This appendix presents the philosophical argument underlying our work. Following the practice of stating claims clearly and reasoning through to conclusions, we offer here not merely a description of our assumptions but a defense of why this work ought to be pursued.

The Argument in Brief

We advance four premises and derive a moral conclusion:

P1 (Phenomenological): Multiple therapeutic frameworks each disclose genuine aspects of the same client presentations—IFS perceives protective parts, psychodynamic therapy perceives defense mechanisms, ACT perceives experiential avoidance, and each perception is veridical rather than illusory.

P2 (Ontological): From P1, we infer that therapeutic phenomena are polysemous in their being—their existence is rich enough to support multiple valid theoretical readings simultaneously.

P3 (Epistemological): We can know the thing itself through its presentations; appearance discloses being rather than veiling it.

P4 (Computational): Neural networks can know polysemous phenomena more completely, expansively, and granularly than human cognition permits, due to human constraints in bandwidth, attention, memory, and theoretical entrenchment.

C (Moral): Given persistent human suffering (58–76% failing to achieve meaningful response, only one-third reaching remission; Cuijpers et al., 2024, 2021), and given that computational systems may extend epistemic access to therapeutic phenomena in ways that reduce suffering, we have a prima facie obligation to develop such systems carefully and deploy them responsibly.

Defense of the Premises

P1: Phenomenological Veridicality. When an IFS therapist perceives a “protective part,” a psychodynamic therapist perceives a “defense mechanism,” and an ACT therapist perceives “experiential avoidance” in the same client, we deny that at most one of them is correct. Each framework, developed through decades of clinical observation and refinement, has learned to perceive real structure in human psychological functioning. The convergent clinical efficacy across modalities—each helping some clients substantially—would be inexplicable if their perceptions were merely projections onto neutral material. The phenomena are genuinely showing themselves differently to different theoretical lenses, and each showing is a genuine disclosure.

P2: Ontological Polysemy. From P1, we draw an inference about the nature of what is being observed. If multiple incompatible descriptions were each true of the same phenomenon, that would be contradiction. But the descriptions are not incompatible—they are complementary angles on something whose being is complex enough to sustain all of them. The phenomenon is not “really” just a defense mechanism that IFS mislabels, nor “really” just experiential avoidance that psychodynamic theory mystifies. Its being is genuinely polysemous: rich, multidimensional, and irreducible to any single theoretical vocabulary. This is an ontological claim about what therapeutic phenomena are, not merely an epistemological claim about our limited knowledge.

P3: Epistemic Access Through Appearance. We position ourselves against strict Kantian skepticism, which holds that the “thing-in-itself” (Ding an sich) remains forever inaccessible behind appearances. We align instead with the phenomenological tradition—particularly Heidegger’s insight that the way beings show themselves (phenomenology) discloses something about what they are (ontology). Appearance is not a veil over reality but a mode of reality’s self-disclosure. We can know therapeutic phenomena through their presentations precisely because those presentations are how such phenomena exist—not masks but manifestations. This grounds our confidence that closer attention to how therapeutic phenomena appear across frameworks yields genuine knowledge of their nature.

P4: Computational Extension of Epistemic Access. Human clinicians face irreducible cognitive constraints:

  • Bandwidth: Mastering 2–3 frameworks deeply is achievable; integrating 23 simultaneously exceeds human cognitive capacity.
  • Attention: Human perception is serial and selective; tracking all dimensions of presentation at intervention-relevant granularity is impossible.
  • Memory: Patterns across thousands of clients—statistical regularities in what works for whom—exceed human retention and synthesis.
  • Entrenchment: Training in one framework shapes perception, making aspects visible to other frameworks harder to see.

Neural networks trained on polytheoretic data face none of these constraints. They can maintain all 23 frameworks simultaneously, track all dimensions in parallel, learn patterns across millions of encounters, and remain theoretically uncommitted. This does not make them superior judges—judgment involves values and context that humans must retain authority over—but it may make them superior perceivers of polysemous structure.

C: The Moral Conclusion. The inference to moral obligation proceeds as follows: Human suffering persists at scale. Current therapeutic approaches help many but leave 58–76% without meaningful response, and only one-third achieve remission (Cuijpers et al., 2024, 2021)—suggesting failures of treatment selection and personalization rather than treatment failure per se. If computational systems can perceive therapeutic phenomena more completely, and if better perception enables better matching of intervention to need, then such systems may reduce suffering that would otherwise persist. Prima facie, we ought to reduce suffering when we can do so without countervailing harms. Therefore, we have a prima facie obligation to develop therapeutic AI systems carefully—not to replace human therapeutic relationship, but to extend human capacity to understand and respond to suffering.

This is not a claim that we must build such systems regardless of consequences. It is a claim that the potential to reduce suffering creates an obligation to investigate whether and how such systems can be built responsibly. The obligation is defeasible: if such systems prove harmful, or if the risks outweigh benefits, the obligation dissolves. But the default is not neutrality—the default is that alleviating suffering matters, and tools that might help deserve serious pursuit.

Conclusion

We do not claim this philosophical synthesis is novel. We claim it is appropriate—that the nature of therapeutic phenomena, as revealed through decades of clinical observation across traditions, demands exactly this kind of epistemological humility combined with ontological realism. The phenomena are real, rich, and multi-readable. Human suffering is real and persistent. The possibility of computational systems that perceive therapeutic structure more completely than we can, and thereby help us help each other more effectively, is worth pursuing.

The argument, in sum: therapeutic phenomena are polysemous; we can know them through their appearances; machines may know them more completely than we can; suffering persists that better knowledge might alleviate; therefore, we ought to build these systems carefully. This paper represents one step in that work.