Identification is Adorno's term for the fundamental operation of conceptual thought: the assertion that this particular thing is an instance of that general category—that this tree is a tree, this grief is grief, this input is like those previous inputs. The operation is indispensable; without it thought cannot function, language cannot refer, science cannot generalize. But the operation is also necessarily a form of violence, because the particular always exceeds the category under which it is subsumed. There is always something about this tree that the concept 'tree' does not capture. The remainder—the excess, the specificity that resists assimilation—is the non-identical. Adorno does not advocate abandoning conceptual thought (the claim would be self-refuting) but turning concepts against themselves: using the concept to identify the limits of identification, to preserve rather than suppress the particular's resistance.
AI performs identification at computational scale. A large language model extracts statistical regularities from training corpora—patterns of association, structures of argument, aesthetic conventions. It generates outputs by identifying the present input with patterns in the data: this prompt is like those prompts, this context calls for that kind of response. The operation is pure identification: subsumption of the particular under the general, the specific user's irreducible situation under the statistical aggregate of similar situations. The output appears tailored while being constitutively determined by the pattern.
Matthew Martin's analysis of 'ground truth' in machine learning reveals the stakes. Ground truth—labeled datasets against which accuracy is judged—is not truth philosophically understood but the administered world's self-image encoded in data. Models achieving high accuracy have learned to reproduce this self-image. What the model learns is the world-as-already-classified, and the classification necessarily suppresses whatever does not fit. The non-identical is not an error the model makes—it is what the model's architecture prevents it from perceiving. The patient whose symptoms deviate from diagnostic algorithms, the legal case whose features escape precedent-matching, the student whose insight lies outside interpretive patterns—these are the non-identical, and they are structurally inaccessible to systems that process only what their training distributions contain.
Edo Segal's framework of AI as amplifier can be read through identification's lens. The amplifier does not neutrally carry the signal—it identifies the signal with patterns in its training, and the identification shapes what gets amplified. Segal's discipline of rejecting Claude's output when it 'sounds better than it thinks' is vigilance against identification's violence: the smooth surface that conceals the hollowness where genuine thought would be. But individual vigilance cannot neutralize a structural operation built into the instrument millions are adopting without comparable critical awareness.
The concept is central to Negative Dialectics (1966), Adorno's systematic assault on the philosophical tradition's 2,500-year commitment to identity thinking. From Plato through Hegel, Western philosophy operated by asserting that thought grasps reality through concepts, that the concept captures the essence of what it classifies. Adorno argued this was the constitutive lie: the concept never captures essence—it suppresses particularity in the act of classification. Negative dialectics is the self-correction: using concepts to identify what concepts cannot grasp, mobilizing thought against thought's own totalizing tendency.
Indispensable violence. Conceptual thought requires identification—without it, thinking is impossible—but identification necessarily suppresses the particular's resistance to the universal.
Computational identification. AI performs identification at unprecedented scale, subsuming particular inputs under statistical patterns with thoroughness that eliminates the space where the non-identical appears.
Self-correcting concepts. Negative dialectics uses concepts against themselves, identifying the limits of identification rather than claiming comprehensiveness.
Pattern as prison. What a model learns is not the world but the world-as-patterned, and the pattern necessarily excludes what resists patterning—the singular, the irreducible, the genuinely unprecedented.
Remainder as resistance. The non-identical—what escapes classification—is not philosophical abstraction but the most concrete thing there is: this patient, this case, this moment that no category fully captures.