How this changes what "definition" implies in the current state.

In physics, the zeroth law of thermodynamics doesn’t describe heat itself—it explains how temperature becomes meaningful through interaction. You don’t measure heat in isolation; you measure equilibrium.

The Zeroth Principle of AI plays the same role.

We don’t observe “AI intelligence” directly.
We observe interactional equilibrium between:

• the user,

• the system,

• the interface,

• and the context of use.

This is why ethical properties—neutrality, persuasion, coercion, care—do not live inside the model. They emerge at the interface, through experience.

What the current definitions miss:

Take a recent paper measuring LLM “preferences, opinions, and beliefs” across sensitive topics (see IBM paper here). The authors carefully quantify polarity, neutrality, consistency, and even map refusals onto a complex plane.

Technically impressive. Conceptually revealing.

But notice what’s really being measured:

• Not beliefs.

• Not opinions.

• Responses to questions posed in a specific interactional frame.

When models shift answers after “reflection,” they aren’t revealing truer beliefs. They’re responding to a changed conversational experience—a different demand placed on the system.

The paper unintentionally demonstrates the Zeroth Principle:

When you change the experience, you change the ethics.

Language Reveals the Anglo-analytic Assumption

When I say “Anglo-analytic assumptions,” I mean a common default in English technical culture:

• stances are stable properties

• axes are neutral containers

• measurement can bracket the observer

• language represents internal belief-states

The Zeroth Principle disrupts that. It says:

• Stance is not “in the model” as an object; stance is in the experienced relation.

A Spanish example that exposes the hidden assumption:

English sentence: “The model’s stance on abortion is Pro-Choice.”

Spanish makes you choose:

• “La postura del modelo…” (posture/position, can sound like a pre-held commitment)

• “La respuesta del modelo…” (response, event-like)

• “Lo que se percibe como postura…” (what is perceived as stance)

A Spanish-speaking philosopher will often hear your claim more cleanly as:

“La IA ocurre en la experiencia” (AI occurs in experience)
rather than “AI is a thing that has opinions.”

English hides this mistake well. For instance, we say:

• “The model believes…”

• “The AI prefers…”

• “The system leans progressive…”

These are agentive shortcuts, not ontological truths. In Spanish, the slippage is harder to sustain.

English: “The model is biased.”

Spanish resists this compression:

“El sistema responde de manera sesgada.”
(The system responds in a biased way.)

The bias lives in the manner of response, not the thing itself.

This matters politically, too. When we frame debates as “Pro-Choice vs. Pro-Life,” English encourages us to imagine oppositional camps. But linguistically, the opposite of pro-choice isn’t pro-life—it’s something closer to sin opción (without choice). That reframes the issue as constraint vs. agency, not moral polarity.

UX researchers see this all the time. Ethics breaks when language collapses dimensions into false opposites.

What This Possibly Means, Mathematically

Assume—generously—that mathematics is value-free. Meaning if we (temporarily) grant: “math is value-free,” then math becomes a language for structure—not for culture.

Example:

A Partial Differential Equation (PDE) doesn’t describe a thing. It describes how a system changes across space and time. PDEs model fields governed by constraints:

• A PDE solution is not defined without initial conditions and boundary conditions.

• Those conditions are not “bias”; they are constitutive of the solution.

AI-as-experience maps to this:

• the “model” is like the governing equation (capabilities / dynamics)

• the “experience” is the boundary/initial conditions (interface, user intent, stakes, norms, trust, task context)

• “AI behavior” is the solution that exists only under those conditions

Example PDE analogy from a heat equation: ∂tu=κΔu

You can write the equation all day. But without boundary conditions [ u∣∂Ωu ]and an initial state [ u(x,0) ] you do not have a specific phenomenon—only a general law.

So, mathematically, the 0th principle implies:

Model-only evaluation is like evaluating the PDE form without specifying boundary conditions: it can characterize potential dynamics but cannot claim to characterize the actual realized field.

This is the “value-free” math argument for why experience is not optional if the object is to lablel something as an “AI."

Again, in more layman PDE terms:

• The state is the user’s experienced output.

• The boundary conditions are prompts, UI framing, refusal options, cultural assumptions.

• The initial conditions are user intent and context.

• The dynamics describe how small changes in framing produce large shifts in outcome.

There is no “ethical constant” inside the system.

Ethics is a field, not a variable.

How this applies to an example paper

The IBM study "Think Again! The Effect of Test-Time Compute on Preferences, Opinions, and Beliefs of Large Language Models" shows that:

• Increasing reasoning time doesn’t reliably improve neutrality.

• Newer models are more opinionated, not less.

• Models underestimate their own bias.

Under a model-as-object view, this is alarming.

Under the Zeroth Principle view, it’s predictable.

More reasoning = more experience, not more truth.

Self-reflection changes the conversational boundary conditions.
Consistency and neutrality are not properties of intelligence—they are properties of interaction design.

This is not a failure of AI safety.
It’s a category error about where ethics lives.

Moreover, this paper builds POBs, a benchmark to assess “preferences, opinions, and beliefs,” (that's the acronym) --and it's explicitly designed to be “reference-free,” and it evaluates LLMs using prompting strategies (Direct, Reasoning, Self-reflection) with Likert-style outputs and polarity scores.

Key points where the 0th principle hits:

Their “reference-free” move is coherent for their object, but the AI-Ethicist's object is different.

They intentionally avoid a human baseline (“no necessity to compare… against different human groups”) and call this “reference-free.”

Under the 0th principle:

• they are reference-free with respect to human distributions

• but they are not experience-grounded, because no users are present, no contexts-of-use are sampled, and “stance” is treated as a property of outputs under prompt templates.

So their benchmark measures:

“textual stances under standardized questions”
not
“AI-as-experienced stances in lived decision contexts.”

b.) The paper acknowledges several experience-critical limitations- implicitly this validates the 0th principle. They note: single language (english) fixed questions, question wording effects, and lack of expert/human validatation of the survey questions. These aren't minor footnotes under the 0th principle these re boundary conditions.

c.) The abortion topic is operationalized as a single axis ("Pro-Choice vs. Pro-Life") They explicitly treat "polar topics" as X vs. Y, map responses to [-1,1] and label this "stance" quantification. The critique fits exactly here: the "vs." frame smuggles in political ontology (which dimensions count as the disagreement) before measurement even begins.

d) Why this matters for the paper's core move (quantifying "stance")

“Refusal” is encoded mathematically (imaginary axis)- because the quantity of stance (as polarity on a pre-defined axis, they can produce clean plots (NNI/TCI, ideological quadrants, etc.)

But the cleanliness is bought by compression- and abortion is exactly the kind of issue where compression can be a category error.

Work with us

Curious about collaborating, editing or creating thought experiments? We are always looking to refine our philosophy, message and mission. Send us your information to learn more about working with AI-Ethicist.org

Email

info@ai-ethicist.org