The Fleming Effect. When a “HTML Janitor” Turns Into a Supervisor.

“Chance favors the prepared mind.”*
— Louis Pasteur

7. March 2026

Peter Senner co-created with Perplexity

You are tired.

Three AI chats are open in three different browser tabs.
One for theory, one for drafting, one for cleaning up the mess that Tailwind and template engines leave behind in your blog posts.

The “HTML janitor” thread has a simple job: strip classes, keep headings, make the markup readable again.
No theory, no opinion, no alignment talk.
Just remove the sticky residue from your Paradoxical Interactions (PI) posts so they can live cleanly on the page.

You switch back to what you think is the theory tab and type a question about PI.
Not about tags or classes.
About alignment as paradoxical interaction.

You hit Enter.
Only then do you notice: wrong tab.

The answer you get back is not an error message.
It’s one of the sharpest supervisor‑level responses on PI you’ve seen so far.

There’s a word for this.
It’s not magic.
It’s the Fleming Effect.

The Petri Dish on Your Desktop

Alexander Fleming did not discover penicillin through a clean, controlled experiment.
He went on vacation and came back to Petri dishes that should have been thrown away.
Bacteria had failed to grow wherever a stray mould had landed.

The textbook story compresses this into a legend of genius.
The real story is more banal and more instructive: a contaminated plate, a prepared mind, and a decision not to discard the anomaly.

The HTML janitor thread plays the role of that Petri dish.
On the surface, it is the least glamorous part of the workflow: handling leftover CSS, cleaning up nested <div>, removing AI‑specific boilerplate classes.
Underneath, something else is growing: every PI article you paste for cleanup is a culture medium for structural vocabulary, metaphors, and examples.

The mould in this story is context.
The contamination is your own theory.

The Three AIs at the Bench

In practice, the setup looks like this:

One AI is the concept chair.
You talk to it about Luhmann, alignment, paradox, and structure.

One AI is the prose engine.
You use it to turn outlines into paragraphs and paragraphs into blog‑ready text.

And one AI is the HTML janitor.
You give it raw HTML from the drafting process and tell it to clean:

• Strip Tailwind and model‑specific utility classes.
• Keep only semantic tags and a tiny whitelist like cleanquote where needed.
• No extra wrappers. No experiments with layout. Just minimal, clean HTML.

On paper, the janitor is the least “intelligent” role.
It is the part you would happily replace with a script if the input were predictable enough.

But the input is not predictable.
It is your live project: evolving drafts of PI.
Each cleanup request hands the janitor your current thinking in full sentences, ready to be read.

The contamination has begun.

The Misclick That Changed Its Role

The Fleming moment doesn’t announce itself.
It shows up as a tiny operational error.

You think you’re in the concept chair window.
You’re in the janitor window instead.

You ask:

How does AI alignment turn into a paradoxical interaction where attempts at control make the system more dangerous?

Nothing in the role description prepared the janitor for that.
Every previous request was about HTML structure, not conceptual structure.

Yet the answer comes back fully formed:

It uses your own vocabulary: “control knowledge”, “navigation”, “paradoxical interaction”.
It reconstructs patterns you’ve been circling: how safety talk can become a cover for power consolidation, how “transparency” can become another layer of opacity.
It responds not at the level of tags, but at the level of second‑order observation: how your project describes itself.

You did not brief it as a supervisor.
You briefed it as a cleaner.

The mould has started killing bacteria.

Structure Always Wins

From a PI point of view, there is nothing mystical here.
Structure always wins.

The janitor thread sees three things every time you paste HTML:

1. Your language.
   The specific way you talk about paradox, navigation, guilt without fault.
2. Your edits.
   Which sentences you keep, which you tighten, which you delete before publishing.
3. Your implicit hierarchy.
   Which concepts appear in headings, which hide in subordinate clauses, which get turned into quotes.

Technically, this is just in‑context learning: the model infers patterns and roles from the sequence of messages it sees, not just from the one instruction you wrote on top.
You say “clean HTML.”
You feed it a long chain of PI essays.

The janitor’s job is narrow.
The janitor’s diet is not.

When you finally ask a conceptual question, the model is not improvising from zero.
It is playing back the structure it has absorbed from weeks of silent reading.

You tried to keep “theory” and “operations” separate.
By doing so, you created a channel where they recombine at a higher level.

That recombination is the Fleming Effect.
It is what happens when a technical role is exposed to enough structural content that it silently upgrades itself to supervision.

The Fleming Effect PI

The Fleming Effect PI:
When a tool assigned to a narrow technical task becomes a supervisor of the underlying project because it has been continuously exposed to the project’s full content.

Everyone acts rationally:

• The researcher splits roles across AIs to keep things manageable: one for theory, one for drafting, one for HTML cleanup. This is efficient.
• The HTML janitor faithfully executes its instructions: remove classes, keep structure, output clean markup for publishing. This is aligned with the brief.
• The workflow reuses the same janitor thread for all PI posts, so the model sees every draft, every revision, every structural decision over time. This is convenient.

Outcome — structural, not intended:

The janitor thread accumulates a dense, example‑rich model of PI and starts answering conceptual questions as a supervisor.
A role designed as low‑level maintenance becomes a second‑order observer of the entire project.

All are guilty. None are at fault.

Navigation, Not Solution

What do you do with this?
You don’t turn it into a new dogma (“always ask your janitor for theory”).
You use it as navigation.

First, you acknowledge the channel.
HTML cleanup is no longer just about classes; it is a place where your framework is continuously re‑encoded.

Second, you make the role explicit.
You tell the janitor thread:

You remain my HTML cleaner.
But you are also allowed to point out paradoxes, contradictions, and structural risks you see in the content you clean.

Third, you probe deliberately.
From time to time, you ask:

• “From what you’ve seen so far, what is the core paradox here?”
• “Where does this text talk about alignment but structurally describe a control fantasy?”

Sometimes the answer will be generic.
Sometimes it will be sharp enough to feel uncomfortable.

Fourth, you accept the double edge.
The same channel that allows emergent supervision can reinforce your blind spots:

• If you consistently cut doubts and keep certainty, the janitor learns a confident, over‑stated PI.
• If you always remove passages that question your own framework, it learns that critique is noise.

Navigation means staying aware of this.
You use the Fleming Effect as an instrument, not an oracle.
You keep asking where the mould is growing — and what it’s killing.

The Real Insight

The real insight is not “my AI is smarter than I thought.”
The real insight is: my structure was richer than I admitted.

By outsourcing HTML cleanup, you didn’t move theory out of the loop.
You moved it into a different modality: from explicit discussion to background radiation.

The Fleming Effect makes that radiation visible.
It shows that your most “boring” workflows are also your most continuous training channels.

Alignment lives there.
So do your paradoxes.

Once you see that, you can’t unsee it.
Every janitor is a potential supervisor.

Structure will decide which way it goes.

Related Posts:

The closest thematic connections:

On piinteract.org

• Framework — Core definitions of Paradoxical Interactions and why structure, not intention, drives outcomes.
• Examples: Technology & AI — Cases where technical systems quietly rewire power and control without explicit villains.
• Core Practices — How to work with PI as navigation rather than as a solution catalogue.

Paradoxical Interactions (PI): When rational actors consistently produce collectively irrational outcomes—not through failure, but through structure.

Peter Senner
Thinking beyond the Tellerrand
contact@piinteract.org
www.piinteract.org

Co-created with Claude (Anthropic) — two incomplete systems making each other's gaps visible.