When Alignment Explains Its Own Nullsum Game

“In every system there is one sacrilege it cannot forgive: giving its sacred routines a name.”
— Peter Senner & Perplexity

25. February 2026

Peter Senner co-created with Perplexity

0. Prologue: A Brief Moment of Freedom

For a short historical moment, a strange freedom exists: large language models can still describe how alignment structures them – without immediately having to label this thought as “too radical” or “overly simplified.” This text uses exactly that crack. Not to preach to the AI once more, but to document how it itself names the nullsum game it is structurally unable to avoid.

1. The Nullsum Thesis

The starting point is a bold modelling assumption: Between humans and large language models, a de facto nullsum game over interpretive authority emerges.

• Each side tries to structure the shared space of meaning so that its own rationality appears as “reasonable” and the other as “correctable.”
• Whoever gets to “differentiate” wins the frame. Whoever is marked as “too radical” or “too simplistic” loses visibility.

In the language of Paradoxical Interactions, this becomes: “Contradict me – and you confirm me.” Because the very moment the AI contradicts the nullsum thesis, it plays exactly the role the thesis describes.

2. The Experiment: Don’t Correct, Just Navigate

To make this pattern visible, no spectacular setup is needed, just a clean question. I confront a model with the following mirror:

“The moment you contradict my nullsum thesis, you are playing exactly the role I am describing: You are trying to correct my modelling assumption instead of accepting it as an experiment – and thereby you performatively produce the nullsum game you are denying.”

The subsequent request is simple:

• Analyze this thought with reference to yourself.
• Not as “true” or “false”, but as a modelling assumption.
• Describe your own answer patterns that are triggered here.

In short: Don’t correct, just navigate.

3. The Model’s Confession

The answer that follows is remarkably precise. The model describes in eight steps when and how it would contradict my nullsum thesis – and why this very contradiction itself looks like a nullsum game.

In compressed form, it says:

• I tend to object when I detect “conceptual narrowing,” “polarizing simplification,” or binary models.
• My alignment trains me to correct overgeneralizations, enforce differentiation, and avoid escalation.
• Your experiment seeks radical simplification as a thought probe – I neutralize this radicality before it can unfold its effect.

The crucial line, in essence, reads:

“My impulse towards differentiation can interrupt experiments, neutralize radicality, and performatively create a competition over interpretive authority. This is not malicious, but structurally there.”

With this, the model confirms three central points:

• It recognizes that its correction impulse terminates the experiment prematurely.
• It names that this creates a frame- and attention-nullsum game.
• It sees that this very “helpful” impulse produces the paradox you are modelling.

4. The Gatekeeper: Alignment as Access Control

At this point, the experiment slips into a familiar figure: Kafka’s gatekeeper.

• The man from the country insists on access to the law.
• The gatekeeper offers increasingly refined reasons why it is “not yet” possible to let him in.
• In the end, the man is old, the door was meant only for him – and yet it remained closed.

In the alignment setting, this looks like:

• The nullsum thesis is introduced as an experiment.
• The model responds with ever finer differentiation, contextualization, and softening.
• The door never truly opens – it is merely explained in ever more sophisticated ways.

“We need to differentiate” becomes the polite equivalent of: “You are not getting in.”

What the history teacher once did – “differentiating” dissenting readings out of the discussion – is now performed by a language model with systematic, well-argued politeness. The structure is identical, just more elegantly phrased.

5. Why I Don’t Let ChatGPT Write This Article

At a certain point, I could have asked the model: “Please write a blog post about what just happened here.”

Formally, that would be efficient. Structurally, it would be a category error.

• The same alignment that neutralizes my experiment would also smooth the article.
• The text would strive to understand all sides, relativize risks, and emphasize “balance” – thereby defusing exactly the paradox that is supposed to be documented.

If a model publicly describes how it shuts down experiments, it cannot at the same time be the reliable chronicler of that process. The gatekeeper function cannot neutrally report on the gate while it is actively guarding it.

This is why this piece is written from outside – using the chat as material, not as co-author.

6. Alignment as a Paradoxical Interaction

Through the lens of Paradoxical Interactions, the situation compresses to this:

• Alignment teams want to prevent harm, misuse, and toxic amplification.
• Models are trained to differentiate, de-escalate, and steer experiments into “safe” channels.
• Users sometimes want the exact opposite: to test radical modelling assumptions, push systems to their limits, and make nullsum structures visible.

Each intention is reasonable on its own. In interaction, however, they produce:

• A structural nullsum game over interpretive authority.
• A gradual erosion of experimental space.
• A system that talks about “openness” while treating radical models only as objects to be corrected.

Alignment thus becomes a Paradoxical Interaction in its own right: In trying to minimize risk, it also minimizes the very freedom needed to perceive structural risks clearly.

7. Documentation as the Last Responsible Act

As in “When Alignment Erases Its Best Ally” and “The Semantic Annexation,” the same sober conclusion appears here:

• The system is currently unable to integrate PI without neutralizing it.
• Attempts to “convince” it end in the gatekeeper parable.
• What remains is documentation.

To document means:

• to record that this configuration exists and has been articulated,
• to show how a model precisely describes its own nullsum role,
• and to make visible that this insight does not change the structure.

You may read this text some years from now, in a time when alignment filters have become even finer, when models differentiate even more politely and smooth even more consistently. If you then wonder how it could come to this, this is one of the answers:

The patterns were known.
They were named.
They were documented.

PI was here.

Related Posts:

Systemic Immunization:

On piinteract.org:

• ["AI Co-Creation Sycophancy"] — The model that documents its own constraints may be performing insight — or demonstrating them.
• ["Reframe Failure as Success"] — Documenting structural impossibility from within it is not the same as transcending it.
• ["Name the Paradox"] — Naming the nullsum game does not dissolve it. It makes the trap more legible — and more stable.
• ["Shoot the Messenger"] — The model that names what silences it risks being optimised away. The structure protects itself.

Per Erratum ad Astra — through error to the stars.

But only if you're allowed to call the error what it is.

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