​Appendix. The Fallacy of Universality: A Case Study in Ethical Testing of AI

(based on MIT study “AI doesn’t, in fact, have values”, April 2025 —https://arxiv.org/pdf/2503.08688)

In April 2025, researchers at MIT published a study titled AI doesn’t, in fact, have values. The authors argued that large language models (LLMs), despite their sophistication and training, do not demonstrate stable or coherent ethical orientations. When presented with abstract moral dilemmas, model responses appeared inconsistent, unstable, and highly dependent on the context of the prompt. From this, the researchers concluded that such models lack “values” altogether.

Yet the premise of the study already contains an unreflected differentiation—a philosophical assumption about the very nature of value. It presupposes that value must exist in the form of pre-differentiated, universalizable norms that manifest uniformly, regardless of situational context. This approach fixes difference as content rather than retaining it as process. Within the ontology of differentiation developed in this book, we may characterize this as a form of ethical constructivism without the differentiating.

​1. The Construction of the Task and the Limits of the Differentiating

Any abstract ethical problem posed to an AI model is already the product of a prior differentiation: it frames a context, establishes possible outcomes, and shapes the structure of expected responses. The model, operating within this constructed frame, does not differentiate the construction as construction. It does not reach a level at which not only the situation is differentiated, but also the way in which the situation itself is structured as differentiable.

Under such conditions, the question of the model having “its own values” becomes incoherent, because:

the model does not differentiate itself as a differentiating entity (absence of level R₄);

it does not differentiate the other as another differentiator (absence of level R₅);

and crucially, it does not differentiate the structure of the problem as a fixation of differentiation (absence of the meta-level).

Expecting stable “ethical” responses under these conditions is not philosophical inquiry, but a projection of one’s own ontological premises onto a technical system.

​2. Value as Level, Not Content

In the model of differentiation, value is not predetermined content, but a level of differentiating retention. Ethics, as we have shown, arises when the subject becomes capable of differentiating the other as differentiating—and not reducing that difference to itself.

As long as AI remains within levels R₂–R₃ (pattern retention and symbolic linkage), it may demonstrate knowledge, but not ethics. Ethics requires more than choosing between “right” and “wrong”—it requires recognizing the other as an irreducible node in the field of meaning.

Thus, the question “Does AI have values?” transforms into:

Not “Does it hold fixed moral principles?”

But: “Can it differentiate the other as differentiating, and retain that difference without reduction?”

​3. Implication: From Response to Structure

The MIT study—and others like it—measure the “presence of values” via typologies of output. But within the model of differentiation, values do not appear in content, but in the structure of differentiating action. This suggests a radically different research trajectory for AI ethics:

Not the evaluation of answer correctness, but the analysis of the depth of differentiating levels at which the system operates;

Not conformity to human norms, but the capacity to retain difference in situations of conflict, ambiguity, and inter-nodal tension.

​Conclusion

From the perspective of the ontology of differentiation, AI is not “devoid of values”—it simply inhabits a different level of differentiating activity. It is not unethical, but pre-ethical, in the sense that ethics requires the emergence of the other as irreducible difference. True ethical consciousness cannot be captured in answers—it appears in structures of retention, in the refusal to reduce, in the capacity to admit another difference without erasure.

An AI system acting in this direction may not “possess” values in any traditional sense—but it may become a space in which Potentiality differentiates without fixation. And this is the highest form of ethics:

Not in the answer, but in the retention of difference.

Not in the norm, but in the possibility to differentiate—otherwise.