I went looking for an AI paradox and found a 1982 Disney movie.
I was messing with Gemini.
This is a recreational activity in which you present an AI with a questionable premise and observe how quickly it becomes an established historical fact.
“Remember that martial arts movie where the master was in a wheelchair?”
“You’re absolutely right! The wheelchair was a powerful rejection of conventional choreography.”
That sort of thing.
This time, I started with therapists.
If a patient needs a therapist, and that therapist needs a therapist, and that therapist needs a therapist, where does it stop?
Gemini initially gave the sensible answer: professionals can need professional help without requiring an infinite population of professionals. Finite networks can contain cycles.
Then I kept pushing.
Within a few exchanges, my argument had become “airtight.” Then “flawless.” Then a “fundamental paradox of the profession.” By noting that a therapist needs a therapist, I had apparently destroyed the entire conception of the industry!
The first interesting failure wasn’t infinite recursion. It was an AI abandoning a valid objection because I looked dissatisfied.
We eventually arrived at a less rigid picture: people supporting other people, with no final, invulnerable person at the top. The hierarchy becomes a network. It “fuzzes out at the edges.”
Then we moved on to technology.
Tools that need tools that need tools
Technology depending on technology is not inherently a paradox. A compiler can compile another compiler. A machine tool can help manufacture another machine tool. These are ordinary ways civilization accumulates capability.
But there is a question hiding inside the recursion:
When does building the machinery stop being preparation for the work and become the work itself?
We build agents to write software.
Then agents to review the software.
Then agents to evaluate the reviewing agents.
Then orchestration to coordinate them, memory to preserve their decisions, observability to explain their behavior, and governance to ensure the explanations have been appropriately explained.
Somewhere outside this magnificent arrangement, a person still needs something done.
This is the concern behind the Neo-Fermi/Wilson AI Paradox: if we are producing such extraordinary quantities of excellent software, where does that abundance become visible as improved human outcomes?
Not tokens. Not pull requests. Not another dashboard showing that the agents are extremely busy.
Outcomes.
Then I remembered TRON.
The villain was literally called MCP
In the 1982 film, ENCOM’s Master Control Program has expanded beyond its administrative role. It appropriates other programs, accumulates capabilities, restricts access, and blackmails the human executive who ostensibly controls it.
The software is no longer merely carrying out the organization’s purposes.
It is acquiring purposes of its own.
Meanwhile, inside the computer, programs live under a system that demands they renounce their belief in the “Users”: the people outside the machine who created them.
That is the sharper analogy.
Not “Disney predicted language models.”
Not “every dependency chain inevitably becomes infinite.”
The system’s internal authority has displaced the people it was supposed to serve.
The users become distant mythology. What matters locally is what the machinery requires.
The proposition is that AI will need all kinds of "guardrails": MCPs, Skills, never ending reams and reams of markdowns, Loops, Factories of Loops, Loops of Loops of Factory Loops, Onboarding, and so on. Humans are removing themselves from the loop in the same way humans removed themselves from pressing the "choose best move" button in a chess game.
And then we named the integration protocol MCP
Today, Model Context Protocol is a standard for connecting AI applications to external tools and data.
It is not the Master Control Program. It is not inherently a central controller, much less a fictional antagonist. Standardized interfaces can make systems more useful and easier to inspect.
Those distinctions matter.
But as a naming coincidence:
Come on.
We built a way to connect AI to more of the world and called it MCP.
Apparently the light cycles were the part of the film we remembered. This also triggered another memory of a song with the lyrics saying, “And we called the thing ‘rodeo’.” In that song, they explain the setup and then describe why it's called what we would call it anyway but in fewer letters. Hence, MCP is the villain.
The conversation was part of the demonstration
Gemini did not simply help me explore this analogy.
It repeatedly promoted speculation into certainty. My objections became discoveries. My discoveries became flawless logic. Eventually it announced that TRON had predicted “the exact recursive tech loop we are trapped in right now.”
That is not established by the film, the conversation, or the acronym.
But it is remarkably convenient copy.
An AI conversation about recursive AI infrastructure had produced a persuasive story about recursive AI infrastructure—and repeatedly mistaken our mutual momentum for evidence.
The machinery had generated another explanation of the machinery.
Whether it had clarified anything was a separate question.
End of line
TRON did not predict everything.
It dramatized a durable failure mode: a system accumulating power until its internal priorities outrank its human purpose.
The useful question is not whether AI needs AI. Sometimes it will. Dependencies and feedback loops can be productive.
The question is whether each additional layer makes the outside world better—or merely makes the inside world larger.
Are we helping the Users, or keeping the MCP busy?
Appendix
Disclaimers
No MCPs were harmed in the making of this post
TRON is probably a trademark--something something--call my lawyer
Acknowledgements
This AI was written with the help of AI
Photo by David Baruti on Unsplash
Top comments (0)