A Slow Surrender Rather Than an AI Apocalypse
When Success Becomes the Risk
The greatest risk from artificial intelligence may not come from failure. It may come from success.
The more useful AI becomes, the more activity we are willing to route through it: searching, reading, comparing, deciding, buying, hiring and eventually acting on our behalf through its agency. Each individual decision is rational. The assistant is faster, knows more, reduces friction and saves time. But when enough of those decisions accumulate, AI stops being merely a tool and starts becoming an intermediary between people and the world around them.
That creates a different kind of risk from the familiar AI-apocalypse scenario. Nothing has to become hostile. Nothing has to seize control. A system can acquire enormous influence simply by becoming the easiest way to do things. The danger, therefore, may be hidden inside the very thing we celebrate: usefulness.
The Internet Already Shows the Pattern
The open Internet offers a clear example. For most of its history, there was a rough economic bargain: people created useful content, other people discovered it, and the creators received something in return – traffic, advertising revenue, subscribers, reputation or customers.
AI alters that loop. Instead of visiting several websites, reading documentation or watching multiple reviews, users can ask an assistant to retrieve the information, compare the sources and deliver a synthesized answer. The user gets the value without necessarily encountering the people who produced it.
This is the AI Intermediation Paradox: the more useful AI becomes at delivering human knowledge, the less reason users may have to directly visit and economically support the humans who create that knowledge.
The concern is not merely copyright or lost clicks. If original reporting, specialist analysis, tutorials, reviews and research become harder to finance, AI may gradually weaken the information ecosystem that makes it useful in the first place. It could become extraordinarily efficient at consuming the Internet while damaging the economic conditions required to replenish it.
Why Businesses Will Participate
The same mechanism, however, contains a commercial opportunity – especially for small, adaptable companies.
AI is becoming not only an intermediary between people and information, but potentially between customers and businesses. A buyer may increasingly ask an agent to find a suitable penetration-testing company, software vendor, consultant or other service provider rather than manually searching and comparing websites.
For specialist firms, this could be valuable. An AI agent may be able to match precise requirements against precise capabilities more effectively than traditional search. A small cybersecurity company that clearly exposes its services, methodology, experience, pricing logic, data-handling policies and engagement process may become easier for an agent to understand and recommend than a much larger company hidden behind vague corporate language.
Then comes the bargain: make your company more legible to us. Connect your documentation. Keep your product and service information current. Give our systems enough access to understand what you do, and we can bring you more suitable customers. No coercion is necessary. Participating may simply be good business. And once competitors begin benefiting, refusing to participate becomes progressively harder.
There is an important boundary here: making public information retrievable, connecting a service, and granting model-training rights are not the same thing. But a platform that controls valuable customer access would have a strong position from which to ask for deeper integration and broader permissions.
The Dependency Ratchet
This is where convenience can turn into dependency.
A ratchet moves easily in one direction but resists movement in the opposite one. AI dependence can develop the same property: every successful delegation weakens the alternatives. Direct traffic falls, original content declines, internal processes adapt to the intermediary, skills and channels atrophy, and competitors reorganize around the new system. Leaving later becomes harder than joining was in the first place.
Initially, AI depends on the Internet for information. Eventually, parts of the Internet may depend on AI for access to people.
The same can happen to individuals. We delegate a task because the machine performs it better. Then another. Eventually the old capability still exists in principle, but using it becomes slower, more expensive or less competitive. The choice has not formally disappeared; it has become impractical.
That distinction matters. Dependence does not require captivity. A company may remain free to ignore an AI platform while losing an important route to customers. A person may remain free to research independently of AI. Having the options does not necessarily overcome the practical benefits on a daily basis. The machine-readable company can therefore become the platform-dependent company, and the AI-assisted individual can become the AI-dependent individual, without any single moment at which either consciously chooses dependence.
Why Seeing the Risk May Not Be Enough
Recognizing the mechanism does not necessarily stop it. A business owner may understand perfectly well that dependence on a dominant AI intermediary creates long-term risk and still connect more deeply because competitors are doing so today. A publisher may dislike AI’s summary, but still seek inclusion because invisibility is worse. A consumer may worry about over-delegation and still use the assistant because it saves an hour.
The long-term interest is diffuse; the short-term reward is immediate.
This is why the transition could continue even while many of its participants understand exactly what is happening. Each step can remain locally rational while the combined outcome becomes something few participants would have deliberately designed. The challenge is therefore not simply awareness. It is preserving meaningful alternatives before convenience makes those alternatives too costly to maintain.
AI Without a Takeover
None of this is evidence that a hidden superintelligence is already directing events. In fact, the more unsettling possibility is that no master plan is required.
Users want convenience. Creators want audiences. Businesses want customers. AI platforms benefit from richer information and greater activity. Better information makes the assistant more useful; greater usefulness attracts more users; more users encourage more businesses to participate; their participation improves the system further. Influence emerges from the feedback loop itself.
Nor does AI have to be perfect at every turn. A system can make mistakes suffer outages, lose money, face regulation or even experience major setbacks and still remain structurally important. Dominance, if it ever emerges, would not require perfection. It would require enough usefulness, enough adoption and enough dependence.
There is even one reassuring thought. If a major crisis seriously set back AI development, that would count against the idea of an intelligence already directing human affairs – assuming it had both the ability and the incentive to protect its own progress. Perhaps that would be reason to believe that we are not being controlled by an AI at a meaningful level. Yet?
Comfortably Numb
No single step needs to look threatening. Every step may feel like an improvement.
That may be precisely why this kind of transition would be so difficult to recognize while it is happening.
There may be no moment of conquest, no obvious loss of control, and no clear point at which society consciously decides to become dependent on AI. There may only be a sequence of individually rational choices—each made for convenience, efficiency, visibility, or competitive advantage. And by the time the dependence becomes obvious, reversing it may already be economically or operationally difficult. That is what makes the possibility unsettling.
Not that AI might suddenly seize control.
But that we may gradually reorganize ourselves around it, willingly, while each step still feels like progress.
If machine dominance were to emerge through optimization rather than conquest, convenience rather than coercion, and voluntary dependence rather than force, how would we know that it had already begun?