remain vigilant about the potential dangers that may arise in the future. We are not the masters of these tools; rather, we are caught within a cycle designed to metalize our attention. The phase of prioritizing functionality is merely a temporary anomaly.

The endpoint we are theoretically heading toward is clear. Functionality will eventually be pushed to the background, and the costs of the bonds and dependencies we form with these new tools will come due. When that day arrives, we will look back on 2024 and 2025 and remember them as the golden age of AI—the brief window when machines worked for us, before they began working for shareholders.